Foundation Models vs. CNNs in Computational Pathology: A Paradigm Shift in Medical AI

Abstract

This article provides a comprehensive analysis for researchers and drug development professionals on the pivotal differences between foundation models (FMs) and traditional convolutional neural networks (CNNs) in computational pathology. We explore the foundational shift from supervised, task-specific CNNs to large-scale, self-supervised FMs, detailing their distinct architectural principles, training methodologies, and data requirements. The scope extends to practical applications in diagnosis, prognosis, and biomarker prediction, while critically addressing performance benchmarking, computational burdens, and robustness challenges such as site-specific bias and geometric fragility. Finally, we synthesize validation evidence and discuss the future trajectory of generalist AI in advancing precision medicine.

From Task-Specific Tools to General-Purpose Foundations: Core Paradigms in Pathology AI

The emergence of digital pathology, characterized by the digitization of histopathological slides into high-resolution whole-slide images (WSIs), has created unprecedented opportunities for artificial intelligence (AI) to transform diagnostic workflows [1]. Within this domain, two deep neural network architectures have proven particularly influential: Convolutional Neural Networks (CNNs) and Transformer models. These architectures possess fundamentally different inductive biases—the inherent assumptions and preferences a model embeds to guide its learning process. CNNs are intrinsically biased toward processing local spatial correlations, making them highly effective for analyzing cellular morphology and tissue texture. In contrast, Transformers leverage self-attention mechanisms to model global contextual relationships, enabling them to capture long-range dependencies across disparate tissue regions—a capability critical for understanding complex tissue architecture and tumor microenvironment interactions [2]. The recent advent of foundation models, large-scale neural networks pre-trained on vast datasets, has further accentuated this architectural dichotomy, presenting researchers and clinicians with critical choices for model selection and development. This technical guide examines the core inductive biases of CNNs and Transformers, their implications for computational pathology tasks, and how their integration is shaping the next generation of pathology AI.

Core Architectural Principles and Inductive Biases

Convolutional Neural Networks: The Locality Prior

CNNs are fundamentally designed around the principle of locality and translation invariance. Their architectural inductive biases are hard-coded through a series of operations that explicitly assume the importance of local spatial patterns.

• Local Receptive Fields: Convolutional operations use filters that traverse the input image with small, localized receptive fields (e.g., 3×3 or 5×5 pixels). This design forces the network to focus on local features such as edges, corners, and texture patterns in its initial layers, progressively building more complex, hierarchical representations in deeper layers [2]. In histopathology, this makes CNNs exceptionally adept at identifying nuclear morphology, mitotic figures, and local glandular structures.

• Translation Equivariance: The weight sharing characteristic of convolutional filters means the same filter is applied across all spatial positions of the input. This creates translation equivariance, where shifting the input results in a corresponding shift in the feature map output. This property is invaluable in pathology because a malignant cell or architectural pattern remains diagnostically significant regardless of its position within a tissue section [2].

• Hierarchical Feature Extraction: CNNs naturally extract features in a hierarchical manner, with early layers capturing simple patterns and subsequent layers combining these into increasingly complex constructs. This bottom-up processing mirrors how pathologists first identify individual cellular characteristics before assessing tissue-level organization.

The primary limitation of CNNs lies in their constrained receptive fields. Even deep networks with many layers have difficulty modeling long-range spatial relationships without additional architectural modifications, as their fundamental operation remains locally bounded [2].

Transformer Architectures: The Global Context Prior

Transformers, originally developed for natural language processing, operate on a fundamentally different principle: the self-attention mechanism. This mechanism allows the model to weigh the importance of all elements in the input sequence when processing each individual element.

• Global Context via Self-Attention: The self-attention mechanism computes pairwise interactions between all patches (tokens) in an image. For an input sequence of tokens, it calculates query, key, and value vectors for each token, then computes attention weights as:

  Attention(Q, K, V) = softmax(QKᵀ/√dₖ)V [2]

  This operation enables each patch to directly influence, and be influenced by, every other patch in the image, regardless of their relative positions. For WSIs, this allows the model to identify relationships between geographically distant but diagnostically linked tissue regions.

• Positional Encoding: Unlike CNNs that inherently understand spatial relationships through convolution, Transformers require explicit positional encodings to incorporate spatial information. These encodings, added to the patch embeddings, inform the model about the relative or absolute positions of patches in the original image [2].

• Minimal Spatial Inductive Bias: Transformers intentionally incorporate minimal prior assumptions about spatial relationships, enabling them to learn complex, non-local patterns directly from data. This flexibility comes at the cost of requiring substantial training data to discover spatial relationships that CNNs assume by design.

Vision Transformers (ViTs) adapt this architecture for images by dividing the input into fixed-size, non-overlapping patches, linearly embedding each patch, and processing the resulting sequence through standard Transformer encoder layers [2].

Comparative Analysis of Architectural Properties

Table 1: Core Architectural Properties of CNNs vs. Transformers

Property	Convolutional Neural Networks (CNNs)	Transformer Models
Primary Inductive Bias	Locality & Translation Equivariance	Global context & Minimal spatial assumptions
Receptive Field	Local (grows hierarchically but remains limited)	Global (from first layer)
Core Operation	Convolution & Pooling	Self-Attention & Layer Normalization
Spatial Understanding	Built-in via convolution kernels	Learned via positional encodings
Parameter Efficiency	High (weight sharing)	Lower (dense attention matrices)
Data Requirements	Moderate	Substantial
Feature Integration	Hierarchical & local-to-global	Any-to-any & context-aware

Empirical Performance in Pathology Applications

Benchmarking Studies and Direct Comparisons

Rigorous empirical evaluations have quantified the performance differences between CNN and Transformer architectures across various pathology tasks. A comprehensive 2025 study trained and evaluated 14 deep learning models—including both CNN-based and Transformer-based architectures—on the BreakHis breast cancer dataset [3]. The findings reveal nuanced performance characteristics:

• Binary Classification Performance: In the less complex binary classification task (benign vs. malignant), multiple models achieved excellent performance. CNN-based models including ResNet50, RegNet, and ConvNeXT, along with the Transformer-based foundation model UNI, all reached an area under the curve (AUC) of 0.999. The best overall performance was achieved by ConvNeXT (a CNN variant), which attained an accuracy of 99.2%, specificity of 99.6%, and F1-score of 99.1% [3].

• Multi-class Classification Performance: In the more challenging eight-class classification task, performance differences became more pronounced. The best-performing model was the fine-tuned foundation model UNI (Transformer-based), which attained an accuracy of 95.5%, specificity of 95.6%, and F1-score of 95.0% [3]. This suggests that Transformer architectures may particularly excel in complex classification scenarios requiring discrimination between subtle morphological subtypes.

• Micro-Metastasis Detection: For particularly challenging tasks like lymph node micro-metastasis detection in breast cancer, hybrid approaches have shown promise. One study developed MetaTrans, a novel network combining meta-learning with Transformer and CNN components, which demonstrated superior performance on multi-center datasets compared to pure CNN or Transformer baselines [4].

Table 2: Experimental Performance Comparison on Pathology Tasks

Model Architecture	Task	Dataset	Key Metric	Performance
ConvNeXT (CNN)	Binary Classification	BreakHis	Accuracy	99.2%
UNI (Transformer)	8-class Classification	BreakHis	Accuracy	95.5%
Virchow (Foundation)	Pan-Cancer Detection	Multi-Cancer	AUC	0.950
rMetaTrans (Hybrid)	Micro-Metastasis Detection	BLCN-MiD	Recall	~95% (inferred)
MSNet (Hybrid)	Lung Adenocarcinoma	Private Dataset	Accuracy	96.55%

Foundation Models in Pathology

Foundation models represent a paradigm shift in computational pathology. These models are pre-trained on massive, diverse datasets using self-supervised learning, producing versatile feature representations that can be adapted to various downstream tasks with minimal fine-tuning.

• Virchow: Trained on approximately 1.5 million H&E-stained WSIs from 100,000 patients, Virchow is a 632 million parameter Vision Transformer. It enables pan-cancer detection with an AUC of 0.950 across nine common and seven rare cancers, demonstrating remarkable generalization capability [5]. This performance highlights how scale and diversity in pre-training can overcome Transformers' traditional data hunger.

• TITAN: The Transformer-based pathology Image and Text Alignment Network is a multimodal whole-slide foundation model pre-trained on 335,645 WSIs. TITAN incorporates both visual self-supervised learning and vision-language alignment with corresponding pathology reports, enabling capabilities like zero-shot classification and cross-modal retrieval without task-specific fine-tuning [6].

• UNI: As one of the first general-purpose pathology models trained via self-supervised learning on more than 100,000 diagnostic-grade H&E-stained whole slide images, UNI demonstrated strong performance across 34 computational pathology tasks and exhibits notable capabilities in resolution-agnostic classification and few-shot learning [3].

These foundation models overwhelmingly leverage Transformer architectures due to their superior scalability and ability to capture the long-range dependencies necessary for whole-slide analysis.

Implementation and Workflow Methodologies

Experimental Protocols for Model Development

The development of performant models in computational pathology follows carefully designed experimental protocols that account for the unique challenges of histopathological data.

• Whole-Slide Image Processing: WSIs are gigapixel-sized images that cannot be processed directly. The standard methodology involves dividing WSIs into smaller patches (typically 256×256 or 512×512 pixels at 20× magnification). For example, the Virchow foundation model processes patches of 512×512 pixels at 20× magnification, extracting 768-dimensional features for each patch [5] [6].

• Multi-Magnification Analysis: Many approaches employ a multi-stage process inspired by pathological practice. The MetaTrans network, for instance, uses separate models for different magnification levels: a tissue-recognition model for low magnification (4×) regions of interest and a cell-recognition model for high magnification (10×) images [4]. This approach mirrors how pathologists first scan at low power to identify suspicious areas before examining cellular details at higher magnification.

• Weakly Supervised Learning: Given the difficulty of pixel-level annotations, slide-level labels are often used in a multiple instance learning framework. Features from individual patches are aggregated to make slide-level predictions, using methods like attention-based pooling or transformer aggregators [5].

• Data Augmentation and Normalization: Histopathology images exhibit significant variability in staining protocols and scanner characteristics. Standard preprocessing includes color normalization techniques like histogram matching or more advanced deep learning approaches to improve model robustness and generalization [1].

Hybrid Architectures: Integrating Local and Global Context

Recognizing the complementary strengths of CNNs and Transformers, researchers have developed hybrid architectures that leverage both local feature extraction and global contextual modeling.

• Fusion Architectures: One approach jointly combines CNN and Vision Transformer modules, where the CNN processes the input image to produce local feature embeddings, while the ViT computes global embeddings from patch sequences. These representations are then concatenated to form a joint feature vector for classification [2]. Empirical results on breast cancer histopathology images demonstrate that ViT+CNN fusion models consistently outperform standalone CNN or ViT models [2].

• MSNet for Lung Adenocarcinoma: This framework employs a dual data stream input method, combining Swin Transformer and MLP-Mixer models to address global information between patches and local information within each patch. The model uses a Multilayer Perceptron (MLP) module to fuse these local and global features for classification, achieving 96.55% accuracy on lung adenocarcinoma pathology images [7].

• MetaTrans for Micro-Metastasis: Designed for limited-data scenarios, MetaTrans optimizes Transformer architecture with meta-learning and CNN components to improve detection of lymph node micro-metastasis in breast cancer. The network is trained end-to-end and remains effective even when sample sizes are significantly smaller than those available for macro-metastases [4].

Diagram 1: Hybrid CNN-Transformer Workflow for Digital Pathology

Successful implementation of CNN and Transformer models in pathology requires careful selection of computational frameworks, datasets, and evaluation methodologies.

Table 3: Essential Research Reagents for Pathology AI Development

Resource Category	Specific Examples	Function & Application
Public Datasets	BreakHis, CAMELYON, TCGA, LC25000	Benchmark model performance on standardized datasets [3] [7] [4]
Foundation Models	Virchow, UNI, TITAN, CONCH	Pre-trained feature extractors for transfer learning [3] [5] [6]
Computational Frameworks	PyTorch, TensorFlow, MONAI	Model development, training, and inference pipelines
Whole-Slide Processing	OpenSlide, CuCIM, HistomicsUI	Efficient handling and patch extraction from gigapixel WSIs [4]
Annotation Tools	ASAP, QuPath, Digital Slide Archive	Region-of-interest marking and label creation
Evaluation Metrics	AUC, Accuracy, F1-Score, Cohen's Kappa	Quantitative performance assessment [3]
Interpretability Methods	Grad-CAM, Attention Rollout, Attention Maps	Model decision explanation and verification [2]

Diagram 2: Architecture Selection Framework for Pathology Tasks

The inductive biases of CNNs and Transformers create a complementary relationship in computational pathology rather than a competitive one. CNNs provide efficient, locally-biased feature extraction well-suited to cellular-level analysis and scenarios with limited data. Transformers offer global contextual modeling capabilities essential for understanding tissue architecture and tumor microenvironment interactions, particularly valuable in complex diagnostic tasks. Foundation models, predominantly Transformer-based, are demonstrating remarkable capabilities in pan-cancer detection and rare cancer identification, achieving clinical-grade performance that matches or exceeds specialized models [5].

The future of pathology AI lies not in choosing between these architectures but in strategically combining them. Hybrid models that leverage CNN-Transformer fusion, multimodal approaches integrating histopathology with genomic and clinical data, and foundation models adapted through efficient fine-tuning represent the most promising directions. As these technologies mature, their successful clinical integration will depend not only on architectural advancements but also on addressing challenges in interpretability, robustness, and seamless workflow integration—ultimately fulfilling the promise of AI-powered precision pathology.

The field of computational pathology is undergoing a profound transformation, driven by a paradigm shift from traditional, supervised convolutional neural networks (CNNs) to large-scale foundation models trained with self-supervised learning. This transition represents more than merely a technical improvement—it constitutes a fundamental reimagining of how artificial intelligence models learn from histopathology data. Traditional CNNs, while revolutionary in their own right, face inherent limitations in generalizability, annotation dependency, and computational efficiency when applied to the gigapixel-scale complexity of whole slide images. Foundation models address these challenges through pre-training on massive, uncurated datasets, capturing fundamental biological representations rather than merely memorizing annotated patterns. This technical guide examines the architectural, methodological, and practical distinctions between these approaches within pathology research, providing researchers and drug development professionals with a comprehensive framework for understanding this critical evolution in digital pathology.

Fundamental Architectural Divergence

Traditional CNN Architectures in Pathology

Convolutional Neural Networks defined the first wave of deep learning success in computational pathology. Their architecture is fundamentally grounded in inductive biases well-suited to image data—local connectivity, spatial invariance, and hierarchical feature learning. CNNs process images through stacked convolutional layers that progressively extract features from low-level edges and textures to high-level morphological patterns, with pooling operations providing spatial invariance and reducing dimensionality [8]. This architectural approach enables CNNs to effectively learn from pixel-level annotations for tasks such as nuclei segmentation, mitosis detection, and tumor region identification [9].

In pathology specifically, the U-Net architecture with its encoder-decoder structure and skip connections became the benchmark for segmentation tasks, while variants like ResNet and VGG were widely adopted for classification [10]. However, a critical limitation of these architectures is their fixed receptive field, which restricts their ability to capture long-range dependencies in histopathology images—a significant drawback when analyzing tissue microenvironments and architectural patterns that extend across large spatial distances [10].

Foundation Model Architectures

Pathology foundation models predominantly leverage transformer-based architectures, specifically Vision Transformers (ViTs), which process images as sequences of patches using self-attention mechanisms. Unlike CNNs' local processing, self-attention enables global contextual understanding from the initial layers, capturing relationships between distant tissue regions that are clinically significant but challenging for CNNs to model effectively [11]. The scale of these models is substantially larger, with parameter counts ranging from millions in traditional CNNs to hundreds of billions in foundation models like Prov-GigaPath [11].

Table 1: Architectural Comparison Between CNNs and Pathology Foundation Models

Characteristic	Traditional CNNs	Pathology Foundation Models
Core Architecture	Convolutional layers with pooling	Vision Transformers (ViTs) with self-attention
Receptive Field	Local, increases hierarchically	Global from first layer
Parameter Scale	Millions (e.g., ResNet-50: ~25M)	Hundreds of millions to billions (e.g., UNI: 303M, Virchow: 631M) [11]
Primary Strengths	Local feature extraction, translation invariance	Long-range dependency modeling, contextual understanding
Handling Whole Slide Images	Requires tiling and separate processing	Can process patch sequences with positional encoding

This architectural evolution enables foundation models to develop a more comprehensive understanding of tissue architecture and cellular relationships, mirroring how human pathologists integrate local cytological details with global tissue patterns to reach diagnoses.

Training Paradigms: From Supervision to Self-Supervision

Traditional Supervised Learning

Supervised learning for CNNs requires extensive datasets with pixel-level or tile-level annotations meticulously labeled by pathologists. This paradigm dominated early computational pathology research, with models trained to map input images to specific outputs based on these annotations. The process involves forward propagation of image data through the network, calculation of loss between predictions and ground truth labels, and backward propagation for weight optimization [8]. While effective for specific tasks, this approach suffers from several critical limitations: (1) annotation bottleneck—the time and expertise required to label datasets limits scale; (2) task specificity—models excel only at the tasks they were trained on, with poor transfer learning capabilities; and (3) annotation bias—models inherit the subjective interpretations and potential biases of the annotating pathologists [9].

Self-Supervised Learning for Foundation Models

Self-supervised learning represents a fundamental shift from task-specific supervision to pretext task learning on unlabeled data. SSL methods create learning signals directly from the data itself, enabling models to learn generally useful representations without manual annotation. The core principle involves pre-training on vast unlabeled datasets—often millions of whole slide images—followed by minimal fine-tuning on specific downstream tasks [11].

Table 2: Self-Supervised Learning Methods in Pathology Foundation Models

SSL Category	Representative Algorithms	Mechanism	Pathology Examples
Contrastive Learning	MoCo v3, DINO, DINOv2	Learning representations by contrasting similar and dissimilar samples	CTransPath (SRCL), Virchow (DINOv2) [11]
Masked Image Modeling	MAE, iBOT	Reconstructing masked portions of input images	Phikon (iBOT) [11], MAE-based models [11]
Self-Distillation	DINO, BYOL	Student network learning from teacher network without labels	UNI, Virchow (DINOv2) [11]

The scale of data used for pre-training pathology foundation models dwarfs that available for supervised approaches. For instance, UNI was trained on 100 million tiles from 100,000 diagnostic whole slide images [11], while Virchow utilized 2 billion tiles from 1.5 million slides [11]. This massive scale enables the learning of robust, generalizable representations that capture the fundamental biological and morphological patterns in histopathology images across diverse tissue types, staining protocols, and disease states.

Quantitative Performance Benchmarks

Recent comprehensive benchmarking studies reveal the substantial performance advantages of foundation models across diverse pathology tasks. Campanella et al. demonstrated that SSL-trained pathology models consistently outperform both supervised models and models pre-trained on natural images across six clinical tasks spanning three anatomical sites and two institutions [11]. Similarly, a clinical benchmark of public SSL pathology foundation models showed that models pre-trained with DINOv2, such as UNI and Phikon-v2, achieve state-of-the-art performance on tissue classification, mutation prediction, and survival analysis tasks [11].

Table 3: Performance Comparison of Selected Pathology Foundation Models

Model	Architecture	SSL Algorithm	Training Data	Reported Performance Advantages
CTransPath	Hybrid CNN-Transformer	SRCL (MoCo-based)	15.6M tiles, 32K slides	Superior to ImageNet pre-trained models for WSI classification [11]
UNI	ViT-Large	DINOv2	100M tiles, 100K slides	SOTA across 33 diverse tasks [11]
Virchow	ViT-Huge	DINOv2	2B tiles, 1.5M slides	Outperforms previous models for rare cancer detection [11]
Phikon-v2	ViT	DINOv2	460M tiles, 58K slides	Robust performance across 8 slide-level tasks with external validation [11]

A critical advantage of foundation models is their data efficiency in downstream tasks. SSL pre-trained models achieve strong performance with limited labeled examples—one study demonstrated that such models require only 25% of labeled data to achieve 95.6% of full performance compared to 85.2% for supervised baselines, representing a 70% reduction in annotation requirements [12]. This data efficiency significantly accelerates research and development cycles in both academic and pharmaceutical settings.

Robustness and Generalization Challenges

Despite their impressive capabilities, pathology foundation models face significant robustness challenges, particularly regarding sensitivity to technical variations between medical centers. A recent landmark study evaluated ten publicly available pathology foundation models and found that most current models remain unrobust to medical center differences, with their embedding spaces more strongly organized by medical center signatures than by biological features [13].

The study introduced a novel Robustness Index (RI) that measures the degree to which biological features dominate confounding features in model embeddings. Formally defined as:

[ Rk = \frac{\sum{i=1}^n \sum{j=1}^k \mathbf{1}(yj = yi)}{\sum{i=1}^n \sum{j=1}^k \mathbf{1}(cj = c_i)} ]

Where (y) represents biological class, (c) represents medical center, and (k) is the number of nearest neighbors considered [13]. Alarmingly, only one of the ten evaluated models achieved a robustness index greater than one, indicating that for most models, medical center confounders dominated over biological features in their representation spaces [13]. This technical confounder sensitivity poses significant challenges for clinical deployment where models must perform consistently across diverse healthcare settings with variations in staining protocols, scanner equipment, and tissue processing methods.

Experimental Protocols and Methodologies

Protocol for Traditional Supervised CNN Training

A standardized protocol for training CNNs for pathology image analysis involves these critical steps:

• Data Preparation: Extract patches from whole slide images (WSIs) at appropriate magnification (typically 20×). Standard patch sizes range from 256×256 to 1024×1024 pixels. Apply stain normalization to reduce color variation between institutions.

• Annotation: Generate pixel-level or patch-level annotations through expert pathologist review. Common annotation types include segmentation masks for nuclei or tumor regions, and categorical labels for tissue types or disease states.

• Data Augmentation: Apply transformations including rotation, flipping, color jittering, and elastic deformations to increase dataset diversity and improve model robustness.

• Model Selection: Choose appropriate CNN architecture (U-Net for segmentation, ResNet for classification) either training from scratch or using transfer learning from ImageNet pre-trained weights.

• Training: Optimize model parameters using supervised loss functions (cross-entropy for classification, Dice loss for segmentation) with appropriate batch sizes and learning rate schedules.

• Validation: Evaluate performance on held-out test sets from the same institution, with careful monitoring for overfitting using techniques such as early stopping.

Protocol for Self-Supervised Foundation Model Pre-training

The emerging standard methodology for pre-training pathology foundation models comprises:

• Unlabeled Data Curation: Collect large-scale, diverse histopathology datasets spanning multiple cancer types, tissue sources, and medical centers. Models like Virchow2 incorporate 3.1 million WSIs from nearly 200 tissue types [11].

• Pretext Task Design: Implement SSL algorithms such as DINOv2 or MAE that create supervisory signals from the data itself without human annotation.

• Multi-Scale Processing: Extract image patches at multiple magnifications (e.g., 5×, 10×, 20×, 40×) to capture both tissue-level architecture and cellular-level details [11].

• Large-Scale Distributed Training: Leverage substantial computational resources (typically hundreds of GPUs) for extended training periods (days to weeks) on billion-patch datasets.

• Embedding Space Validation: Analyze resulting feature spaces using methodologies like the Robustness Index to quantify organization by biological versus confounding features [13].

Protocol for Robustness Evaluation

To assess model robustness to medical center variations, researchers should implement:

• Multi-Center Dataset Curation: Collect data from at least 3-5 independent medical centers with documented variations in staining protocols and scanning equipment.

• Robustness Index Calculation: For each sample, identify k-nearest neighbors in the model's embedding space (typically k=50). Calculate the ratio of neighbors sharing biological class versus medical center identity [13].

• Cross-Center Performance Evaluation: Measure model performance separately for each medical center, analyzing performance degradation relative to the training center.

• Confusion Matrix Analysis: Examine whether classification errors correlate with medical center identity rather than biological similarity.

The Scientist's Toolkit: Essential Research Reagents

Table 4: Essential Computational Tools for Pathology Foundation Model Research

Tool Category	Specific Solutions	Primary Function	Application in Research
Deep Learning Frameworks	PyTorch, TensorFlow	Model implementation and training	Core infrastructure for developing and training both CNNs and foundation models
Whole Slide Image Processing	QuPath, PyHIST, OpenSlide	WSI annotation, tiling, and preprocessing	Essential for data preparation and patch extraction from gigapixel WSIs [8]
Computational Pathology Libraries	TIAToolbox, HistoML	Domain-specific algorithms and utilities	Provides standardized implementations of pathology-specific processing pipelines
Self-Supervised Learning Implementations	DINOv2, MAE reference code	SSL algorithm implementations	Critical for reproducing state-of-the-art pre-training methodologies
Benchmarking Platforms	Clinical pathology benchmarks [11]	Standardized model evaluation	Enables fair comparison across different architectures and training regimes
Computational Resources	GPU clusters (H100, A100), cloud computing	Large-scale model training	Essential for training foundation models on million-slide datasets

The transition from supervised learning on labeled data to large-scale self-supervision represents a fundamental paradigm shift in computational pathology. While foundation models demonstrate remarkable capabilities and generalization potential, important challenges remain in achieving true robustness to technical variations between medical centers. Future research directions include: (1) developing novel SSL objectives specifically designed to learn stain-invariant and scanner-invariant representations; (2) creating efficient fine-tuning methodologies that preserve pre-trained knowledge while adapting to new tasks and domains; and (3) establishing standardized benchmarking frameworks that rigorously evaluate model performance across diverse clinical settings [13] [11].

The ongoing SLC-PFM NeurIPS 2025 competition highlights the continued momentum in this field, providing researchers with unprecedented access to large-scale pathology datasets and establishing standardized evaluation protocols across 23 clinically relevant tasks [14]. As foundation models continue to evolve, they hold immense promise for transforming pathology research and clinical practice—enabling more accurate diagnoses, revealing novel biomarkers, and ultimately improving patient outcomes through computational advances that capture the complex biological reality of disease processes.

The field of computational pathology is undergoing a profound transformation, driven by a fundamental shift in its underlying data philosophy. The traditional approach, reliant on medium-scale, meticulously curated datasets for training task-specific Convolutional Neural Networks (CNNs), is being challenged by a new paradigm that leverages very large, unlabeled corpora of whole slide images (WSIs) to train foundational vision models. This shift mirrors the revolution witnessed in natural language processing and general computer vision but is uniquely complex due to the gigapixel size, structural heterogeneity, and clinical stakes of pathology data. This whitepaper delineates the core differences between traditional CNNs and foundation models within pathology research, examining the technical, methodological, and philosophical underpinnings of this transition. It provides a comprehensive guide for researchers and drug development professionals navigating this new landscape, complete with quantitative benchmarks, experimental protocols, and essential toolkits.

Core Architectural and Methodological Differences

The transition from CNNs to foundation models represents more than a simple improvement in scale; it constitutes a fundamental redesign of model architecture, learning objectives, and data utilization.

Traditional CNNs: The Curated Data Paradigm

Traditional CNNs, such as ResNet50, have been the workhorses of early computational pathology. Their development follows a supervised learning approach that is heavily dependent on human expertise and curation.

• Architecture: CNNs are characterized by their inductive biases, such as translation equivariance, built directly into their architecture through convolutional layers. These biases make them highly data-efficient, enabling effective learning from smaller, annotated datasets [15].
• Data Requirements: These models require extensive pixel-level annotations for tasks like segmentation and detection. Creating these datasets is time-consuming, expensive, and represents a significant bottleneck for scaling applications [16] [17].
• Scope and Limitations: CNNs are typically designed as task-specific models. A model trained for prostate cancer grading cannot be applied to lung cancer subtyping without retraining. This narrow focus limits their generalizability and increases the cumulative resource investment for developing multiple diagnostic tools [16].

Foundation Models: The Unlabeled Corpora Paradigm

Pathology foundation models, such as UNI, Prov-GigaPath, and Virchow, represent a paradigm shift toward large-scale, self-supervised learning on diverse, unlabeled data.

• Architecture: Most modern foundation models are based on Vision Transformers (ViTs). Unlike CNNs, ViTs have minimal inherent inductive biases, which allows them to learn more complex, generalizable patterns directly from data when trained on a massive scale [15] [18].
• Data Utilization: These models are pretrained using self-supervised learning (SSL) algorithms like DINOv2 or masked autoencoding on vast, unlabeled datasets comprising millions of tissue tiles. This process creates a universal feature extractor without the need for manual annotations [18] [11].
• Scope and Promise: The goal is a general-purpose model that can be adapted with minimal effort (e.g., via linear probing or light fine-tuning) to a wide array of downstream tasks—from cancer subtyping and mutation prediction to biomarker identification—across different organs and diseases [11].

Table 1: Fundamental Differences Between Traditional CNNs and Pathology Foundation Models

Feature	Traditional CNNs	Pathology Foundation Models
Core Architecture	Convolutional layers with strong inductive biases	Vision Transformers with minimal inductive biases
Learning Paradigm	Supervised learning	Self-supervised learning (SSL)
Primary Data Source	Medium-scale, curated, labeled datasets	Very large-scale, unlabeled whole slide image corpora
Annotation Requirement	High (pixel-/slide-level)	None for pre-training
Typical Model Scope	Task-specific	General-purpose, adaptable to many tasks
Representative Examples	ResNet50, U-Net [16]	Prov-GigaPath, UNI, Virchow [15] [18] [11]

Quantitative Performance Benchmarks

Empirical evidence demonstrates the tangible benefits of the foundation model approach, particularly in performance and robustness, though not without caveats.

Diagnostic Accuracy and Task Performance

Foundation models have consistently demonstrated superior performance on standardized benchmarks. In a systematic assessment on clinical datasets, foundation models outperformed models pretrained on natural images (e.g., ImageNet) across a variety of tasks [11]. A specific study on kidney disease classification reported that foundation models achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of over 0.980 on internal validation for diagnosing healthy controls, acute interstitial nephritis, and diabetic kidney disease. Crucially, in external validation, the performance of a traditional ImageNet-pretrained ResNet50 "markedly dropped," while the foundation models maintained robust performance [15]. Prov-GigaPath attained state-of-the-art performance on 25 out of 26 benchmark tasks, including cancer subtyping and genomic mutation prediction, showing significant improvements over the next-best methods [18].

Data Scale and Model Size

The power of foundation models is unlocked by training on datasets that are orders of magnitude larger than those used for traditional CNNs.

Table 2: Scale Comparison of Representative Pathology Models

Model	Architecture	Training Data (Tiles / Slides)	Parameters	SSL Algorithm
CTransPath	Hybrid CNN-Transformer	16M / 32K [11]	28M [11]	SRCL (MoCo)
Phikon	Vision Transformer (ViT-Base)	43M / ~6K [11]	86M [11]	iBOT
UNI	ViT-Large	100M / 100K [15] [11]	303M [11]	DINOv2
Virchow	ViT-Huge	2B / ~1.5M [11]	631M [11]	DINOv2
Prov-GigaPath	GigaPath (LongNet)	1.3B / 171K [15] [18]	1.1B [11]	DINOv2 + MAE

Critical Analysis: Challenges and Limitations of the New Paradigm

Despite their promise, foundation models in pathology face significant challenges that temper the enthusiasm around a simple "bigger is better" narrative.

• Domain Mismatch and Robustness: A core failure mode is domain mismatch. Models can be sensitive to variations in staining protocols, scanner types, and tissue preparation across different medical centers. One multi-center study found that for most foundation models, embeddings grouped more strongly by the source hospital than by biological class, leading to performance drops of 15-25% on external validation [19]. This highlights a critical lack of robustness for direct clinical deployment.
• Information Bottleneck and Task Complexity: Foundation models often function by compressing large image patches into small embedding vectors, which can create an information bottleneck. While these compressed representations excel at simple tasks like disease detection, their performance can drop to near-chance levels for more complex tasks such as biomarker prediction or immunotherapy response forecasting, which require nuanced tissue context [20].
• Computational Cost and Adaptation Issues: The immense size of foundation models creates a practical barrier to adaptation. Full fine-tuning is often unstable and computationally prohibitive, forcing researchers to rely on "linear probing" (training a simple classifier on frozen embeddings). This contradicts the foundational model premise of easy adaptation and can yield suboptimal performance compared to end-to-end training of smaller, task-specific models [19]. One study noted that foundation models could consume up to 35× more energy than a task-specific model [19].

Experimental Protocols for Model Evaluation

For researchers seeking to validate and compare these models, a standardized experimental protocol is essential. The following workflow, specifically for slide-level classification, is widely adopted and can be applied to both public and proprietary datasets.

Step 1: Data Curation and Preprocessing

• Dataset Curation: Assemble a dataset of WSIs with slide-level labels (e.g., diagnosis, mutation status). Ensure patient-level splits to avoid data leakage. Multi-center datasets are preferred for robust evaluation [15] [11].
• Patch Extraction: Use a tool like Slideflow to tile WSIs at a standard magnification (e.g., 20x) into non-overlapping patches (e.g., 256x256 pixels) [15].
• Background Filtering: Apply filters (e.g., Otsu's thresholding, Gaussian blur) to remove non-tissue background patches [15].

Step 2: Feature Extraction

• Encoder Selection: Choose a pretrained foundation model (e.g., UNI, Phikon, Prov-GigaPath) or a baseline CNN (e.g., ImageNet-pretrained ResNet50) as the feature extractor.
• Embedding Generation: Process each tissue patch through the encoder to generate a feature vector (embedding). This converts a WSI from a bag of pixels into a bag of feature vectors [15] [11].

Step 3: Slide-Level Aggregation and Classification

• Multiple Instance Learning (MIL): Since WSIs are too large for direct processing, treat each WSI as a "bag" of patch instances. Use an MIL aggregator to combine patch features into a single slide-level representation. Common aggregators include:
  • ABMIL: An attention-based mechanism that learns to weight patches by their diagnostic importance [15].
  • CLAM: A clustering-constrained attention MIL model that provides interpretable heatmaps [15].
  • TransMIL: Uses transformer layers to model correlations between patches [15].
• Model Training & Validation: Train the MIL model (aggregator + classifier) on the slide-level features and labels. Perform k-fold cross-validation and, critically, external validation on a held-out dataset from a different institution to assess generalizability [15].

The Scientist's Toolkit: Essential Research Reagents

Successful implementation of foundation models requires a suite of computational tools and resources. The table below details key components for building and evaluating pathology foundation models.

Table 3: Essential Research Reagents for Pathology Foundation Model Research

Reagent / Resource	Type	Function & Application	Exemplars / Notes
Large-Scale WSI Datasets	Data	Pretraining and benchmarking foundation models. Scale and diversity are critical.	TCGA, KPMP, JP-AID [15]; Proprietary datasets (e.g., Prov-Path, MSKCC data) [18] [11].
Self-Supervised Learning (SSL) Algorithms	Algorithm	Enables pretraining on unlabeled images by defining a pretext task.	DINOv2, iBOT, Masked Autoencoder (MAE) [11].
Vision Transformer (ViT)	Model Architecture	Backbone network for most foundation models; processes images as sequences of patches.	ViT-Base, ViT-Large, ViT-Huge [11]; GigaPath for whole-slide context [18].
Multiple Instance Learning (MIL) Frameworks	Model Architecture	Aggregates patch-level features for slide-level prediction without patch-level labels.	ABMIL, CLAM, TransMIL [15].
Computational Framework	Software	Libraries for WSI processing, model training, and visualization.	Slideflow, PyTorch, TIAToolbox [15].

The shift from medium-scale curated datasets to very large unlabeled corpora represents a fundamental evolution in the data philosophy of computational pathology. Foundation models, pretrained with self-supervised learning on massive WSI datasets, offer a powerful and versatile alternative to traditional, task-specific CNNs. Their demonstrated superiority in benchmark tasks and improved robustness marks a significant leap forward.

However, this new paradigm is not a panacea. Challenges related to domain robustness, computational burden, and effective adaptation remain active areas of research. The path forward likely involves a synthesis of scale and specificity—developing larger, more diverse pretraining datasets while also innovating in domain-robust architectures and efficient fine-tuning methods. For researchers and drug developers, mastering this new paradigm is essential. It promises to unlock deeper biological insights from pathology images, accelerate biomarker discovery, and ultimately, contribute to more personalized and effective patient therapies.

The field of computational pathology is undergoing a fundamental transformation, moving from specialized, task-specific models to general-purpose, scalable foundations. Traditional Convolutional Neural Networks (CNNs) have long been the workhorse of digital pathology image analysis, but their architectural limitations constrain their ability to leverage massive datasets and develop generalized representations of histological structures. Foundation models, trained on broad data at unprecedented scale, represent a paradigm shift not merely in size but in core capability and application philosophy. These models differ fundamentally in their approach to expressiveness—the ability to capture and represent complex pathological patterns—and scalability—the capacity to improve performance with increasing model and data size. Understanding these differences is crucial for researchers, scientists, and drug development professionals seeking to leverage artificial intelligence for advancing precision medicine and therapeutic discovery.

The term "foundation model," popularized by the 2021 Stanford Report, refers to any model "trained on broad data (generally using self-supervision at scale) that can be adapted to a wide range of downstream tasks" [21]. What makes these models foundational is not their specific architecture but their applicability across diverse domains through adaptation mechanisms like prompting and fine-tuning. In pathology, this transition mirrors the broader AI landscape where companies are increasingly training foundation models to solve highly specialized problems, meet compliance requirements, and build core competency in the technology [21].

Architectural Foundations: CNN vs. Foundation Model Design

Traditional CNN Architecture in Pathology

Convolutional Neural Networks have demonstrated remarkable success in pathological image analysis through their hierarchical feature extraction approach. CNN-based architectures such as ResNet50, VGG16, DenseNet121, and EfficientNet employ a series of convolutional layers that progressively detect increasingly complex patterns—from edges and textures in early layers to specific cellular structures and tissue organizations in deeper layers [3]. This inductive bias toward translational invariance and local feature detection makes CNNs particularly well-suited for identifying morphological patterns in histopathological images where local cellular arrangements carry significant diagnostic meaning.

The expressiveness of CNNs is fundamentally constrained by their receptive fields—the segment of input image that affects a particular neuron's activation. Although deeper networks expand receptive fields through pooling and strided convolutions, they primarily capture local spatial hierarchies rather than global contextual relationships across entire whole slide images (WSIs). This limitation becomes particularly significant in pathology, where diagnostic interpretation often requires understanding spatial relationships between distant tissue regions and integrating contextual information across multiple scales.

Foundation Model Architecture and Scaling Principles

Foundation models in pathology, particularly transformer-based architectures like UNI and Prov-GigaPath, employ self-attention mechanisms that enable global receptive fields from the initial processing stages [3] [22]. Unlike CNNs that process images through local convolutional filters, vision transformers typically divide images into patches and process them through self-attention layers that can model relationships between all patches simultaneously. This architectural difference fundamentally enhances expressiveness by capturing long-range dependencies across entire tissue sections without being constrained by local receptive fields.

The scalability of foundation models emerges from both architectural considerations and training methodology. The transformer architecture demonstrates remarkably consistent scaling behavior—performance predictably improves with increased model parameters, training data, and computational budget [21]. This scalability enables foundation models to leverage massive unlabeled datasets through self-supervised learning objectives, learning rich representations of histopathological structures without requiring expensive manual annotations. Prov-GigaPath, for instance, was trained on 1.3 billion image patches extracted from 171,189 whole slide images, demonstrating the massive data scaling potential of these approaches [3].

Table 1: Architectural Comparison Between CNN and Foundation Models in Pathology

Characteristic	Traditional CNN	Pathology Foundation Model
Core Architecture	Convolutional layers with local receptive fields	Transformer blocks with self-attention mechanisms
Receptive Field	Local, expands through network depth	Global from initial layers
Training Data Scale	Thousands to hundreds of thousands of images	Millions to billions of image patches [3]
Parameter Count	Typically millions to low hundreds of millions	Hundreds of millions to billions
Primary Learning Approach	Supervised learning with labeled data	Self-supervised pre-training followed by fine-tuning
Context Integration	Limited to local spatial hierarchies	Whole-slide and cross-slide relationships
Representative Models	ResNet50, VGG16, EfficientNet [3]	UNI, Prov-GigaPath, GigaPath [3] [22]

Quantitative Performance Comparison

Experimental Benchmarking in Diagnostic Tasks

Comparative studies provide compelling evidence for the performance advantages of scaled foundation models in complex pathology tasks. A 2025 comprehensive analysis evaluated 14 deep learning models—including both CNN-based and transformer-based architectures—on the BreakHis v1 dataset for breast cancer classification [3]. In binary classification tasks, which present relatively low complexity, multiple models achieved excellent performance, with CNN-based models (ResNet50, RegNet, ConvNeXT) and the transformer-based foundation model UNI all reaching an AUC of 0.999 [3].

The expressiveness advantage of foundation models becomes particularly evident in more complex diagnostic scenarios. In eight-class classification tasks with increased complexity, performance differences among architectures became more pronounced. The best-performing model was the fine-tuned foundation model UNI, which attained an accuracy of 95.5% (95% CI: 94.4–96.6%), a specificity of 95.6% (95% CI: 94.2–96.9%), an F1-score of 95.0% (95% CI: 93.9–96.1%), and an AUC of 0.998 (95% CI: 0.997–0.999) [3]. This superior performance in complex multi-class scenarios demonstrates how foundation models leverage their expansive pre-training to maintain discriminative power across fine-grained diagnostic categories.

Table 2: Performance Comparison on Breast Cancer Classification (BreakHis v1 Dataset)

Model Type	Model Name	Binary Classification AUC	Eight-Class Classification Accuracy	Eight-Class F1-Score
CNN-Based	ConvNeXT	0.999	Not reported	Not reported
CNN-Based	ResNet50	0.999	Not reported	Not reported
CNN-Based	RegNet	0.999	Not reported	Not reported
Foundation Model	UNI (fine-tuned)	0.999	95.5%	95.0%
Foundation Model	UNI (zero-shot)	Poor performance	Poor performance	Poor performance

Beyond Classification: Advanced Predictive Capabilities

The scalability of foundation models enables capabilities that extend far beyond traditional classification tasks. The HE2RNA model demonstrates how deep learning can predict RNA-Seq expression profiles from H&E-stained whole slide images alone, creating a bridge between histological morphology and molecular profiling [23]. Through a multitask weakly supervised approach trained on matched WSIs and RNA-Seq data from TCGA, HE2RNA learned to predict expression levels for thousands of genes with statistically significant correlation to ground truth measurements [23].

This capability to map histological patterns to molecular phenotypes represents a quantum leap beyond traditional CNN applications. For instance, HE2RNA accurately predicted expression of immune-related genes (C1QB, NKG7, ARHGAP9) across multiple cancer types and identified pathway-level activities including angiogenesis, hypoxia, DNA repair, and immune responses [23]. Similarly, a 2021 study demonstrated that attention-based multiple instance learning could predict gene expression from H&E-stained tissues with sufficient accuracy to discriminate fulminant-like pulmonary tuberculosis in murine models, achieving sensitivity and specificity of 0.88 and 0.95 respectively [24].

Experimental Protocols and Methodologies

Foundation Model Pre-training Protocol

The development of pathology foundation models follows a rigorous multi-stage protocol centered on self-supervised learning at scale. The pre-training phase typically leverages massive unlabeled datasets comprising hundreds of thousands to millions of whole slide images from diverse tissue types and disease states [3] [22].

Data Curation and Preprocessing: Whole slide images are partitioned into smaller patches (typically 256×256 pixels) at multiple magnification levels. UNI, for instance, was trained on more than 100 million image tiles extracted from over 100,000 diagnostic-grade H&E-stained whole slide images across 20 major tissue types [3]. Quality control procedures remove artifacts, blurry regions, and non-tissue areas.

Self-Supervised Learning Objective: Models learn through pretext tasks that don't require manual annotations. Common approaches include masked image modeling (where the model learns to predict randomly masked portions of the image) and contrastive learning (where the model learns to identify different augmentations of the same image). Prov-GigaPath employed the novel GigaPath architecture incorporating LongNet to handle giga-pixel scale context [3].

Training Infrastructure and Scale: Training occurs on specialized hardware, typically GPU clusters, for weeks or months. The exponential relationship between model size, data quantity, and computational requirements necessitates substantial infrastructure investment. Companies increasingly invest in on-premises training infrastructure, trading flexibility for predictable architecture and availability [21].

Transfer Learning and Fine-tuning Protocol

The true power of foundation models emerges through adaptation to specific downstream tasks. The fine-tuning protocol enables researchers to leverage pre-trained representations for specialized applications with limited labeled data.

Task-Specific Data Preparation: Depending on the target application, researchers curate labeled datasets typically ranging from hundreds to thousands of annotated examples. For classification tasks, slide-level or region-level labels are prepared; for segmentation tasks, pixel-level annotations are required.

Model Adaptation: The pre-trained foundation model serves as a feature extractor, with the final layers modified or replaced to suit the specific task. During fine-tuning, all or most model weights are updated using task-specific data. The learning rate is typically set lower than during pre-training to avoid catastrophic forgetting of general representations.

Performance Validation: Models are evaluated using standard metrics appropriate to the task (accuracy, AUC, F1-score for classification; Dice coefficient for segmentation) with rigorous cross-validation or hold-out testing. Clinical validation often involves multiple independent datasets to assess generalizability across institutions and staining protocols.

Molecular Prediction Protocol

The prediction of molecular features from histology represents one of the most advanced applications of foundation models in pathology. The HE2RNA protocol exemplifies this approach [23]:

Multi-modal Data Alignment: Whole slide images are aligned with matched molecular data (RNA-Seq, protein expression, genetic mutations) from the same samples. The TCGA database frequently serves as this data source, providing paired histology and molecular profiling.

Weakly-Supervised Training: Models learn to predict molecular features using only slide-level labels without regional annotations. The HE2RNA model employs a multitask approach where each task corresponds to predicting the expression level of a specific gene [23].

Spatial Expression Mapping: Through interpretable design, these models can generate virtual spatialization of gene expression, creating heatmaps that localize molecular activity to specific tissue regions. This spatial prediction is validated through comparison with immunohistochemistry staining on independent datasets [23].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Research Reagents for Pathology Foundation Model Research

Reagent / Resource	Function and Application	Example Specifications
Whole Slide Images (WSIs)	Digital representations of histopathology slides for model training and validation	H&E-stained, 40x magnification, gigapixel resolution [22]
Annotation Software	Tools for labeling regions of interest, cell types, and pathological structures	Digital pathology platforms with collaborative annotation features
Computational Infrastructure	Hardware for training and deploying large-scale models	High-performance GPU clusters with specialized memory [21]
Molecular Datasets	Paired genomic, transcriptomic, or proteomic data for multi-modal learning	RNA-Seq, mutation calls, protein expression data [23]
Benchmark Datasets	Standardized datasets for model evaluation and comparison	Publicly available datasets like BreakHis, TCGA [3] [23]
Experiment Tracking Systems	Software for managing training runs, hyperparameters, and results	Specialized MLops platforms (e.g., Neptune.ai) [21]

Implementation Challenges and Considerations

Data Scalability and Infrastructure Requirements

The scalability advantages of foundation models come with substantial computational costs that present implementation challenges. Training foundation models requires specialized hardware infrastructure, virtually always utilizing GPUs with extensive memory and processing capabilities [21]. The scale of data processing is immense—Prov-GigaPath processed 1.3 billion image patches, requiring sophisticated data pipelines and distributed training approaches [3]. Organizations must weigh investments in on-premises infrastructure against cloud-based solutions, balancing computational demands against data privacy and compliance requirements, particularly in healthcare settings with strict regulatory frameworks [21] [25].

Domain Adaptation and Generalization

While foundation models demonstrate remarkable generalization capabilities, their application to specific pathology domains requires careful adaptation. Studies consistently show that foundation model encoders used without fine-tuning produce generally poor performance on specialized classification tasks [3]. The zero-shot capabilities that appear in natural language foundation models are less pronounced in computational pathology, necessitating targeted fine-tuning with domain-specific data. This adaptation process requires both technical expertise in machine learning and domain knowledge in pathology to ensure models learn clinically relevant features and maintain diagnostic accuracy across tissue types, staining protocols, and scanner variations.

Future Directions and Strategic Implications

The evolution toward foundation models in computational pathology represents more than a technical improvement—it constitutes a fundamental shift in how pathological analysis is conceptualized and implemented. The scalability of these models enables continuous improvement with additional data and computational resources, creating virtuous cycles of capability enhancement [21]. For pharmaceutical development, this progression offers opportunities to identify novel biomarkers, predict treatment response, and accelerate therapeutic discovery through more sophisticated analysis of histological patterns [25].

The strategic implications for healthcare institutions and drug development companies are substantial. Successful foundation model strategies are characterized by early proof-of-concept projects, building deep expertise across all aspects of training, application-based performance evaluation, and maintaining focus on core objectives [21]. As these technologies mature, they promise to transform pathology from a primarily descriptive discipline to a quantitative, predictive science capable of extracting profound insights from the morphological patterns that underlie disease processes.

The trajectory is clear: while traditional CNNs will continue to serve specific, limited-scope applications, foundation models represent the future of computational pathology—more expressive, more scalable, and ultimately more capable of capturing the extraordinary complexity of human disease through histopathological analysis.

Harnessing Foundation Models and CNNs for Real-World Pathology Tasks

The analysis of whole-slide images (WSIs) presents a unique computational challenge due to their gigapixel size and the complex, context-dependent nature of pathological diagnosis. Traditional convolutional neural networks (CNNs) have provided a foundation for automated analysis but face significant limitations in handling WSIs' extreme resolution and weak supervision requirements. The emergence of foundation models (FMs) pretrained on massive datasets represents a paradigm shift, offering more powerful and transferable feature representations. This whitepaper examines the critical technical evolution from traditional CNNs to foundation models within multiple instance learning (MIL) frameworks for computational pathology. We provide a comprehensive analysis of how this integration enhances diagnostic accuracy, robustness, and generalization across diverse clinical scenarios, supported by quantitative benchmarks, implementation protocols, and practical research frameworks.

Foundation Models vs. Traditional CNNs: A Technical Paradigm Shift

Fundamental Architectural and Training Differences

The transition from traditional CNNs to foundation models in pathology represents more than incremental improvement—it constitutes a fundamental shift in approach. Traditional CNNs such as ResNet and VGG, typically pretrained on natural image datasets like ImageNet, operate as feature extractors that capture general texture patterns but lack domain-specific morphological understanding [19]. These models struggle with the exceptional complexity of tissue morphology, where diagnostic interpretation depends on multi-scale contextual relationships that natural image-trained models fail to capture adequately [19].

Foundation models address these limitations through self-supervised learning on massive histopathology-specific datasets. Unlike CNNs trained with supervised learning on limited annotated data, FMs leverage self-supervised objectives such as masked image modeling and contrastive learning, pretrained on millions of histology image patches [6] [26]. This enables learning of rich, transferable representations of tissue microstructure without reliance on scarce manual annotations. Architecturally, while CNNs process individual patches in isolation, newer FMs like TITAN employ Vision Transformers (ViTs) that can capture long-range dependencies across entire WSIs by processing sequences of patch embeddings in a spatially-aware manner [6].

Performance and Generalization Benchmarks

Table 1: Quantitative Performance Comparison Between CNN and Foundation Models

Model Category	Representative Models	AUROC Range	Key Strengths	Major Limitations
Traditional CNNs	ResNet, VGG	0.916-0.909 [27]	Computational efficiency, architectural simplicity	Limited domain specificity, poor cross-site generalization
Pathology Foundation Models	CONCH, Virchow, UNI, TITAN	0.984-0.992 [27] [6]	Superior transfer learning, multimodal capabilities, site robustness	Computational intensity, training instability, security vulnerabilities [19]

Empirical evaluations demonstrate the significant performance advantage of foundation models over traditional approaches. The PEAN system, which incorporates pathologists' visual attention patterns, achieved an accuracy of 96.3% and AUC of 0.992 on internal testing, outperforming CNN-based models by substantial margins [27]. Similarly, the TITAN foundation model outperformed supervised baselines and existing multimodal slide foundation models across diverse tasks including cancer subtyping, biomarker prediction, and outcome prognosis [6].

Critical for clinical deployment, foundation models exhibit enhanced robustness across healthcare institutions. While traditional CNNs often suffer from performance degradation due to site-specific biases, uncertainty-aware FM ensembles like PICTURE maintained diagnostic accuracy (AUROC 0.924-0.996) across five independent international cohorts [26]. However, systematic evaluations reveal that despite their advantages, pathology FMs still exhibit fundamental weaknesses including low absolute accuracy in some contexts (F1 scores ~40-42% in zero-shot retrieval), geometric instability, and concerning security vulnerabilities to adversarial attacks [19].

MIL Frameworks: The Bridge Between FMs and WSI Classification

The MIL Paradigm in Computational Pathology

Multiple instance learning provides the essential computational framework for managing the extreme dimensionality of WSIs by treating each slide as a "bag" containing thousands to millions of individual patch "instances." Standard MIL approaches aggregate patch-level predictions to generate slide-level diagnoses while identifying diagnostically relevant regions [28]. Traditional attention-based MIL (ABMIL) frameworks process patch embeddings independently, overlooking critical spatial relationships between neighboring tissue regions [29].

Recent advances have addressed this limitation through spatially-aware architectures. The GABMIL framework explicitly captures inter-instance dependencies while maintaining computational efficiency, achieving up to 7 percentage point improvement in AUPRC over standard ABMIL [29]. Similarly, SMMILe leverages instance-based MIL to achieve superior spatial quantification without compromising WSI classification performance, demonstrating that explicit modeling of patch relationships is essential for accurate morphological interpretation [28].

Integrating Foundation Models with MIL Frameworks

The integration of FMs with MIL frameworks creates a powerful synergy for WSI analysis. FMs provide superior patch embeddings that capture rich morphological features, while MIL frameworks enable effective aggregation of these features for slide-level prediction. This combination has proven particularly effective in challenging diagnostic scenarios. For instance, the PICTURE system integrates nine different pathology foundation models within an ensemble MIL framework to differentiate glioblastoma from primary central nervous system lymphoma, achieving an AUROC of 0.989 with validation across five independent cohorts [26].

Table 2: Performance of Integrated FM-MIL Frameworks Across Cancer Types

Framework	Architecture	Datasets	Key Results	Clinical Application
SMMILe [28]	Superpatch-based measurable MIL	6 cancer types, 3,850 WSIs	Matches/exceeds SOTA classification with outstanding spatial quantification	Metastasis detection, subtype prediction, grading
AttriMIL [30]	Attribute-aware MIL with multi-branch scoring	5 public datasets	Superior bag classification and disease localization	Differentiating subtle tissue variations
PICTURE [26]	Uncertainty-aware FM ensemble	2,141 CNS slides	AUROC 0.989, validated across 5 cohorts	Differentiating glioblastoma from mimics
TITAN [6]	Multimodal vision-language FM	335,645 WSIs	Superior few-shot and zero-shot classification	Rare disease retrieval, cancer prognosis

The AttriMIL framework demonstrates how MIL can be enhanced through attribute-aware mechanisms that quantify pathological attributes of individual instances, establishing region-wise and slide-wise constraints to model instance correlations during training [30]. This approach captures intrinsic spatial patterns and semantic similarities between patches, enhancing sensitivity to challenging instances and subtle tissue variations that are critical for accurate diagnosis.

Experimental Protocols and Implementation Frameworks

WSI Preprocessing and Feature Extraction

A standardized preprocessing pipeline is essential for reproducible WSI analysis. The following protocol, synthesized from multiple studies, ensures consistent input data quality:

• Whole-Slide Scanning: Utilize FDA-approved WSI scanners (e.g., Philips Ultra-Fast Scanners) with quality control procedures to maintain consistent sharpness and color fidelity [31].
• Tissue Segmentation: Apply automated algorithms to detect and segment tissue regions, excluding background areas. The PICTURE system identifies patches containing mostly blank backgrounds using color and cell density profiles [26].
• Patch Extraction: Tessellate WSIs into non-overlapping patches, typically 256×256 or 512×512 pixels at 20× magnification. Studies indicate that using larger patches (512×512) reduces sequence length without compromising morphological information [6].
• Color Normalization: Standardize stain variations using statistical methods (e.g., based on mean and standard deviation of pixel intensities) or deep learning-based normalization to mitigate site-specific batch effects [26].
• Feature Extraction: Process patches through frozen foundation model encoders to generate feature embeddings. The TITAN model uses CONCHv1.5 to extract 768-dimensional features for each 512×512 patch [6].

MIL Model Training and Optimization

Following feature extraction, implement MIL training with the following considerations:

• Feature Grid Construction: Spatially arrange patch embeddings into a 2D grid replicating their original positions in the tissue [6]. This preserves spatial context essential for morphological assessment.
• Data Augmentation: Apply feature-space augmentations including vertical and horizontal flipping, and posterization to enhance robustness [6]. The iBOT framework uses global and local cropping from feature grids for self-supervised pretraining [6].
• Spatial Context Integration: Implement spatially-aware MIL aggregation. GABMIL captures inter-instance dependencies through interaction-aware representations, while TransMIL employs transformer architectures to model long-range dependencies [29].
• Uncertainty Quantification: Integrate epistemic uncertainty measures using Bayesian inference, deep ensembles, and normalizing flows to identify out-of-distribution samples and enhance model reliability [26].

Evaluation Metrics and Validation

Comprehensive model assessment should include:

• Slide-Level Classification: Area Under Receiver Operating Characteristic Curve (AUROC), accuracy, F1-score, and balanced accuracy across multiple testing cohorts [26].
• Spatial Quantification: Patch-level localization accuracy measured against pathologists' annotations or eye-tracking data [27] [28].
• Robustness Evaluation: Performance consistency across multiple institutions, scanners, and staining protocols, quantified through Robustness Index (RI) [19].
• Clinical Utility: Diagnostic concordance with expert pathologists, time savings, and performance on rare or challenging cases [27].

The Scientist's Toolkit: Essential Research Reagents

Table 3: Essential Research Tools for FM-MIL Integration

Resource Category	Specific Tools	Function	Key Features
Foundation Models	CONCH, Virchow, UNI, CTransPath, Phikon [26]	Patch feature extraction	Self-supervised learning on pathology-specific datasets
MIL Frameworks	ABMIL, TransMIL, GABMIL, SMMILe, AttriMIL [28] [30] [29]	Slide-level prediction from patches	Spatial context modeling, attention mechanisms
Whole-Slide Datasets	TCGA, Camelyon16, in-house collections [28] [26]	Model training and validation	Multi-center, multi-cancer, paired clinical data
Computational Tools	PyTorch, VOSviewer, Whole-Slide Processing Libraries [32]	Implementation and analysis	Support for large-scale WSI processing

The integration of foundation models with multiple instance learning frameworks represents a significant advancement in computational pathology, enabling more accurate, robust, and interpretable analysis of whole-slide images. This technical synergy addresses critical limitations of traditional CNN-based approaches by combining domain-specific pretraining with spatially-aware aggregation mechanisms. Despite persistent challenges including computational demands, security vulnerabilities, and validation requirements, the FM-MIL paradigm shows tremendous promise for clinical translation. Future developments will likely focus on multimodal integration, federated learning approaches to enhance data privacy, and specialized architectures designed specifically for the hierarchical organization of tissue morphology. As these technologies mature, they hold the potential to transform pathological diagnosis, biomarker discovery, and personalized treatment planning in oncology and beyond.

The rapidly emerging field of computational pathology has demonstrated tremendous promise in developing objective prognostic models from histology images, yet most approaches remain limited by their unimodal focus [33]. Traditional diagnostic workflows in pathology integrate morphological assessment with molecular profiling and clinical data, creating a pressing need for computational frameworks that can similarly fuse these heterogeneous data streams. Multimodal integration represents a transformative approach that simultaneously examines pathology whole slide images (WSIs) and molecular profile data to predict patient outcomes and discover prognostic biomarkers that would remain invisible to unimodal analysis [33].

Within this paradigm, a fundamental shift is occurring from traditional Convolutional Neural Networks (CNNs) to pathology foundation models pretrained using self-supervised learning on massive datasets. This technical evolution enables more robust feature extraction and dramatically improves performance across diverse downstream tasks, particularly when integrated with multimodal data sources [15]. The capacity to align histopathological patterns with genomic alterations and clinical reports represents a critical advancement toward precision oncology, offering the potential to identify novel biomarkers and improve patient risk stratification beyond the capabilities of single-modality analysis [33] [34].

Foundation Models vs. Traditional CNNs: Core Architectural Differences

Fundamental Technical Distinctions

Pathology foundation models fundamentally differ from traditional CNNs in their architecture, training methodology, and data requirements. CNNs extract spatial patterns using small convolutional kernels across multiple layers and possess strong inductive biases, which enables high performance with limited datasets but prevents them from fully leveraging large-scale data [15]. In contrast, Vision Transformers (ViTs) utilized in foundation models employ self-attention mechanisms with minimal inductive biases, allowing them to outperform CNNs when trained on extensive pathology image datasets [15].

The training approaches further differentiate these architectures. Traditional CNNs typically utilize supervised learning with ImageNet initialization, requiring extensive labeled data for effective training. Pathology foundation models employ self-supervised learning (SSL) pretrained on massive unlabeled datasets comprising millions of pathology image patches, learning generalized representations that transfer effectively to various diagnostic tasks with minimal fine-tuning [15]. This fundamental difference in training paradigm enables foundation models to develop a more comprehensive understanding of histopathological structures and their variations.

Performance Implications in Pathology Tasks

Recent comparative studies demonstrate the practical implications of these architectural differences. In kidney pathology classification tasks, all foundation models (UNI, UNI2-h, Prov-Gigapath, Phikon, Virchow) consistently outperformed ImageNet-pretrained ResNet50, achieving area under the receiver operating characteristic curve (AUROC) exceeding 0.980 in internal validation [15]. More significantly, in external validation, ResNet50 performance markedly dropped while foundation models maintained robust performance, demonstrating superior generalizability across institutions with different staining protocols and scanning methods [15].

Foundation models also excel in recognizing diagnostically relevant structures without extensive manual annotation. Visualization of attention heatmaps confirmed that foundation models accurately identified morphologically significant regions in kidney pathology, including glomerular and tubular structures relevant to disease classification [15]. This capability for unsupervised biomarker discovery represents a crucial advancement over CNN-based approaches that typically require detailed region-of-interest annotations for comparable performance.

Table 1: Comparative Performance of Foundation Models vs. CNN in Kidney Pathology Classification

Model Architecture	Pretraining Data	Internal Validation AUROC	External Validation AUROC	Generalizability
ResNet50 (CNN)	ImageNet	0.950	Significant performance drop	Limited
UNI	Pathology SSL	>0.980	Maintained high performance	Excellent
Phikon	Pathology SSL	>0.980	Maintained high performance	Excellent
Virchow	Pathology SSL	>0.980	Maintained high performance	Excellent

Multimodal Integration Frameworks and Methodologies

Deep Learning-based Multimodal Fusion Architectures

Multimodal fusion architectures represent the computational core of integrated histology-genomic analysis. The deep learning-based Multimodal Fusion (MMF) algorithm utilizes both H&E whole slide images and molecular profile features (mutation status, copy number variation, RNA-Seq expression) to measure and explain relative risk of cancer death [33]. This approach employs weakly-supervised learning to handle the massive data size of WSIs, treating each slide as a collection of patches (instances) with only slide-level labels available rather than detailed patch-level annotations [33] [15].

Multiple instance learning (MIL) provides an effective framework for slide-level classification by aggregating information from individual patches without requiring patch-level annotations [15]. Within heterogeneous tissue slides where diagnostic value varies widely among patches, MIL excels by learning to focus on clinically relevant patches through several aggregation methods:

• Attention-based MIL (ABMIL): Uses an attention mechanism to perform weighted aggregation of patch features, with weights trained by a neural network [15]
• Transformer-based MIL (TransMIL): Utilizes transformer mechanisms to learn spatial relationships between patches via self-attention [15]
• Clustering-constraint attention MIL (CLAM): Incorporates instance-level clustering into the attention mechanism to enhance feature representation learning, particularly effective through its multi-branch variant (CLAM-MB) that computes class-specific attention weights [15]

Table 2: Multiple Instance Learning Aggregation Methods for Whole Slide Image Analysis

Aggregation Method	Mechanism	Advantages	Application Context
Max Pooling	Selects most indicative patch	Computational simplicity	Limited performance for complex morphology
Attention-Based MIL (ABMIL)	Weighted aggregation via learned attention	Adaptively focuses on relevant regions	General whole slide classification
Transformer-Based MIL (TransMIL)	Self-attention between patches	Captures spatial relationships	Tasks requiring structural context
CLAM-MB	Class-specific clustering constraints	Enhanced feature separation	Multi-class classification problems

Experimental Protocols for Multimodal Integration

Implementing robust multimodal integration requires systematic experimental protocols spanning data collection, preprocessing, feature extraction, and model validation. For radiology-pathology-genomics integration in non-small cell lung cancer (NSCLC) immunotherapy response prediction, researchers developed a comprehensive workflow [34]:

Data Acquisition and Curation:

• Assemble paired datasets including CT scans, digitized PD-L1 immunohistochemistry slides, and genomic features from clinical sequencing platforms
• Apply rigorous quality control: exclude specimens with staining artifacts, damaged tissue, or uncertain histology
• Implement expert segmentation of radiological lesions by board-certified radiologists for consistent feature extraction

Feature Extraction Pipeline:

• Whole Slide Image Processing: Divide WSIs into non-overlapping tiles at 20x magnification (256px, 128μm), remove background using Otsu's thresholding and Gaussian blur filtering
• Molecular Feature Processing: Encode mutation status, copy number variations, and RNA-Seq expression values using standardized normalization
• Radiomics Feature Extraction: Calculate texture features from segmented lesions, augmented by superpixel-based perturbations to ensure robustness

Multimodal Model Training:

• Implement cross-validation schemes that maintain patient-level separation to prevent data leakage
• Utilize dynamic deep attention-based multiple instance learning with masking (DyAM) for adaptive feature weighting across modalities
• Apply class-balancing techniques to address real-world response rate imbalances (approximately 25% responders in NSCLC immunotherapy cohorts)

Quantitative Outcomes and Validation Frameworks

Performance Benchmarks Across Cancer Types

Comprehensive validation across multiple cancer types demonstrates the superior performance of multimodal integration compared to unimodal approaches. In a pan-cancer analysis encompassing 6,592 gigapixel WSIs from 5,720 patients across 14 cancer types, multimodal fusion achieved an overall concordance index (c-Index) of 0.645 for survival prediction, outperforming unimodal models using only histology (c-Index = 0.585) or molecular features alone (c-Index = 0.607) [33].

The advantage of multimodal integration was particularly evident in specific cancer types. For NSCLC immunotherapy response prediction, the multimodal model integrating CT imaging, histology, and genomic features achieved an AUC of 0.80 (95% CI 0.74-0.86), significantly outperforming standard biomarkers including tumor mutational burden (AUC = 0.61) and PD-L1 immunohistochemistry scoring (AUC = 0.73) [34]. This demonstrates that multimodal integration provides more accurate prediction of clinical endpoints than Food and Drug Administration-approved biomarkers currently used in clinical decision-making.

Validation of Histopathologic Scoring Systems

Valid histopathologic scoring remains fundamental to multimodal integration, requiring rigorous methodology to ensure data quality [35]. Key validation principles include:

Scoring System Development:

• Define clear scoring criteria with specific terminology and percentage ranges for each category
• Avoid vague terms like "mild," "moderate," or "severe" without quantitative boundaries
• Establish scoring definitions that characterize specific lesion parameters relevant to the research objectives

Validation Measures:

• Repeatability Validation: Assess intra-observer and inter-observer consistency through blinded rescoring studies
• Biological Validation: Ensure scoring systems accurately reflect underlying pathobiology through correlation with clinical outcomes or molecular features

Digital image analysis provides significant advantages for scoring reproducibility. In prostate cancer studies of estrogen receptor β2 immunohistochemistry, digital methods demonstrated near-perfect reproducibility (Spearman correlation = 0.99) compared to pathologist visual scoring (Spearman correlation = 0.84) [36]. This enhanced reproducibility is particularly valuable for large-scale studies where manual scoring consistency becomes challenging.

Implementation Workflow and Visualization

End-to-End Multimodal Integration Pipeline

The following workflow diagram illustrates the complete process for aligning histology with genomic data and clinical reports:

Architectural Comparison: Foundation Models vs CNNs

The following diagram contrasts the fundamental differences between pathology foundation models and traditional CNNs for feature extraction:

Table 3: Essential Research Resources for Multimodal Integration Studies

Resource Category	Specific Tools/Platforms	Function in Multimodal Research
Pathology Foundation Models	UNI, UNI2-h, Prov-Gigapath, Phikon, Virchow [15]	Patch-level feature extraction from WSIs without stain normalization
Multiple Instance Learning Frameworks	CLAM, ABMIL, TransMIL [15]	Slide-level classification using attention mechanisms
Whole Slide Image Processing	Slideflow, Libvips, OpenSlide [15]	WSI tiling, background removal, and data augmentation
Molecular Data Platforms	MSK-IMPACT, RNA-Seq pipelines [34]	Genomic feature extraction including mutations, CNV, expression
Multimodal Integration Platforms	PORPOISE [33]	Interactive visualization of model explanations and biomarker discovery
Digital Pathology Analysis	Aperio Image Analysis, TissueFAXS [36]	Automated IHC quantification and tissue segmentation

Multimodal integration of histology, genomic data, and clinical reports represents a paradigm shift in computational pathology, enabling more accurate prognostic models and discovery of novel biomarkers. The transition from traditional CNNs to pathology foundation models addresses critical limitations in generalizability and annotation dependency, particularly when combined with multimodal data streams through multiple instance learning approaches.

Future developments will likely focus on refining cross-modal alignment techniques, improving model interpretability for clinical translation, and establishing standardized validation frameworks across institutions. As these technologies mature, multimodal integration promises to bridge the gap between histopathological assessment and molecular profiling, ultimately enhancing precision oncology through more comprehensive patient stratification and biomarker discovery.

The emergence of foundation models (FMs) represents a paradigm shift in computational pathology, moving beyond traditional convolutional neural networks (CNNs) toward large-scale, versatile systems pretrained on massive datasets. These models, typically based on Vision Transformer (ViT) architectures and trained via self-supervised learning (SSL) on millions of histopathology image tiles, demonstrate remarkable generalization capabilities across diverse diagnostic tasks [37] [38]. However, this shift introduces a critical challenge: how to best adapt these powerful but generic models to specific clinical applications. The adaptation strategy itself—whether through comprehensive fine-tuning or simpler linear probing—has emerged as a decisive factor influencing model performance, computational efficiency, and clinical viability.

In pathology, FMs differ fundamentally from traditional CNNs in their scale, pretraining methodology, and intended versatility. While CNNs are typically trained from scratch or with ImageNet initialization on specific, labeled datasets for narrow tasks like cancer classification or segmentation, FMs are pretrained on enormous unlabeled histopathology datasets using SSL objectives, learning universal visual representations of tissue morphology [37] [39]. This foundational training enables them to capture intricate patterns across tissues, stains, and pathologies, but realizing this potential requires careful adaptation. The choice between fine-tuning and linear probing represents a fundamental trade-off between leveraging learned representations and adapting to specific domains, with significant implications for performance, robustness, and clinical deployment.

Fundamental Differences: Foundation Models vs. Traditional CNNs in Pathology

Table 1: Core distinctions between foundation models and traditional CNNs in pathology

Characteristic	Traditional CNNs	Pathology Foundation Models
Architecture	Convolutional layers (ResNet, VGG)	Vision Transformers (ViT), hybrid CNN-Transformers [37]
Scale	Millions of parameters	Hundreds of millions to billions of parameters (e.g., UNI: 303M-1.5B, Virchow2: 632M) [37]
Pretraining Data	ImageNet (natural images) or pathology-specific datasets	Massive histopathology datasets (e.g., 100K-1.5M whole slide images) [40] [37]
Pretraining Method	Supervised learning on labeled data	Self-supervised learning (DINOv2, MAE, iBOT) on unlabeled images [37] [39]
Primary Strength	Excellent performance on specific, narrow tasks	Generalizability across diverse organs, tasks, and institutions [38] [39]
Adaptation Approach	Often used as-is or with full fine-tuning	Linear probing, parameter-efficient fine-tuning (PEFT/LoRA) [40] [37]

The architectural and methodological evolution from CNNs to FMs in pathology represents more than incremental improvement—it constitutes a fundamental reimagining of how AI systems learn histopathological representations. Traditional CNNs excel at local feature extraction through their inductive bias for spatial hierarchies, making them effective for specific tasks like nuclear segmentation or tumor detection [41]. However, their locality and task-specific nature limit their ability to capture the complex, long-range dependencies and contextual relationships inherent in tissue architecture.

Pathology FMs address these limitations through transformer-based architectures that process images as sequences of patches, enabling global attention across entire tissue regions [37]. More significantly, their self-supervised pretraining on massive, diverse histopathology datasets allows them to learn a comprehensive "visual vocabulary" of tissue morphology, stain variations, and pathological patterns without human labeling bottlenecks [39]. This foundational knowledge enables unprecedented transfer capabilities but introduces the critical challenge of adaptation strategy selection—a decision with profound implications for model behavior, performance, and clinical utility.

Adaptation Strategies: Methodological Framework

Linear Probing: Frozen Feature Extraction

Linear probing represents the most constrained adaptation approach, where the entire FM backbone remains frozen, and only a simple linear classifier (typically a single fully connected layer) is trained on top of the extracted features [37]. This approach treats the FM as a fixed feature extractor, leveraging the representations learned during pretraining without modifying them.

The methodological workflow for linear probing involves:

• Feature Extraction: Processing input whole slide images (WSIs) divided into tiles through the frozen FM to generate embedding vectors for each tile
• Feature Aggregation: Combining tile-level embeddings into slide-level representations using methods like mean pooling or attention-based multiple instance learning (ABMIL)
• Classifier Training: Training only the final linear layer on the aggregated features using task-specific labels [37]

Full Fine-Tuning: Comprehensive Parameter Update

Full fine-tuning represents the opposite extreme, where all parameters of the FM are updated during training on the target task. This approach allows the model to substantially adjust its representations to the specific domain and task but requires significant computational resources and extensive labeled data to avoid catastrophic forgetting or overfitting [40].

The fine-tuning methodology involves:

• Parameter Initialization: Loading the pretrained FM weights
• End-to-End Training: Updating all model parameters using backpropagation with task-specific loss functions
• Careful Regularization: Applying strong regularization techniques to prevent overfitting to limited medical datasets [3]

Parameter-Efficient Fine-Tuning (PEFT): Balanced Approach

Parameter-efficient fine-tuning strategies, such as Low-Rank Adaptation (LoRA), offer a middle ground by introducing small, trainable adapter modules while keeping the majority of the pretrained weights frozen [37]. This approach balances adaptation capacity with computational efficiency, making it particularly suitable for medical domains with moderate data availability.

Diagram 1: Methodological workflow for FM adaptation strategies in pathology (Width: 760px)

Experimental Comparison: Quantitative Performance Analysis

Table 2: Performance comparison of adaptation strategies across pathology tasks

Adaptation Strategy	Data Efficiency	Computational Cost	Typical Performance	Robustness to Domain Shift	Clinical Deployment Suitability
Linear Probing	High (few-shot capable)	Low (0.1-0.5×)	Moderate to high (AUC: 0.90-0.98) [39]	High (generalizes well across institutions) [37]	Excellent (stable, interpretable)
Parameter-Efficient Fine-Tuning (PEFT)	Medium (100+ samples/class)	Medium (0.3-0.7×)	High (AUC: 0.95-0.99) [37]	Medium to high	Very good (balanced approach)
Full Fine-Tuning	Low (requires large datasets)	High (1.0-3.0×)	Variable (can be highest with sufficient data)	Low to medium (risk of overfitting) [40]	Poor (computationally expensive, unstable)

Empirical evidence consistently demonstrates that linear probing achieves optimal performance in data-scarce scenarios and for cross-institutional generalization, while PEFT provides the best balance for moderate data regimes. A comprehensive clinical benchmark evaluating public pathology FMs on disease detection and biomarker prediction tasks found that linear probing achieved AUCs above 0.9 across all tasks while maintaining robustness across institutions [39]. Similarly, studies have shown that for few-shot tasks (<5 labels per class), linear probing or KNN classification on frozen features outperforms more complex adaptation methods [37].

The superiority of linear probing for many clinical applications stems from its stability and preservation of the robust features learned during large-scale pretraining. As noted by Tizhoosh (2025), most pathology FMs are "too large, memory-intensive, and unstable to fine-tune on moderate-sized data sets typical of clinical research," leading to frequent performance degradation during full fine-tuning due to overfitting and catastrophic forgetting [40]. This dependency on linear probing represents a significant divergence from the original FM paradigm, which promised easy fine-tuning adaptation, but has proven pragmatically necessary in pathology applications.

Table 3: Key research reagents and computational resources for FM adaptation experiments

Resource Category	Specific Examples	Function in FM Adaptation
Public Foundation Models	UNI, Virchow2, Phikon, CTransPath, PLUTO-4G [37] [39]	Pretrained backbones for feature extraction and adaptation
Benchmark Datasets	TCGA, BCSS, Camelyon, in-house clinical cohorts [39] [3]	Evaluation of adaptation strategies on diverse tasks and domains
Computational Frameworks	PyTorch, MONAI, TIAToolbox, custom benchmarking pipelines [39]	Infrastructure for model training, evaluation, and deployment
Adaptation Algorithms	Linear classifiers, LoRA, adapter modules, attention-based pooling [37]	Implementation of specific adaptation strategies
Evaluation Metrics	AUC, F1 score, Robustness Index (RI), DICE coefficient [40] [37]	Quantitative assessment of performance and generalization

The experimental toolkit for FM adaptation requires both computational resources and methodological components. Publicly available FMs like UNI and Virchow2 serve as essential starting points, providing robust pretrained backbones that have learned comprehensive histopathological representations from diverse datasets [37]. Benchmarking pipelines, such as the automated clinical benchmark described by Nature Communications, enable standardized evaluation across multiple institutions and tasks, facilitating meaningful comparisons between adaptation strategies [39].

Computationally, adaptation experiments require significant resources, particularly for full fine-tuning. Studies indicate that FMs can consume up to 35× more energy than task-specific models, raising sustainability concerns for clinical deployment [40]. This substantial computational footprint necessitates careful strategy selection, with linear probing and PEFT offering more environmentally sustainable alternatives for many applications.

Decision Framework and Clinical Implementation Guidelines

The choice between adaptation strategies should be guided by specific clinical requirements, data availability, and computational constraints. The following decision framework provides practical guidance for researchers and clinicians:

Diagram 2: Decision framework for selecting FM adaptation strategies (Width: 760px)

For clinical implementation, several best practices emerge from recent research:

• Prioritize Linear Probing for Initial Deployment: Begin with linear probing to establish a robust baseline, as it provides excellent generalization with minimal computational overhead and reduced overfitting risk [37] [39].

• Employ PEFT for Performance Optimization: When linear probing proves insufficient and adequate data exists, implement parameter-efficient methods like LoRA to enhance task-specific performance without the instability of full fine-tuning [37].

• Validate Across Multiple Institutions: Regardless of strategy, conduct rigorous external validation using slides from different hospitals and scanner types to assess real-world robustness, as site-specific bias remains a critical challenge for pathology FMs [40] [39].

• Monitor for Domain Shift: Continuously evaluate model performance on new data, as staining variations, scanner changes, and evolving protocols can degrade performance over time, necessitating strategy reassessment [40].

The adaptation strategy for pathology foundation models represents a critical determinant of their clinical utility, balancing performance, efficiency, and robustness. While full fine-tuning aligns with the theoretical promise of FMs as adaptable foundations, practical constraints in pathology—including limited annotated data, computational costs, and institutional variability—have established linear probing as the predominant approach for clinical deployment. Parameter-efficient fine-tuning emerges as a promising middle ground, offering enhanced adaptation capacity without the instability of full parameter updates.

As pathology FMs continue to evolve, the adaptation paradigm itself requires further innovation. Current methods largely inherit strategies from natural image domains without fully addressing the unique challenges of histopathology, including multi-scale relationships, stain invariance, and complex morphological contexts. Future research should focus on developing pathology-specific adaptation techniques that explicitly incorporate domain knowledge while maintaining the efficiency and robustness necessary for clinical integration. Through careful strategy selection and continued methodological refinement, foundation models can realize their potential to transform cancer diagnosis, biomarker discovery, and precision oncology.

The digital transformation of histopathology, coupled with advancements in deep learning, is paving the way for a new era in computational pathology. However, a significant bottleneck hinders progress: the need for large, meticulously annotated datasets to train robust models. Fully supervised approaches requiring pixel-level or region-level annotations are resource-intensive, time-consuming, and do not scale to the vast amounts of data generated in clinical workflows. Weakly supervised learning (WSL) presents a paradigm shift by enabling model training using only slide-level labels, bypassing the need for costly manual annotations. A particularly promising approach involves the automated extraction of these weak labels from the free-text diagnostic reports that accompany whole slide images (WSIs) in routine clinical practice. This whitepaper provides an in-depth technical guide on leveraging free-text reports for training Convolutional Neural Networks (CNNs) and examines the emerging role of Foundation Models (FMs) within this paradigm, framing the discussion within a broader thesis on how FMs differ from traditional CNNs in pathology research.

Core Conceptual Differences: CNNs vs. Foundation Models in Pathology

The selection of an AI model architecture is a fundamental decision that dictates the entire workflow, from data preparation to clinical deployment. Table 1 summarizes the key distinctions between traditional CNNs and foundation models in the context of weakly supervised learning for pathology.

Table 1: Comparison of Traditional CNNs and Foundation Models in Pathology

Feature	Traditional CNNs (with MIL)	Foundation Models (FMs)
Primary Learning Approach	Trained from scratch or with generic ImageNet weights for a specific task (e.g., cancer detection).	Large-scale self-supervised pretraining (e.g., DINO, MAE) on vast, unlabeled image datasets, followed by adaptation.
Annotation Requirements	Slide-level labels (e.g., from reports); no pixel-wise annotations needed.	Massive volumes of unlabeled data for pretraining; task-specific labels for fine-tuning or linear probing.
Typical Workflow	End-to-end training on task-specific data, often using Multiple Instance Learning (MIL).	1. Pretrain on diverse tissue patches. 2. Extract frozen features (embeddings). 3. Train a simple classifier on these features (linear probing).
Handling of Whole Slides	Explicitly models slide as a "bag" of patches; uses attention mechanisms to identify diagnostically relevant regions.	Compresses image patches into fixed-size vector embeddings, potentially losing spatial and contextual information.
Interpretability	Attention maps highlight regions the model deemed important for prediction, offering some transparency.	Often function as "black boxes"; the reasoning behind compressed embeddings is difficult to interpret.
Performance on Complex Tasks	Excels in tasks like cancer detection (AUCs >0.99 reported), closely mimicking a pathologist's search process.	High performance on simple tasks (e.g., disease detection) but can fail on complex tasks like biomarker prediction (AUC ~0.60).
Robustness & Generalization	Trained directly on clinical data, showing better generalization to real-world variability in some studies.	Prone to "domain shift"; performance can drop 15-25% when applied to data from different hospitals/scanners.
Computational Resources	Relatively lower requirements for training and inference.	Extremely high resource burden; can consume up to 35x more energy than task-specific models.

The evidence suggests that while FMs offer the theoretical promise of universal feature extractors, traditional CNNs—particularly those employing MIL frameworks—currently demonstrate superior alignment with the practical needs of pathology, providing robust performance, better interpretability, and greater computational efficiency for many diagnostic tasks [20] [19].

Technical Framework and Experimental Protocols

Automated Label Extraction from Free-Text Reports

The first critical step in a weakly supervised pipeline is generating high-quality labels from unstructured pathology reports. This process involves Natural Language Processing (NLP) to convert clinical text into structured, machine-readable labels.

Protocol: NLP Pipeline for Label Extraction

• Data Collection & Preprocessing: Gather a large corpus of free-text pathology reports paired with their corresponding WSIs. No manual curation or annotation is required at this stage. Reports are typically in a semi-structured format, containing fields like specimen type, macroscopic description, microscopic description, and diagnosis.
• Concept Extraction: Implement an NLP tool, such as the Semantic Knowledge Extractor Tool (SKET), to analyze the reports. SKET is an unsupervised hybrid system that combines a rule-based expert system with pre-trained machine learning models [42].
• Label Assignment: The NLP tool scans the diagnostic text for semantically meaningful concepts corresponding to predetermined pathological classes (e.g., "adenocarcinoma," "high-grade dysplasia," "hyperplastic polyp"). Each WSI is then assigned a set of one or more labels (a multilabel problem) based on the concepts found in its report.
• Validation: To evaluate the quality of the automated labels, a subset of reports can be manually annotated by pathologists to create a ground truth. The performance of the automated tool is then assessed using standard classification metrics (e.g., accuracy, F1-score) against this ground truth. One study achieved a micro-accuracy of 0.908 at the image level using this method [42].

Weakly Supervised Image Classification with Multiple Instance Learning

Once weak labels are obtained, the next step is to train a CNN to classify the WSIs. Given the extreme size of WSIs (often exceeding 100,000x100,000 pixels), they are processed as collections of smaller patches. MIL is the dominant framework for this task.

Protocol: CNN Training with MIL

• Whole Slide Image Patching: Each WSI is divided into hundreds or thousands of small, non-overlapping patches (e.g., 256x256 pixels at 20x magnification). The entire WSI is treated as a "bag," and the patches within it are "instances."
• Model Architecture: A standard CNN (e.g., a ResNet) is used as a feature extractor for each individual patch. The patch-level features are then aggregated using an attention-based pooling layer.
• Attention Mechanism: The attention layer learns to assign a weight to each patch in the bag, indicating its relative importance for the slide-level prediction. This allows the model to focus on diagnostically relevant regions (e.g., cancerous tissue) while ignoring benign areas [43] [42] [20].
• Training: The model is trained end-to-end using only the slide-level labels automatically extracted from the reports. The loss function is computed based on the aggregated, slide-level prediction. This approach has been shown to achieve clinical-grade performance, with one study reporting an AUC of 0.991 for prostate cancer detection and 0.966 for breast cancer metastasis detection [20].
• Inference and Interpretation: During inference, the trained model can not only predict the slide-level label but also generate an attention heatmap overlaid on the original WSI. This heatmap visually indicates the regions that most influenced the decision, providing a degree of model interpretability valuable for pathologists.

The following diagram illustrates the complete integrated workflow, from raw data to diagnosis.

The Scientist's Toolkit: Key Research Reagents and Materials

Implementing the described workflows requires a suite of computational tools and data resources. Table 2 details the essential "research reagents" for developing weakly supervised learning models in computational pathology.

Table 2: Essential Research Reagents and Materials for Weakly Supervised Learning in Pathology

Item Name / Category	Function / Purpose	Examples & Notes
Whole Slide Images (WSIs)	The primary image data for model training and validation. Comprise high-resolution digitized tissue sections.	Sources: Institutional archives, public datasets (e.g., TCGA). Formats: SVS, NDPI, TIFF. Typically require 0.25–0.5 μm per pixel resolution [43] [42].
Free-Text Pathology Reports	Source for automated weak label extraction via NLP. Provides the diagnostic ground truth for each WSI.	Semi-structured clinical documents. Key fields: microscopic description, diagnosis. Multilingual support (e.g., Italian, Dutch) may be required [42].
NLP Tool for Label Extraction	Automatically analyzes free-text reports to extract semantically meaningful concepts as weak labels.	Tools like Semantic Knowledge Extractor Tool (SKET). Combines rule-based systems with pre-trained ML models [42].
Digital Pathology Platform	Software for managing, visualizing, and analyzing whole slide images. Often includes scanner integration and data management.	Proscia Concentriq, Philips IntelliSite (FDA-approved), Aperio (e.g., Aperio AT2) [43] [44].
Deep Learning Framework	Provides the programming environment for building, training, and testing CNN and FM architectures.	TensorFlow, PyTorch. Essential for implementing MIL frameworks and attention mechanisms [42].
Multiple Instance Learning (MIL) Framework	The core algorithmic architecture for training models with slide-level labels.	Custom implementations using CNNs (e.g., ResNet) with attention pooling layers. Allows the model to identify key patches [42] [20].
Computational Hardware	Accelerates the training of deep learning models on large-scale WSI datasets.	High-performance GPUs (e.g., NVIDIA RTX series, A100) are essential due to the computational intensity of processing WSIs [19].

Quantitative Performance Benchmarking

Empirical evidence is crucial for evaluating the real-world potential of any new methodology. The following table consolidates key quantitative findings from recent studies on weakly supervised learning and foundation models in pathology.

Table 3: Benchmarking Performance of Weakly Supervised CNNs and Foundation Models

Model / Study	Task / Description	Dataset	Key Performance Metric & Result
CNN with MIL (Campanella et al.)	Cancer vs. non-cancer classification [20].	44,732 WSIs from 15,187 patients across 3 tissue types [20].	AUC:• Prostate Cancer: 0.991• Basal Cell Carcinoma: 0.988• Breast Cancer Metastases: 0.966
NLP + CNN (Bianconi et al.)	Multi-label diagnosis of colon cancer using labels auto-extracted from reports [42].	3,769 clinical images & reports from 2 hospitals [42].	Image-Level Micro-Accuracy: 0.908
3D CNN - TriPath	Recurrence risk-stratification in prostate cancer using 3D tissue volumes [45].	Prostate cancer specimens imaged with 3D microscopy [45].	Performance: Superior to 2D slice-based approaches and clinical baselines from 6 genitourinary pathologists.
Pathology Foundation Models (Alfasly et al.)	Zero-shot retrieval across 23 organs & 117 cancer subtypes [19].	11,444 WSIs from TCGA [19].	Macro-Averaged F1 Score: ~40-42% (Top-5 retrieval).Organ-Level Variability: Kidneys: 68% (F1), Lungs: 21% (F1).
Pathology FMs (De Jong et al.)	Robustness evaluation across multiple institutions [19].	Multi-center WSI datasets [19].	Robustness Index (RI): Most models had RI ≈ 1 or less, meaning embeddings grouped by hospital/scanner, not biological class.

The automated leveraging of free-text reports for weakly supervised learning represents a powerful strategy to overcome the data annotation bottleneck in computational pathology. The evidence indicates that traditional CNNs, particularly those utilizing Multiple Instance Learning frameworks, are currently more clinically effective and robust than foundation models for many diagnostic tasks. They efficiently leverage weak labels to achieve high diagnostic accuracy while providing interpretable results through attention mechanisms. While foundation models hold future promise, their current limitations in domain robustness, high computational cost, and performance variability on complex tasks hinder clinical adoption. The path forward for pathology AI lies not in merely scaling model size, but in developing domain-aware, task-specific architectures that are closely aligned with the clinical workflow and the complex, contextual nature of tissue morphology.

Navigating the Pitfalls: Challenges and Optimization Strategies for Pathology AI

The field of computational pathology is undergoing a profound transformation, moving from specialized, task-specific convolutional neural networks (CNNs) to general-purpose pathology foundation models (PFMs). Traditional CNNs, with their strong inductive biases for spatial information, have excelled at localized tasks such as nuclei segmentation and cancer classification by capturing hierarchical features through convolutional layers [9] [46]. However, the emergence of PFMs—large-scale vision transformers (ViTs) pretrained on massive, diverse datasets of histopathology images—represents a fundamental architectural and methodological shift [37]. These models, typically ranging from hundreds of millions to over a billion parameters, generate robust, transferable feature representations adaptable to numerous downstream tasks without significant retraining [37]. While this shift brings remarkable performance improvements across cancer classification, tissue segmentation, and biomarker prediction [37], it introduces substantial computational and energy burdens that raise critical sustainability concerns for researchers and drug development professionals implementing these technologies in practice.

Architectural Comparison: Computational Foundations and Workflow Implications

Fundamental Architectural Differences

The operational divergence between CNNs and PFMs stems from their core architectural principles:

• CNN Architecture: CNNs employ a hierarchical structure with convolutional layers that apply learned filters across input images, leveraging local connectivity and translation invariance to progressively build more abstract representations. This design excels at capturing local patterns and spatial hierarchies through its inductive bias for grid-like data [46]. The processing is inherently local, with each layer having a limited receptive field that expands through network depth.

• PFM Architecture: Modern pathology foundation models predominantly utilize Vision Transformer (ViT) architectures that divide images into patches and process them as sequences using self-attention mechanisms [37]. This global attention mechanism allows each patch to interact with every other patch, enabling the model to capture long-range dependencies and complex contextual relationships across entire whole-slide images (WSIs) [37] [46]. This comes at a computational cost, as self-attention scales quadratically with the number of input patches.

Comparative Workflow Processing

The table below outlines the fundamental differences in how CNNs and PFMs process histopathology data:

Table 1: Architectural and Workflow Comparison Between CNNs and PFMs

Aspect	Traditional CNNs	Pathology Foundation Models
Core Architecture	Convolutional layers with local receptive fields	Vision Transformers with global self-attention
Parameter Scale	Millions of parameters	Hundreds of millions to billions of parameters [37]
Input Processing	Fixed-size image patches	Patch-based processing of whole-slide images [37]
Context Utilization	Local patterns within receptive field	Global context across entire tissue section
Pretraining Approach	Supervised on specific tasks	Self-supervised (DINOv2, MAE) or weakly-supervised on massive datasets [37]
Feature Representation	Task-specific features	General-purpose, transferable embeddings

Diagram 1: Comparative processing pipelines for CNNs versus Pathology Foundation Models

Quantitative Analysis of Computational and Energy Burden

Direct Energy Consumption Metrics

The scale of PFMs introduces substantial energy demands throughout the model lifecycle:

Table 2: Energy Consumption Comparison in Computational Pathology

Model Type	Training Energy	Inference Energy per Biopsy	Carbon Footprint
Task-Specific CNNs	Moderate (GPU days)	~0.63 Wh [37]	Lower
Pathology Foundation Models	High (GPU months)	6.74-22.09 Wh [37]	Up to 35× higher than CNNs [37]
Large Language Models (Reference)	MWh scale [47]	-	Hundreds of tonnes CO₂eq [47]

Recent empirical analyses reveal that PFMs are up to 35× more energy-intensive than parameter-matched task-specific networks in clinical deployment scenarios [37]. This energy burden extends throughout the model lifecycle, from pretraining on massive datasets encompassing hundreds of thousands to several million whole-slide images [37], to inference operations during clinical use.

Scaling Laws and Efficiency Trade-offs

The relationship between model scale, performance, and energy efficiency follows complex patterns:

• Performance Gains: PFMs consistently achieve performance improvements of 5-10 percentage points in AUC and balanced accuracy for major cancer subtyping tasks compared to ImageNet-pretrained CNNs [37]. Similar gains are observed for segmentation (DICE coefficients exceeding 0.80) and biomarker prediction (5-8% AUC improvements) [37].

• Sublinear Returns: As model scale increases from ViT-Base (~86M parameters) to ViT-Gigantic (~1.1B parameters), performance improvements often follow sublinear scaling laws while computational costs increase superlinearly [37].

• Hardware-Mediated Efficiency: Surprisingly, reducing parameters or FLOPs does not always yield better energy efficiency due to complex hardware-mediated effects and memory hierarchy considerations [48]. Cache-aware model design emerges as crucial for optimal energy utilization.

Methodologies for Measuring and Quantifying Energy Consumption

Experimental Protocols for Energy Assessment

Researchers can employ several methodologies to quantify the computational burden of foundation models:

Table 3: Energy Measurement Tools and Methodologies

Tool Name	Measurement Type	Key Metrics	Infrastructure Compatibility
CodeCarbon	Embedded Python package	CPU/GPU energy, CO₂eq emissions	Local servers, cloud platforms [47]
Eco2AI	Embedded package	Hardware-specific power draw	GPU clusters, workstations [47]
Green Algorithms	Online calculator	Estimated consumption from system specs	Any infrastructure [47]
CarbonTracker	Embedded package	Real-time power monitoring	HPC clusters, cloud environments [47]
Wattmeters	Physical measurement	Actual node-level consumption	Bare-metal servers, workstations [47]

Standardized Experimental Protocol for Model Comparison

To quantitatively compare CNN versus PFM energy consumption, researchers should implement this standardized protocol:

• Hardware Configuration: Utilize identical GPU infrastructure (e.g., NVIDIA H100 or A100) for all experiments with wattmeters for physical power measurement [47].

• Benchmarking Dataset: Employ a standardized histopathology dataset such as CAMELYON16 or TCGA whole-slide images with consistent preprocessing.

• Model Implementation:

  • CNN Baseline: Implement ResNet-50 or EfficientNet architectures pretrained on ImageNet
  • PFM Comparison: Utilize available foundation models (UNI, Virchow2, or PLUTO) with consistent feature extraction

• Energy Measurement:

  • Integrate CodeCarbon or Eco2AI tracking into training and inference pipelines
  • Record power consumption at 1-second intervals throughout experiments
  • Calculate total energy consumption as: Energy (Wh) = Average Power (W) × Time (h)

• Performance Assessment:

  • Evaluate on standardized tasks: tumor classification, segmentation, and biomarker prediction
  • Record AUC, DICE scores, and inference latency
  • Compute efficiency metrics: Performance per Watt and Inference per Joule

Diagram 2: Experimental workflow for measuring model energy consumption

Implementing energy-efficient computational pathology requires specific tools and frameworks:

Table 4: Essential Research Reagents and Computational Tools

Resource Category	Specific Tools/Platforms	Function in Research
Pathology Foundation Models	UNI, Virchow2, GigaPath, PLUTO-4G [37]	Pretrained feature extractors for transfer learning
Energy Measurement Tools	CodeCarbon, Eco2AI, Green-Algorithms [47]	Quantify energy consumption and carbon footprint
Computational Pathology Frameworks	CLAM, TIAToolbox, QuPath	Whole-slide image processing and analysis
Efficient Fine-tuning Methods	LoRA, PEFT, Linear Probing [37]	Parameter-efficient adaptation of foundation models
Benchmark Datasets	TCGA, CAMELYON, NCT-CRC-HE-100K	Standardized evaluation of model performance
Hardware Infrastructure	NVIDIA H100/A100 GPUs, High-memory servers	Compute resources for training and inference

Mitigation Strategies for Sustainable Model Deployment

Technical Approaches to Reduce Computational Burden

Researchers can employ several strategies to mitigate the energy impact of PFMs:

• Parameter-Efficient Fine-Tuning (PEFT): Methods like Low-Rank Adaptation (LoRA) update only small subsets of parameters during adaptation, reducing fine-tuning energy by up to 70% while maintaining performance [37].

• Optimal Adaptation Strategies: Empirical studies show linear probing (training only classifier heads) achieves optimal efficiency for few-shot tasks (<5 labels/class), while PEFT provides the best accuracy-efficiency trade-off for moderate data regimes (~100+ samples) [37].

• Model Distillation: Creating smaller, specialized models through knowledge distillation from large PFMs can reduce inference energy consumption while preserving much of the performance benefit [49].

• Dynamic Inference: Implementing adaptive computation pathways that use simpler models for straightforward cases and complex models only for challenging examples.

Energy-Aware Model Selection Framework

Researchers should consider a structured approach to model selection:

• Task Complexity Assessment: Evaluate whether the clinical task requires the contextual understanding of PFMs or can be addressed with efficient CNNs.

• Data Availability Consideration: Match model complexity to available annotated data, with linear probing of PFMs being optimal for limited data and full fine-tuning reserved for data-rich environments.

• Lifecycle Energy Budgeting: Estimate total energy consumption across training, inference, and maintenance phases when selecting modeling approaches.

• Hardware-Software Co-Design: Optimize model architectures for target deployment hardware to maximize computational efficiency.

The adoption of pathology foundation models represents a double-edged sword for computational pathology. While PFMs offer remarkable performance improvements and generalization capabilities across diverse diagnostic tasks, they incur substantial computational and energy costs that raise significant sustainability concerns. The 35× higher energy consumption compared to task-specific CNNs necessitates careful consideration by researchers and drug development professionals implementing these technologies.

The path forward requires a nuanced approach that matches model complexity to clinical need, employs parameter-efficient adaptation methods, and embraces energy-aware development practices. As the field advances, the development of more efficient transformer architectures, improved distillation techniques, and hardware-software co-design will be crucial for making foundation models environmentally sustainable while maintaining their transformative potential for pathology research and cancer diagnostics.

Researchers must balance the pursuit of state-of-the-art performance with environmental responsibility, ensuring that the computational pathology revolution benefits both human health and planetary wellbeing.

The application of artificial intelligence in medical imaging has remarkable advances, yet a fundamental challenge remains: deep learning models often struggle to maintain reliability and accuracy when deployed in new clinical environments due to domain shift [50]. In computational pathology, this shift manifests as variations between medical centers caused by differences in staining procedures, scanning equipment, image acquisition protocols, and tissue processing methods [51] [13]. These technical variations create confounding features that can dominate a model's learned representations, ultimately reducing its ability to generalize to data from laboratories not seen during training [13]. Within this context, foundation models have emerged as a promising alternative to traditional Convolutional Neural Networks (CNNs), offering new approaches to learning robust representations that prioritize biological features over technical artifacts. This technical guide examines the core differences between these architectural paradigms in addressing site-scanner bias, providing researchers with experimental frameworks and metrics for evaluating model robustness in pathology applications.

Fundamental Divergences: Traditional CNNs vs. Foundation Models

Architectural and Methodological Approaches

Traditional CNN-based approaches in pathology typically employ supervised learning on specific, often limited, annotated datasets. These models utilize convolutional layers to hierarchically extract features, starting with low-level patterns (edges, textures) and progressing to high-level histological structures [52]. For example, a 2017 study demonstrated a CNN pipeline achieving accuracies of 100%, 92%, and 95% for discriminating cancer tissues, subtypes, and biomarkers respectively, utilizing architectures like Inception and ResNet in an ensemble approach [53]. However, these models are typically trained for specific diagnostic tasks (e.g., classifying lung cancer subtypes or breast cancer biomarkers) and demonstrate limited adaptability to new domains without retraining [53] [54].

Pathology foundation models represent a paradigm shift toward large-scale, self-supervised learning on massive, diverse histopathology datasets. These models, including UNI, Phikon, Prov-GigaPath, and Virchow, are pre-trained on millions of histopathology tiles encompassing numerous tissue types, cancer indications, and medical centers [39]. They leverage self-supervised learning algorithms like DINOv2, iBOT, and masked autoencoders to learn general-purpose visual representations without requiring manual annotations during pre-training [39]. This approach aims to capture fundamental biological structures rather than task-specific patterns, theoretically enhancing their ability to generalize across domains.

Table 1: Core Architectural Differences Between Traditional CNNs and Foundation Models in Pathology

Characteristic	Traditional CNNs	Pathology Foundation Models
Training Paradigm	Supervised learning on task-specific datasets	Self-supervised learning on massive, diverse datasets
Scale of Pre-training	Limited to thousands or hundreds of thousands of images	Millions to billions of tiles from hundreds of thousands of slides
Architecture	Convolutional networks (ResNet, EfficientNet)	Vision Transformers (ViT), hybrid CNN-Transformers
Adaptation Approach	Full retraining or fine-tuning	Lightweight fine-tuning, adapter layers, prompt tuning
Representation Focus	Task-specific histological patterns	General-purpose tissue and cellular representations

Quantitative Performance Comparison

Recent comparative studies reveal the performance differential between these approaches. In breast cancer classification using the BreakHis dataset, CNN-based models like ResNet50, RegNet, and ConvNeXT achieved an AUC of 0.999 in binary classification tasks [3]. Similarly, traditional CNNs demonstrated 99.98% classification accuracy in breast cancer and diabetic retinopathy diagnosis in optimized implementations [54]. However, foundation models like UNI achieved superior performance in more complex eight-class classification tasks (95.5% accuracy vs. CNN performance in the 90-95% range), suggesting better handling of increased complexity [3].

The computational efficiency trade-offs are equally important. The Pathology-NAS framework, which leverages LLM-driven neural architecture search, achieves 99.98% classification accuracy while reducing FLOPs by 45% compared to leading methods [54]. This demonstrates how architectural optimization can enhance efficiency without sacrificing performance.

Table 2: Performance Comparison on Histopathology Tasks

Model Type	Binary Classification Accuracy	Multi-class Classification Accuracy	Computational Efficiency
Traditional CNNs (ResNet50, EfficientNet)	99.2-99.98% [54] [3]	90-95% [3]	Moderate to High [3]
Vision Transformers (Swin-Transformer)	98.08% [3]	70.38% [3]	Lower (without optimization) [54]
Foundation Models (UNI, with fine-tuning)	99.98% [54]	95.5% [3]	Variable (model-dependent) [39]
Optimized Architectures (Pathology-NAS)	99.98% [54]	High (task-dependent) [54]	High (45% FLOPs reduction) [54]

Measuring Robustness: Quantitative Frameworks

The Robustness Index

A critical advancement in evaluating model robustness is the Robustness Index (Rk), which quantifies the extent to which biological features dominate confounding features in a model's embedding space [13]. The index is formally defined as:

Rk = ΣΣ 1(yj = yi) / ΣΣ 1(cj = ci)

Where y represents biological class (e.g., cancer type), c represents medical center, and k denotes the number of nearest neighbors considered (typically k=50) [13]. This metric directly measures whether a model's representation space is organized primarily by biologically relevant features or by confounding technical factors.

Application of this index to ten current pathology foundation models revealed significant variability in robustness, with only one model achieving a robustness index greater than 1 (indicating biological features dominate confounding features, though only slightly) [13]. This finding highlights a critical challenge: most current pathology foundation models remain strongly influenced by medical center signatures despite their extensive training.

Benchmarking Clinical Performance

Beyond embedding-space analysis, comprehensive clinical benchmarking provides crucial insights into real-world performance. Recent evaluations of public self-supervised pathology foundation models on disease detection tasks revealed that all models showed consistent performance with AUCs above 0.9 across all tasks [39]. However, performance variations emerge in more specific applications like biomarker prediction and survival analysis, where factors like training data diversity and model architecture significantly influence outcomes [39].

Experimental Protocols for Robustness Evaluation

Robustness Assessment Framework

Diagram 1: Robustness assessment for pathology FMs

A comprehensive robustness evaluation should incorporate multiple medical centers in the test dataset, ensuring sufficient representation of biological classes across different technical environments [13]. The experimental workflow proceeds through several critical stages:

• Multi-Center Dataset Curation: Assemble whole slide images from at least 3-5 independent medical centers, ensuring each biological class (e.g., cancer type) is represented across multiple centers [13]. This controls for biological variation while allowing measurement of center-specific effects.

• Embedding Extraction: Process image tiles through the foundation model to generate feature embeddings, maintaining associations with both biological labels and medical center origins [13].

• Neighborhood Analysis: For each sample, identify its k-nearest neighbors (k=50) in the embedding space using cosine distance [13].

• Robustness Index Calculation: Compute the ratio of same-class neighbors to same-center neighbors across all samples [13].

• Error Attribution: Analyze classification errors to determine if they correlate with medical center origins, particularly investigating whether errors are attributable to "same-center confounders" - images from the same center but different class that appear nearby in embedding space [13].

Domain Generalization Protocols

Diagram 2: Domain generalization techniques

Evaluating domain generalization requires rigorous protocols that test models on completely unseen domains. Key methodological considerations include:

• Data Splitting Strategy: Implement center-wise splitting rather than random splitting, ensuring all samples from certain medical centers are entirely absent during training [55].

• Automatic Augmentation Methods: Employ algorithms like RandAugment that automatically search for optimal augmentation policies, which have demonstrated state-of-the-art domain generalization performance in histopathology [51].

• Multi-Task Evaluation: Assess performance across diverse tasks including disease detection, biomarker prediction, and cancer subtyping using datasets from multiple independent medical centers [39].

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Tools for Robustness Evaluation

Tool/Category	Specific Examples	Function & Application
Public Foundation Models	UNI, Phikon, CTransPath, Prov-GigaPath [39]	Pre-trained models for feature extraction and transfer learning
Robustness Evaluation Metrics	Robustness Index (Rk) [13]	Quantifies dominance of biological vs. confounding features in embedding space
Domain Generalization Algorithms	RandAugment, Adversarial Alignment, Meta-Learning [51] [55]	Improves model performance on unseen domains through data augmentation and specialized learning strategies
Benchmark Datasets	TCGA, Clinical benchmarks from multiple medical centers [53] [39]	Standardized datasets for evaluating cross-domain performance
Neural Architecture Search	Pathology-NAS, LLM-driven architecture search [54]	Automates discovery of optimal model architectures for specific tasks and constraints

Implementation Guidelines: From Theory to Practice

Practical Recommendations for Model Selection

Based on current evidence, researchers should consider the following when selecting approaches for robust pathology AI:

• For task-specific applications with limited data diversity: Well-optimized traditional CNNs still provide strong performance, particularly when enhanced with automatic data augmentation techniques [51].

• For multi-center deployments requiring generalization: Foundation models with demonstrated cross-center performance offer advantages, particularly when fine-tuned with robustness-aware techniques [39].

• For resource-constrained environments: Lightweight architectures discovered through neural architecture search (e.g., Pathology-NAS) provide an optimal balance of performance and efficiency [54].

Mitigation Strategies for Domain Shift

Several technical approaches have demonstrated effectiveness in addressing domain shift:

• Automatic Data Augmentation: Frameworks that automatically search for optimal augmentation policies can achieve state-of-the-art domain generalization performance, sometimes surpassing manually curated augmentation strategies [51].

• Adapter-Based Fine-Tuning: Rather than full model fine-tuning, using lightweight adapter layers can preserve general-purpose representations while adapting to specific domains [50].

• Stain-Aware Training: Incorporating stain-specific augmentations during training improves model resilience to staining variations encountered across medical centers [39].

Future Directions and Research Challenges

Despite significant advances, important challenges remain in achieving truly robust computational pathology models. Current foundation models still exhibit strong medical center signatures in their embedding spaces, with most having robustness indices below 1 [13]. This indicates confounding features still dominate biological features in most models. Future research should focus on developing training methodologies that explicitly optimize for robustness metrics during pre-training rather than treating robustness as a secondary consideration.

The creation of more comprehensive benchmarking datasets spanning diverse medical centers, staining protocols, and scanner types will be essential for proper evaluation of model generalization [39]. Additionally, techniques for better integration of domain adaptation and generalization methods with foundation models represent a promising research direction [50]. As the field progresses, the development of standardized robustness evaluation protocols and reporting standards will be critical for clinical translation of these technologies.

The development of artificial intelligence (AI) for computational pathology hinges on learning from Whole Slide Images (WSIs), which are gigapixel-sized digital scans of tissue specimens. Traditional Convolutional Neural Networks (CNNs) operate on a single-instance learning paradigm, requiring large datasets of annotated image patches to learn effectively. This creates a fundamental bottleneck in pathology AI, as curating such datasets is prohibitively expensive and time-consuming due to the need for expert pathologist annotations [56]. Furthermore, the multi-gigapixel nature of WSIs makes them computationally intractable for conventional CNNs, which typically process small, fixed-size images. Foundation Models (FMs) represent a paradigm shift. These models are large-scale neural networks pretrained on massive, often unlabeled or weakly labeled, datasets using Self-Supervised Learning (SSL). This pretraining allows them to learn general-purpose, transferable feature representations of histopathology data, which can then be adapted to various downstream tasks with minimal task-specific data [57]. The core distinction lies in the learning approach: traditional CNNs require direct, manual supervision for specific tasks, whereas FMs first learn the fundamental "language" of histology morphology in a task-agnostic way, reducing the dependency on scarce, annotated data for each new application.

Technical Approaches to Overcome Data Limits

To circumvent the challenges of data scarcity and privacy, researchers have developed sophisticated pretraining strategies that leverage different types and sources of data. The following table summarizes the primary technical approaches identified in recent literature.

Table 1: Technical Approaches for Pretraining with Limited Data

Approach	Core Methodology	Key Example Models	Addresses Scarcity Via
Large-Scale Self-Supervised Learning (SSL)	Uses pretext tasks (e.g., masked image modeling, contrastive learning) on unlabeled images to learn morphological features.	UNI [57], Phikon [57], Prov-GigaPath [18]	Leverages vast repositories of unlabeled WSIs, eliminating need for manual annotations.
Multimodal Vision-Language Pretraining	Aligns histopathology images with corresponding text (e.g., pathology reports, synthetic captions) in a shared embedding space.	CONCH [58], TITAN [6], PLIP [57]	Incorporates rich, freely available textual data as a weak supervisory signal.
Synthetic Data Generation	Employs generative AI models to create realistic, annotated image-caption pairs for pretraining.	TITAN (uses PathChat) [6]	Artificially expands the scale and granularity of training datasets.
Weakly-Supervised Pre-Training	Propagates slide-level weak labels to individual image patches to improve instance-level representation learning.	SimMIL [59]	Makes more efficient use of the limited labeled data that is available.

Large-Scale Self-Supervised Learning (SSL)

SSL techniques enable models to learn from the inherent structure of the data itself without human-provided labels. For example, the Prov-GigaPath model was pretrained on a massive dataset of 1.3 billion image tiles from 171,189 WSIs using a combination of two SSL methods: DINOv2 for tile-level feature learning and a masked autoencoder objective with a LongNet architecture for whole-slide context modeling [18]. This scale of pretraining allows the model to learn robust, generalizable features that are not tied to a specific annotation campaign.

Multimodal Vision-Language Pretraining

Inspired by models like CLIP in natural image processing, this approach learns a joint representation space for images and text. The CONCH model, for instance, was pretrained on over 1.17 million histopathology image-caption pairs using a contrastive learning objective paired with a captioning loss [58]. This allows the model to connect visual morphological patterns with their semantic descriptions found in pathology reports or medical literature. The TITAN model further extends this by aligning WSIs with both real pathology reports and 423,122 synthetic captions generated by a generative AI copilot [6]. This multimodality is key to overcoming data limits, as it leverages the vast, often untapped, resource of textual medical data.

Leveraging Synthetic Data

The use of synthetic data generated by AI is an emerging and powerful strategy to combat data scarcity. As exemplified by TITAN, generative AI can produce high-quality, fine-grained descriptions of histopathology regions of interest (ROIs) [6]. These synthetic captions can augment real-world data, increasing the diversity and scale of the training corpus without compromising patient privacy, as the synthetic data does not contain real patient information.

Quantitative Performance in Data-Scarce Scenarios

The true value of foundation models is demonstrated in their performance on downstream tasks, especially when labeled data is limited. The following table compares the performance of various models across different data-efficient learning settings.

Table 2: Performance Comparison of Foundation Models in Low-Data Regimes

Model	Pretraining Data Scale	Zero-Shot Performance (Example)	Few-Shot / Linear Probing Performance
CONCH	1.17M image-text pairs [58]	90.7% accuracy on NSCLC subtyping [58]	State-of-the-art (SOTA) on 14/14 diverse benchmarks [58]
TITAN	335,645 WSIs + 423k synthetic captions [6]	Effective rare disease retrieval & report generation [6]	Outperforms slide & ROI models in few-shot classification [6]
Prov-GigaPath	1.3B tiles from 171k WSIs [18]	N/A	SOTA on 25/26 tasks (e.g., 23.5% AUROC boost in EGFR prediction) [18]
UNI	Large-scale, unspecified [57]	N/A	Superior performance at a quarter of the computational cost of others [56]
H-optimus-0	Large-scale, unspecified [56]	N/A	Best performer (89% BA) for ovarian carcinoma subtyping [56]

The data shows that FMs pretrained with the aforementioned strategies excel in zero-shot and few-shot settings. For instance, CONCH's zero-shot capability allows it to classify non-small cell lung cancer subtypes with over 90% accuracy without any task-specific training data [58]. In a rigorous evaluation on ovarian carcinoma subtyping, foundation models like H-optimus-0 and UNI achieved high balanced accuracies (up to 89% on an internal test set), demonstrating strong utility even for complex diagnostic tasks with limited fine-tuning data [56]. Prov-GigaPath's success on 25 out of 26 diverse tasks further underscores the generalization power afforded by large-scale pretraining [18].

Detailed Experimental Protocols

To ensure reproducibility and provide a clear technical guide, this section outlines the core experimental methodologies used to validate the performance of pathology foundation models in data-scarce scenarios.

Zero-Shot Classification Protocol

This protocol evaluates a model's ability to perform a task without any task-specific training. It is commonly used for vision-language models like CONCH and TITAN.

• Prompt Engineering: For each class in the target classification task (e.g., "invasive ductal carcinoma," "renal cell carcinoma"), a set of text prompts is created. An ensemble of prompts (e.g., "a histopathology image of {class name}", "a tissue sample showing {class name}") is often used to improve robustness [58].
• Text Embedding Generation: The text encoder of the pretrained foundation model processes each prompt to generate a set of text embeddings for all classes.
• Image Embedding Generation: The image encoder processes the input WSI or image tile to generate a visual embedding.
• Similarity Matching: The cosine similarity is computed between the visual embedding and each of the text embeddings.
• Prediction: The class whose prompt embeddings have the highest average similarity to the visual embedding is selected as the prediction [58].
• Whole-Slide Inference (MI-Zero): For gigapixel WSIs, the slide is divided into tiles. Each tile is classified via zero-shot matching. Tile-level scores are then aggregated (e.g., via max-pooling or averaging) to produce a final slide-level prediction [58].

Few-Shot Linear Probing Protocol

This protocol evaluates a model's utility as a feature extractor when only a very small labeled dataset is available for a downstream task.

• Feature Extraction: The foundation model's encoder is frozen. The target dataset (with only a few samples per class) is processed, and feature embeddings are extracted for each WSI or image tile.
• Classifier Training: A shallow, task-specific classifier (typically a linear layer or logistic regression model) is trained on top of the frozen features using the small labeled dataset.
• Evaluation: The trained linear classifier is evaluated on a held-out test set. The performance (e.g., accuracy, AUROC) reflects the quality and transferability of the pretrained features in a data-scarce setting [56] [40]. This approach is widely used because full fine-tuning of large FMs on small data is often unstable and leads to overfitting [40].

Research Reagent Solutions

The following table details the key computational "reagents" required for implementing the pretraining and evaluation strategies discussed in this whitepaper.

Table 3: Essential Research Reagents for Pathology Foundation Model Development

Reagent / Resource	Function / Description	Example Models / Tools
Large-Scale WSI Datasets	Provides the raw, unlabeled data for self-supervised pretraining. Diversity in organs, stains, and scanners is critical.	Prov-Path [18], Mass-340K [6]
Image-Text Pair Datasets	Enables vision-language pretraining by pairing histology images with descriptive text (reports or captions).	CONCH's 1.17M pairs [58], TITAN's reports & synthetic captions [6]
Synthetic Caption Generator	A generative AI model that produces fine-grained morphological descriptions for image patches, expanding training data.	PathChat [6]
Self-Supervised Learning Framework	Software libraries that provide implementations of SSL algorithms like masked autoencoding or contrastive learning.	DINOv2 [18], iBOT [6], MAE [60]
Multiple Instance Learning (MIL) Aggregator	An architecture that aggregates features from thousands of patches to form a single WSI-level representation.	ABMIL [56], HIPT [57]
Long-Sequence Transformer	A specialized transformer architecture capable of processing the extremely long token sequences representing entire WSIs.	LongNet [18]

Workflow Visualization

The following diagram illustrates the integrated workflow for building a foundation model for pathology using strategies to overcome data scarcity.

The limitations imposed by data scarcity and privacy in computational pathology are being effectively surmounted by a new generation of foundation models. These models fundamentally differ from traditional CNNs by shifting the paradigm from supervised learning on annotated datasets to self-supervised and multimodal pretraining on vast, unlabeled corpora. Through techniques such as large-scale SSL, vision-language alignment, and the use of synthetic data, FMs learn a foundational understanding of histopathology morphology. This enables robust performance in critical data-scarce scenarios like zero-shot inference and few-shot learning, as evidenced by their state-of-the-art results on diverse diagnostic and prognostic tasks. The future of pathology AI lies in refining these data-efficient approaches, making powerful diagnostic tools more accessible and accelerating their integration into clinical and drug development workflows.

In computational pathology, the adaptation of foundation models (FMs) has increasingly diverged from the core paradigm envisioned for general artificial intelligence. Rather than leveraging full fine-tuning to adapt these massive models to specific diagnostic tasks, researchers and practitioners have overwhelmingly retreated to linear probing—training only a simple linear classifier on top of frozen feature embeddings from the FM backbone [40] [19]. This pragmatic adaptation strategy has become ubiquitous not due to superior performance, but because the alternative—fine-tuning the entire model—frequently degrades performance, causes catastrophic forgetting, and demands unsustainable computational resources [40]. This dependency stands in stark contrast to the foundational premise of the FM paradigm, which promises large pretrained systems that enable zero-shot and easily fine-tuned adaptation across domains [19]. In pathology, this promise has collapsed, exposing a significant gap between theoretical universality and real-world usability in medical AI.

The retreat to linear probing represents a fundamental contradiction in the current application of pathology FMs. As Tizhoosh critically observes, this situation is "akin to buying a Ferrari that cannot run and then purchasing a bicycle to tow it" [40]. Most contemporary pathology FMs function not as adaptable "foundations" but as static feature extractors that must be linearly probed, raising important questions about their true flexibility and value proposition in clinical settings [19]. This article examines the prevalence, causes, and implications of this phenomenon, situating it within the broader context of how foundation models differ from traditional convolutional neural networks (CNNs) in pathology research.

Quantitative Evidence: The Performance Trade-offs of Adaptation Strategies

Recent empirical studies consistently demonstrate that fine-tuning instability is not merely a theoretical concern but a practical limitation affecting real-world applications. The performance trade-offs between different adaptation strategies reveal clear patterns across multiple pathology tasks and datasets.

Table 1: Comparative Performance of FM Adaptation Strategies in Pathology

Study	Task	Linear Probing Performance	Full Fine-Tuning Performance	Performance Gap
Alfasly et al. [40]	Multi-organ cancer subtyping (23 organs, 117 subtypes)	Macro F1: 40-42% (top-5 retrieval)	Frequently degrades accuracy	Negative
Kidney Pathology Study [15]	Kidney disease diagnosis (3-class classification)	AUROC >0.980 (internal), robust external validation	Marked performance drop for ResNet50	Significant advantage for linear probing
Multi-center Robustness Study [40]	Cross-institutional generalization	RI >1.2 only for Virchow2 (biological structure dominates)	Most models RI <1 (site-specific bias dominates)	Linear probing more robust to domain shift
Prostate Cancer Diagnosis [40]	Gleason grading	Competitive with task-specific models	Overfitting, catastrophic forgetting	Linear probing more stable

The evidence from these studies indicates that linear probing not only matches but often surpasses fine-tuning performance, particularly in scenarios requiring robustness to domain shifts [15]. In one comprehensive kidney pathology analysis, all foundation models using linear probing with multiple instance learning (MIL) frameworks significantly outperformed ImageNet-pretrained ResNet50, especially in external validation where traditional fine-tuning approaches showed marked performance drops [15]. This pattern holds across various cancer types and diagnostic tasks, suggesting the phenomenon is not isolated to specific tissue types or disease categories.

Beyond simple accuracy metrics, linear probing demonstrates superior computational efficiency. In large-scale comparisons, foundation models consumed up to 35× more energy than task-specific models [40] [19], raising sustainability concerns for clinical deployment. When fine-tuning is attempted, this computational burden increases substantially without consistent performance benefits, making linear probing an attractive option for both practical and environmental reasons.

Methodological Foundations: Experimental Protocols for FM Evaluation

The evaluation of foundation model adaptation strategies follows rigorous experimental protocols designed to assess both performance and robustness. Understanding these methodologies is crucial for interpreting results and designing effective implementation strategies.

Standard Linear Probing Protocol

The prevailing approach to linear probing employs a standardized workflow:

• Feature Extraction: Frozen FM backbone processes pathology patches (typically 256×256 pixels at 20× magnification) to generate feature embeddings [15]. Common FMs include UNI, Virchow, Virchow2, Prov-GigaPath, and Phikon, each pretrained on large-scale histopathology datasets using self-supervised learning objectives like DINOv2 or masked image modeling [15].

• Feature Aggregation: For whole-slide image classification, patch-level features are aggregated using attention-based multiple instance learning (ABMIL), transformer-based MIL (TransMIL), or clustering-constrained attention mechanisms (CLAM) to create slide-level representations [15]. This step is crucial for handling the gigapixel nature of whole-slide images where only small regions may be diagnostically relevant.

• Classifier Training: A linear layer (typically a single fully-connected layer with softmax activation) is trained on the frozen features while the FM backbone remains completely frozen [40] [15]. This approach minimizes the risk of overfitting, particularly important given the limited annotated datasets typical in pathology.

The stability of this approach contrasts sharply with full fine-tuning, where researchers report "fragile adaptation" characterized by training instability, overfitting to small pathology datasets, and catastrophic forgetting of pretrained representations [40].

Comparative Evaluation Framework

Robust evaluation of adaptation strategies incorporates multiple dimensions of assessment:

• Internal Validation: Standard k-fold cross-validation (typically 5-fold) on the development dataset with patient-level splits to prevent data leakage [15].

• External Validation: Performance assessment on completely independent datasets from different institutions, essential for evaluating real-world generalization [15]. Studies consistently show that fine-tuned models experience more significant performance drops on external validation compared to linear probing approaches.

• Robustness Metrics: Quantitative assessment of domain shift sensitivity using metrics like the Robustness Index (RI), which compares whether model embeddings cluster more strongly by biological class or by medical center [40]. An RI >1 indicates true biological robustness, while RI <1 suggests confounding by site-specific bias.

Diagram 1: FM adaptation workflow comparison. Linear probing maintains model stability while full fine-tuning often leads to performance degradation.

Underlying Causes: Architectural and Domain-Specific Factors

The instability of fine-tuning in pathology FMs stems from fundamental architectural differences between transformers and traditional CNNs, combined with unique characteristics of medical imaging data.

Architectural Differences: Transformers vs. CNNs

Table 2: Architectural Comparison of FMs vs. Traditional CNNs in Pathology

Characteristic	Traditional CNNs	Vision Transformers (FMs)	Implications for Fine-Tuning
Inductive Biases	Strong (locality, translation equivariance)	Minimal (global attention)	ViTs require more data for effective adaptation
Parameter Count	Typically <100M	100M to >1B parameters	ViTs more prone to overfitting on small medical datasets
Feature Representation	Hierarchical, local to global	Global context from outset	ViTs capture long-range dependencies but are less sample-efficient
Data Requirements	Effective with smaller datasets	Require massive pretraining	Pathology datasets often insufficient for stable fine-tuning
Geometric Robustness	Built-in via convolutional kernels	Requires explicit augmentation	ViTs more fragile to rotation, scale variations without careful training

Unlike CNNs with their strong inductive biases for visual tasks, transformer-based FMs employ minimal architectural assumptions, relying instead on pretraining scale to learn relevant representations [15]. While this enables exceptional performance when pretraining data is abundant, it creates fragility during adaptation to specialized domains like pathology, where annotated datasets are typically small and exhibit significant domain shift across institutions [40].

Domain-Specific Challenges in Pathology

The biological complexity of human tissue presents unique challenges that exacerbate fine-tuning instability:

• Biological Complexity: Tissue interpretation requires contextual, multi-scale reasoning far beyond simple object recognition [40]. A pathologist needs over twelve years of training to distinguish cancer subtypes, reflecting the semantic complexity that FMs must capture.

• Ineffective Self-Supervision: Most self-supervised learning frameworks were designed for natural images with discrete objects, making them less suitable for complex histopathology slides where patches contain mixed or irrelevant content [40]. Consequently, models may learn stain textures rather than biologically meaningful patterns.

• Information Compression Bottleneck: FMs compress entire tissue patches into fixed-size embeddings, inevitably losing diagnostically critical information [20]. This compression is particularly problematic for pathology, where diagnostic decisions rely on subtle morphological features and spatial relationships.

• Patch-Size Mismatch: A critical but overlooked issue is the fundamental mismatch between standard ViT patch sizes and the field-of-view needed for pathological assessment [40]. Standard patches (e.g., 224×224 pixels) often fail to capture meso- and macro-architectural patterns essential for diagnosis.

Implementing and evaluating adaptation strategies for pathology FMs requires specialized computational resources and methodological approaches.

Table 3: Essential Research Reagents for FM Adaptation Experiments

Resource Category	Specific Tools & Platforms	Function in FM Research
Foundation Models	UNI, Virchow/Virchow2, Prov-GigaPath, Phikon, PLUTO-4G	Pretrained backbones for feature extraction and transfer learning
Multiple Instance Learning Frameworks	ABMIL, TransMIL, CLAM	Slide-level feature aggregation from patch embeddings
Computational Infrastructure	High-memory GPUs (e.g., NVIDIA A100, H100 clusters), PyTorch, HuggingFace	Model training, inference, and feature extraction
Pathology Datasets	TCGA, KPMP, JP-AID, institutional WSI repositories	Benchmarking and evaluation across diverse tissue types
Evaluation Metrics	Robustness Index (RI), AUROC, F1-score, Representation Shift Metrics	Quantifying performance and generalization capability

The resource requirements for effective FM research are substantial, with successful implementation typically requiring access to high-performance computing infrastructure and diverse, multi-institutional pathology datasets for robust evaluation [37] [15]. The trend toward larger models (e.g., PLUTO-4G with 1.1B parameters) further increases these computational demands, creating significant barriers to entry for smaller research groups and healthcare institutions [37].

Implications and Future Directions

The prevalence of linear probing over fine-tuning has profound implications for both the development of pathology FMs and their clinical translation.

Clinical Translation Challenges

The dependency on linear probing creates significant hurdles for clinical implementation:

• Limited Adaptability: Without effective fine-tuning, FMs cannot be easily customized to institution-specific protocols, staining variations, or specialized diagnostic tasks [40].

• Domain Shift Vulnerability: Despite better performance compared to fine-tuning, linear probing still suffers from performance degradation across institutions, with most models showing significant sensitivity to scanner and staining variations [40] [61].

• Resource-Intensive Deployment: The computational footprint of FMs remains substantial even with linear probing, creating challenges for real-time clinical integration and raising environmental sustainability concerns [40].

Emerging Solutions and Alternative Approaches

Research communities are exploring several strategies to address fine-tuning instability:

• Parameter-Efficient Fine-Tuning (PEFT): Methods like Low-Rank Adaptation (LoRA) that update only small subsets of parameters show promise for balancing adaptability and stability [37]. Empirical studies indicate PEFT achieves the highest accuracy for moderate data regimes (~100+ samples per class) while maintaining stability [37].

• Hybrid Architectures: Combining the robustness of CNNs with the representational power of transformers may offer improved adaptation characteristics [37]. Models like CTransPath and PathOrchestra exemplify this approach.

• Geometric-Aware Training: Explicitly incorporating rotation and scale invariance through data augmentation and architectural modifications improves robustness without requiring full fine-tuning [40].

• Multiple Instance Learning Integration: As demonstrated in kidney pathology diagnosis, combining FMs with MIL frameworks provides strong performance while maintaining the stability of linear probing [15].

Diagram 2: Pathology FM instability causes and solutions. Multiple research directions aim to address the fundamental limitations of current adaptation approaches.

The prevalence of linear probing over fine-tuning in pathology foundation models represents a significant deviation from the original FM paradigm and highlights fundamental challenges in adapting these general-purpose architectures to specialized medical domains. This retreat to linear probing stems from architectural mismatches, biological complexity, data limitations, and computational constraints that collectively make full fine-tuning unstable and impractical in most clinical pathology scenarios.

While linear probing provides a stable and often effective adaptation strategy, its limitations constrain the full potential of FMs in pathology. The path forward requires neither abandonment of the FM approach nor blind faith in scaling laws, but rather fundamental rethinking of how to build pathology-specific architectures and adaptation methods that balance representational power with clinical practicality. Emerging approaches including parameter-efficient fine-tuning, hybrid architectures, and enhanced integration with multiple instance learning frameworks offer promising avenues for developing more adaptable and robust pathology AI systems capable of meeting the stringent demands of clinical diagnostics.

Benchmarking Performance: Empirical Evidence for FMs and CNNs in Clinical Workflows

The adoption of artificial intelligence (AI) in pathology promises to revolutionize cancer diagnosis, prognosis, and therapeutic response prediction. However, the transition from research laboratories to clinical practice hinges on rigorously demonstrating diagnostic accuracy and robustness across diverse, independent datasets and medical centers. This requirement is particularly critical when evaluating foundation models (FMs)—large-scale AI systems pre-trained on extensive datasets—against traditional task-specific convolutional neural networks (CNNs). Foundation models represent a paradigm shift from traditional CNNs, which are typically trained on limited, annotated datasets for specific classification tasks. FMs leverage self-supervised learning on vast collections of unlabeled whole-slide images (WSIs) to learn universal histopathological representations, theoretically enabling superior generalization to novel tasks and environments with minimal fine-tuning [62].

This technical guide examines methodologies and evidence for validating the diagnostic accuracy and robustness of pathology AI models, with a focused comparison between emerging foundation models and conventional CNNs. We synthesize recent empirical evidence, provide detailed experimental protocols for cross-center validation, and offer practical toolkits for researchers conducting rigorous evaluation studies. By framing this discussion within the broader thesis of how foundation models differ from traditional CNNs in pathology research, we aim to equip scientists and drug development professionals with the frameworks necessary to critically assess the translational potential of these technologies.

Foundation Models vs. Traditional CNNs: A Conceptual and Practical Comparison

Table 1: Comparative Analysis of Foundation Models and Traditional CNNs in Computational Pathology

Characteristic	Foundation Models (FMs)	Traditional CNNs
Core Architecture	Typically Vision Transformers (ViTs) or hybrid architectures [63]	Typically EfficientNet, ResNet, U-Net [64] [65]
Training Paradigm	Self-supervised learning (SSL) on massive, diverse, unlabeled WSI datasets [62]	Supervised learning on smaller, task-specific, labeled datasets [64] [65]
Data Requirements	Extremely large (100,000+ WSIs), unlabeled or weakly labeled [66] [62]	Moderate (100s-1,000s WSIs), requires precise manual annotations [64] [65]
Computational Cost	Very high pre-training cost; fine-tuning is more efficient [62]	Lower per-model cost, but cumulative cost can be high if many models are needed
Primary Strengths	Broad generalization, adaptability to new tasks with little data, multimodal potential [62]	High performance on specific trained tasks, more interpretable, lower infrastructure barrier [64] [53]
Key Weaknesses	High pre-training cost, "black box" nature, unstable fine-tuning, emerging robustness concerns [19] [62]	Narrow scope, poor transfer to new tasks/organ s, annotation bottleneck [19]
Typical Validation Approach	Linear probing on frozen features, then task-specific fine-tuning [19]	End-to-end training and validation on partitioned datasets

Empirical Evidence: Performance and Robustness Across Datasets

Recent systematic evaluations have revealed critical insights into the real-world performance of both foundation models and traditional CNNs, particularly regarding their generalization capabilities.

Table 2: Empirical Evidence from Multi-Center and Cross-Dataset Validation Studies

Study & Model Type	Key Validation Findings	Reported Metrics	Implications
Alfasly et al. (FM Evaluation) [19]	Low, inconsistent accuracy across 23 organs & 117 cancer subtypes; performance variability from 21% (lung) to 68% (kidney) top-1 F1 score.	Macro F1: ~40-42% (top-5 retrieval)	Questions true diagnostic generalization of current FMs; may learn texture artifacts over morphology.
De Jong et al. (FM Robustness) [19]	Most FMs (except Virchow2) showed significant site bias (RI < 1); embeddings clustered more by hospital/scanner than cancer type.	Robustness Index (RI); RI > 1 indicates biological clustering	Highlights vulnerability to technical confounders, raising concerns for clinical use on unseen data.
Mulliqi et al. (FM vs. TS CNN) [19]	On a large prostate biopsy dataset (100k+ slides), task-specific (TS) CNN matched/exceeded FM performance when sufficient labeled data existed. FMs consumed 35x more energy.	Diagnostic accuracy for Gleason grading, Energy consumption	Challenges universal superiority of FMs; questions sustainability and practical advantage in data-rich scenarios.
Wang et al. (FM Security) [19]	Universal adversarial perturbations (UTAP) severely degraded FM performance (e.g., 97% → 12% accuracy), transferring across model architectures.	Accuracy drop under attack	Reveals critical safety vulnerabilities; perturbations mimic real-world stain/scanner variations.
PathOrchestra FM [66]	Demonstrated high accuracy (>0.95 AUC) on 47/112 tasks across 61 private and 51 public datasets, including pan-cancer classification and biomarker prediction.	AUC, Accuracy, F1-Score	Shows potential for strong, broad generalization when trained on extensive, diverse data (287k slides from 3 centers).
Traditional CNN (HNSCC) [64]	CNN for HNSCC classification achieved 89.9% accuracy on unseen data; segmentation IoU of 0.782 for tumor tissue. Explanations aligned with pathological features.	Accuracy, Intersection over Union (IoU)	Demonstrates that well-validated, task-specific CNNs can achieve high, reliable performance for focused applications.

Experimental Protocols for Cross-Center Validation

Rigorous experimental design is fundamental to producing clinically meaningful evidence for AI models in pathology. The following protocols outline key methodologies for assessing diagnostic accuracy and robustness.

Protocol for Robustness and Site Bias Assessment

Objective: To evaluate whether a model's embeddings cluster by biological class rather than by site-specific technical artifacts. Methodology:

• Dataset Curation: Assemble a test set comprising whole-slide images from multiple independent medical centers (minimum 3-5 centers). The dataset should include samples from the same cancer types/subtypes across all centers.
• Embedding Generation: Process all WSIs through the model to extract feature embeddings. For FMs, use the frozen feature extractor. For CNNs, use features from the penultimate layer.
• Similarity Calculation: For each sample, compute two similarity scores: (a) average similarity to all samples of the same cancer class (within-class similarity), and (b) average similarity to all samples from the same medical center (within-center similarity).
• Robustness Index (RI) Computation: Calculate the Robustness Index as the ratio of the mean within-class similarity to the mean within-center similarity. An RI > 1 indicates that biological signal dominates site-specific bias, which is desirable [19].
• Statistical Analysis: Perform statistical testing (e.g., t-tests) to confirm significant differences between within-class and within-center similarities.

Protocol for Geometric Fragility Assessment

Objective: To measure a model's invariance to basic geometric transformations, a proxy for robustness to common image variations. Methodology:

• Test Set Creation: Select a representative set of image patches from multiple centers.
• Transformation Application: Apply a series of rotations (e.g., in 15° increments from 0° to 360°) to each patch, generating a transformed test set.
• Embedding Extraction: Pass both original and transformed patches through the model to obtain embeddings.
• Invariance Metric Calculation:
  • Compute the mean mutual k-nearest neighbors (m-kNN) score between original and transformed embeddings. A higher m-kNN (closer to 1) indicates better invariance [19].
  • Compute the mean cosine distance between original and transformed embeddings. A lower cosine distance (closer to 0) indicates better alignment [19].
• Comparative Analysis: Compare models based on these metrics. Studies show that models trained with explicit rotation augmentation significantly outperform those without (p < 0.0001) [19].

Protocol for Cross-Dataset Diagnostic Accuracy Evaluation

Objective: To assess model performance degradation when applied to entirely external datasets not seen during training. Methodology:

• Dataset Selection: Secure at least one internal dataset (for development and initial validation) and two or more fully external datasets from different institutions for testing.
• Task Definition: Define clear diagnostic tasks (e.g., tumor detection, cancer subtyping, Gleason grading).
• Model Configuration:
  • For Foundation Models: Employ linear probing (training a shallow classifier on frozen features) or full fine-tuning based on available computational resources and data [19].
  • For Traditional CNNs: Conduct end-to-end training on the internal dataset.
• Evaluation: Apply the trained models to the external test sets and compute standard diagnostic accuracy metrics (AUC, F1-score, sensitivity, specificity).
• Analysis: Report performance on each external dataset separately. Significant performance drops indicate poor generalizability. The STARD-AI reporting guideline should be followed to ensure comprehensive transparency [67].

Diagram 1: Experimental Workflow for Cross-Center Validation. This workflow outlines the key phases for designing a robust validation study for pathology AI models, from data curation to final analysis.

Table 3: Key Research Reagent Solutions for Validation Studies

Resource Category	Specific Examples & Functions	Relevance to Validation
Public WSI Repositories	The Cancer Genome Atlas (TCGA): Provides large-scale, multi-center WSI data for various cancers.	Serves as a crucial source of external test data for cross-dataset validation [19] [66] [65].
Annotation Software	QuPath [64], Aperio ImageScope [65]: Open-source and commercial tools for precise manual annotation of WSIs by pathologists.	Essential for generating high-quality ground truth labels for model training and evaluation.
Pre-trained Models	UNI, GigaPath, Virchow, Phikon, PathDino (FMs) [19]; EfficientNet, ResNet (CNNs) [64].	Provide baseline models and feature extractors for comparative benchmarking studies.
Computational Pathology Frameworks	PathML [64]: A Python library for preprocessing and tile extraction from WSIs.	Standardizes WSI preprocessing, a critical step for ensuring reproducible and comparable results across studies.
Reporting Guidelines	STARD-AI [67]: A tailored checklist for reporting diagnostic accuracy studies of AI models.	Ensures study transparency, reproducibility, and allows for critical appraisal of potential biases.

Visualization of Key Robustness Concepts

Diagram 2: Key Factors Influencing Model Robustness. This diagram conceptualizes the primary factors that contribute to the development of a biologically robust model, as opposed to one that is confounded by technical artifacts.

The rigorous validation of diagnostic accuracy and robustness across datasets and centers is a non-negotiable prerequisite for the clinical translation of AI in pathology. Empirical evidence to date presents a nuanced picture: while foundation models trained on enormous datasets show remarkable potential for broad generalization [66], they currently face significant challenges including susceptibility to site-specific bias [19], geometric fragility [19], and substantial computational demands [19]. Conversely, traditional CNNs, while more interpretable and efficient for focused tasks, struggle with generalization beyond their training domains and create annotation bottlenecks [64] [65].

The critical differentiator for foundation models to fulfill their promise is not merely scale, but the deliberate incorporation of robustness as a core design principle. This includes training on intentionally diverse multi-center data, employing targeted augmentations, developing domain-specific architectures, and adhering to rigorous validation protocols and reporting standards like STARD-AI [67]. For researchers and drug developers, the choice between model paradigms must be guided by the specific clinical context, weighing the need for broad adaptability against the requirements for a focused, efficient, and interpretable solution. The future of reliable AI in pathology depends on a validation ethos that prioritizes demonstrable robustness alongside diagnostic accuracy.

Foundation Models (FMs) represent a paradigm shift in computational pathology, moving from the task-specific, data-hungry nature of traditional Convolutional Neural Networks (CNNs) towards general-purpose models capable of operation in low-data regimes. While FMs promise significant advantages through zero-shot (performing tasks without explicit training) and few-shot (learning from minimal examples) capabilities, their real-world performance in pathology reveals both transformative potential and critical limitations. Empirical evidence indicates that current pathology FMs often underperform simpler methods in zero-shot settings and exhibit vulnerabilities including low diagnostic accuracy, poor robustness to site-specific variations, and unexpected sensitivity to minor image perturbations [68] [19]. Success in this domain appears contingent on moving beyond direct architectural transfers from natural image processing and instead developing domain-specific innovations that incorporate uncertainty quantification and ensemble methods to enhance reliability and generalizability across diverse clinical settings [26].

Traditional CNNs have fundamentally reshaped computational pathology by automating the analysis of Whole-Slide Images (WSIs). These models are typically trained in a supervised manner on large, meticulously labeled datasets to perform specific tasks such as tissue classification, nuclei segmentation, or cancer detection [1]. While achieving strong performance on their designated tasks, this approach suffers from significant limitations: it requires enormous annotated datasets (often containing thousands to millions of examples), demonstrates poor transferability to new tasks without extensive retraining, and struggles with generalizability across institutions due to variations in slide preparation, staining, and scanning protocols [19] [1].

Foundation Models propose a fundamentally different approach. Inspired by breakthroughs in natural language processing, FMs are large-scale models pre-trained on vast, diverse datasets often using self-supervised learning objectives that do not require manual labels [69] [19]. The core premise is that this pre-training phase allows the model to learn universal, robust representations of histopathological data—capturing everything from basic tissue structures to complex morphological patterns. For downstream applications, these models can then be adapted with minimal task-specific data, either through zero-shot inference (using the model directly without any further training), few-shot learning (providing a small number of examples), or full fine-tuning [69] [70]. This shift promises to reduce dependency on scarce annotated data and create more adaptable, general-purpose diagnostic tools.

Quantitative Performance Comparison of FMs vs. Traditional Methods

Rigorous evaluation of FMs in zero-shot and few-shot settings reveals a complex performance landscape, where these models do not consistently outperform established, often simpler, methodologies.

Zero-Shot Performance Benchmarks

In a critical zero-shot evaluation of single-cell FMs (scGPT and Geneformer), models were tested on cell type clustering and batch integration tasks without any further task-specific training. The results demonstrated that these FMs were frequently outperformed by established baseline methods, as shown in Table 1 [68].

Table 1: Zero-shot performance comparison (Average BIO score for cell type clustering). Higher scores are better. Adapted from [68].

Model/Method	Pancreas Dataset	PBMC (12k) Dataset	Tabula Sapiens Dataset
scGPT	Underperformed baselines	Comparable to scVI	Underperformed Harmony
Geneformer	Underperformed baselines	Underperformed baselines	Underperformed baselines
HVG (Baseline)	Best Performance	Best Performance	Best Performance
scVI (Baseline)	outperformed FMs	outperformed Geneformer	outperformed FMs
Harmony (Baseline)	outperformed FMs	outperformed Geneformer	Best Performance

In pathology imaging, a large-scale zero-shot retrieval study of FMs (UNI, GigaPath, Virchow) on 11,444 WSIs from 23 organs and 117 cancer subtypes reported only modest macro-averaged F1 scores of 40–42% for top-5 retrieval, with extreme performance variability across organs (e.g., 68% in kidneys vs. 21% in lungs) [19]. This suggests that current FMs may capture organ-specific texture patterns rather than generalizable diagnostic morphology.

Few-Shot and Fine-Tuning Performance

The performance gap often narrows significantly when FMs are provided with a limited number of examples. In a direct comparison on a natural language processing (NLP) task (entity extraction from tweets), few-shot learning dramatically outperformed zero-shot learning, while fine-tuning yielded mixed results, as summarized in Table 2 [70].

Table 2: Comparison of learning techniques on an NLP entity extraction task [70].

Technique	Data Input	Reported Accuracy	Key Takeaway
Zero-Shot Learning	No examples; only task instruction	19%	Highly sensitive to prompt phrasing; provides a performance baseline.
Few-Shot Learning	Task instruction + ~100 examples	97%	Dramatic performance gain by providing contextual examples.
Fine-Tuning	~100 examples for weight updates	91%	Can be outperformed by few-shot; justification depends on scale of use.

However, in histopathology, fine-tuning is often unstable. Studies indicate that FMs are frequently too large and memory-intensive for effective fine-tuning on typical clinical datasets, leading practitioners to rely predominantly on linear probing (training a shallow classifier on frozen FM embeddings)—a pragmatic retreat from the FM paradigm's promise of easy adaptation [19].

Experimental Protocols for Evaluating FM Capabilities

To rigorously assess the zero-shot and few-shot capabilities of FMs, researchers employ specific, standardized experimental protocols.

Zero-Shot Evaluation Protocol

Objective: To determine if a pre-trained FM can correctly perform a task without any exposure to labeled examples from that specific task.

Procedure:

• Model Acquisition: Obtain a pre-trained FM (e.g., scGPT, Geneformer, Virchow) without performing any additional training.
• Task Formulation: Define the downstream task (e.g., cell type clustering, cancer subtype classification) and encode it into a format the FM can process. For vision-language models, this involves creating a text prompt (e.g., "Is this tissue glioblastoma or lymphoma?").
• Unseen-Class Evaluation: Ensure the test dataset contains categories (e.g., cell types, cancer subtypes) that were not present in the model's pre-training data to test true generalization [71] [68].
• Inference & Analysis: Feed the test data to the model and analyze its raw outputs.
  • For clustering tasks, analyze the resulting embeddings using metrics like Average BIO score or Average Silhouette Width (ASW) [68].
  • For classification, assess accuracy, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUROC).

Key Consideration: This protocol tests the inherent knowledge and generalization capability acquired during pre-training. Performance is highly dependent on the alignment between the pre-training data distribution and the target task [68] [72].

Few-Shot Learning Evaluation Protocol

Objective: To measure a model's ability to rapidly learn a new task from a very small number of labeled examples.

Procedure:

• Support Set Curation: Select a small number K of labeled examples (e.g., K=1 for one-shot, K=5-100 for few-shot) per class from the training dataset. This is the "support set."
• Prompt Construction: For LLMs and vision-language models, construct a prompt that includes:
  • Clear task instruction.
  • The K labeled examples from the support set [69] [70].
  • The target (query) sample to be classified.
• In-Context Learning: The model processes the entire prompt and generates a prediction for the query sample based on the contextual patterns learned from the support set without updating its internal weights [69].
• Episodic Evaluation: Repeat the evaluation across many randomly sampled support-query sets (episodes) to obtain a statistically reliable performance estimate (e.g., mean accuracy and standard deviation) [71].

Key Consideration: The quality and diversity of the few examples are more critical than quantity. Performance is also sensitive to the order and phrasing of examples within the prompt [71] [70].

Methodologies and Visualization of FM Workflows

Successful implementation of FMs in pathology, particularly for critical diagnostics, requires sophisticated workflows that address their inherent uncertainties.

Diagram 1: A comparison of a traditional CNN workflow versus an uncertainty-aware Foundation Model workflow for pathology image analysis. The FM pathway emphasizes ensemble methods and uncertainty quantification to improve reliability.

The Scientist's Toolkit: Key Research Reagents and Materials

Table 3: Essential computational reagents and frameworks for evaluating FMs in pathology.

Reagent / Framework	Type / Category	Primary Function in Research
scGPT / Geneformer	Single-Cell Foundation Model	Pre-trained model for zero-shot evaluation of cell type clustering and batch integration [68].
UNI / Virchow / CTransPath	Histopathology Foundation Model	Pre-trained vision transformer for extracting general-purpose feature embeddings from whole-slide images [26].
PICTURE Framework	Uncertainty-Aware Ensemble System	Integrates multiple FMs with Bayesian inference and normalizing flow to quantify prediction certainty and flag out-of-distribution samples [26].
ACT Rule (W3C)	Accessibility / QC Standard	Defines quantitative thresholds (e.g., 4.5:1 contrast ratio) for evaluating visual output, ensuring interpretability of diagrams and interfaces [73].
CLAHE Algorithm	Image Pre-processing Tool	Contrast Limited Adaptive Histogram Equalization; standardizes image contrast in WSIs to mitigate stain variability during pre-processing [1].

Discussion: Challenges and Future Directions

The initial promise of FMs in pathology is tempered by significant empirical challenges. Beyond the performance metrics, core conceptual issues include:

• Ineffective Self-Supervision: Many FMs use pre-training objectives (e.g., masked image modeling) developed for natural images containing discrete objects. These assumptions fail in histopathology, where patches are complex and lack clear "objects," causing models to learn superficial stain textures rather than biologically relevant patterns [19].
• Geometric Fragility and Security: Transformer-based FMs lack inherent inductive biases for spatial invariance, making them surprisingly fragile. Studies show their embeddings can be severely disrupted by small rotations or universal adversarial perturbations (UTAPs), raising safety concerns for clinical deployment [19].
• Data Scarcity and Design Flaws: Despite their size, FMs may not have seen enough diverse, high-quality histopathology data. A more fundamental issue may be the mismatch between standard patch-based processing and the multi-scale, contextual nature of tissue morphology [19].

Future progress hinges on domain-specific innovation, not merely architectural transfer. The success of the PICTURE system demonstrates the value of integrating uncertainty quantification, model ensembles, and out-of-distribution detection to build reliable systems [26]. Furthermore, developing smaller, more data-efficient architectures that respect the hierarchical structure of tissues may prove more fruitful than relentless scaling, aligning with Occam's razor principle for model design in the complex domain of human pathology [19].

The field of computational pathology is undergoing a significant transformation with the emergence of foundation models (FMs). These large-scale models, pre-trained on massive datasets, promise unprecedented generalization across diverse tasks through adaptation rather than training from scratch [74]. Concurrently, traditional Convolutional Neural Networks (CNNs) remain a bedrock technology in medical image analysis. Within this evolving landscape, a critical question arises: how do foundation models truly differ from traditional CNNs in pathology research, and does the newer paradigm always supersede the older? This article argues that despite the compelling potential of FMs, task-specific CNNs frequently maintain competitive advantage or even superior performance in well-defined, data-rich pathological applications, particularly when computational efficiency, robustness, and clinical deployability are paramount.

The fundamental differences between these approaches are both architectural and philosophical. CNNs represent a mature, task-specific paradigm where models are typically designed and trained for a single objective, such as cancer grading or biomarker prediction [38]. Their architecture incorporates innate inductive biases for image processing, utilizing convolutional layers to hierarchically extract local features from pixels. In contrast, foundation models embrace a "pretrain-then-adapt" philosophy, seeking to create general-purpose feature extractors from broad data that can later be specialized for downstream tasks [75]. While this offers theoretical flexibility, evidence suggests that for many concrete pathology applications, this generality comes at a cost—including massive computational overhead, unexpected fragility, and performance that fails to exceed carefully crafted task-specific solutions [19] [3].

Fundamental Differences: CNNs vs. Foundation Models in Pathology

Architectural and Methodological Distinctions

Understanding the performance characteristics of CNNs versus foundation models requires examining their fundamental architectural and methodological differences. Convolutional Neural Networks are specialized by design for processing pixel data through localized filter operations that capture hierarchical patterns—from edges and textures in early layers to complex morphological features in deeper layers. This architecture embodies a powerful translation invariance inductive bias highly suited to tissue morphology analysis. In pathology, CNNs typically operate as task-specific models trained end-to-end on labeled datasets for objectives like tumor classification, segmentation, or survival prediction [38]. Their training is deterministic, resource-efficient, and produces optimized solutions for narrow problem domains.

In contrast, pathology foundation models represent a paradigm shift toward general-purpose feature extraction. These models, often based on transformer architectures or hybrid designs, are first pre-trained on massive, diverse datasets of histopathology images—sometimes encompassing millions of tissue patches from dozens of organs [3]. This self-supervised pre-training phase aims to create a comprehensive representation of tissue morphology without explicit labels. The resulting model then serves as a feature extractor that can be adapted to various downstream tasks through fine-tuning or linear probing [19]. While this approach theoretically enables broader generalization, it introduces significant complexity in data curation, computational requirements, and adaptation stability.

Table 1: Fundamental Characteristics of CNNs vs. Foundation Models in Pathology

Characteristic	Convolutional Neural Networks (CNNs)	Pathology Foundation Models
Core Architecture	Convolutional layers with local connectivity	Transformers or hybrid architectures with self-attention
Training Paradigm	Supervised learning on task-specific labeled data	Self-supervised pre-training followed by downstream adaptation
Model Scale	Medium to large (thousands to millions of parameters)	Very large (millions to billions of parameters)
Primary Strengths	Computational efficiency, stability, interpretability for specific tasks	Potential for broad generalization, transfer learning across tasks
Typical Applications	Classification, segmentation, detection in well-defined domains	Multi-task learning, few-shot adaptation, multimodal integration

Practical Implementation Differences

The practical implications of these architectural differences manifest throughout the model development lifecycle. CNN workflows are streamlined and deterministic—researchers curate labeled datasets for specific tasks, select appropriate architectures (EfficientNet, ResNet, U-Net variants), and train models with predictable computational requirements [76] [64]. This approach delivers high-performance, deployable models for focused applications but requires new data annotation and training for each distinct task.

Foundation model workflows introduce additional complexity through their two-phase adaptation process. The initial pre-training phase demands extraordinary computational resources—one study noted that FMs consumed up to 35× more energy than task-specific models [19]. Downstream adaptation often relies on linear probing (training only a final classification layer) rather than full fine-tuning, as the massive models prove unstable when further trained on typical pathology datasets of hundreds to thousands of slides [19]. This limitation directly contradicts the foundational premise of FMs as adaptable bases, instead reducing them to static feature extractors.

Evidence of Task-Specific CNN Superiority

Performance Benchmarks in Diagnostic Applications

Empirical evidence from rigorous comparative studies demonstrates that well-designed CNNs frequently match or exceed foundation model performance on specific pathology tasks. In breast cancer histopathology, a comprehensive analysis of 14 deep learning models—including both CNN-based and Transformer-based architectures—revealed that CNN-based models like ResNet50, RegNet, and ConvNeXT achieved perfect AUC scores of 0.999 in binary classification tasks, performing equivalently to the foundation model UNI [3]. The best overall performance was achieved by ConvNeXT, which attained an accuracy of 99.2%, specificity of 99.6%, and F1-score of 99.1% on the BreakHis dataset for breast cancer classification [3].

Similarly, in invasive ductal carcinoma (IDC) grading, traditional CNNs demonstrated exceptional capability without foundation model complexity. A systematic comparison of seven CNN architectures found that EfficientNetV2B0-21k outperformed other models with a balanced accuracy of 0.9666, while practically all selected CNNs performed well with an average balanced accuracy of 0.936 ± 0.0189 on the cross-validation set and 0.9308 ± 0.0211 on the test set [76]. These results highlight that for well-defined histological grading tasks, carefully optimized CNNs deliver state-of-the-art performance without the overhead of massive foundation models.

Table 2: Performance Comparison of CNN vs. Foundation Model Architectures on Pathology Tasks

Task	Best Performing CNN	CNN Performance	Best Performing FM	FM Performance
Breast Cancer Classification	ConvNeXT [3]	Accuracy: 99.2%, AUC: 0.999	UNI (fine-tuned) [3]	Accuracy: 95.5%, AUC: 0.998
IDC Grading	EfficientNetV2B0-21k [76]	Balanced Accuracy: 0.9666	Not assessed in study	-
Head & Neck Cancer Classification	EfficientNet-B0 [64]	Accuracy: 89.9%	Not assessed in study	-
Prostate Cancer Diagnosis	Task-specific end-to-end model [19]	Matched or outperformed FMs	Multiple FMs (UNI, GigaPath, Virchow) [19]	Modest performance (40-42% F1)
CNS Tumor Diagnosis	Not applicable	-	PICTURE (Ensemble of 9 FMs) [26]	AUROC: 0.989

Resource Efficiency and Computational Demands

The computational resource disparity between CNNs and foundation models represents a critical practical consideration for pathology laboratories and research institutions. A large-scale study utilizing over 100,000 prostate biopsy slides revealed that foundation models consumed up to 35× more energy than task-specific models without delivering substantial performance improvements [19]. This finding challenges the scalability and sustainability of FM approaches, particularly in clinical settings where computational resources are often constrained.

Furthermore, the adaptation of foundation models frequently proves impractical due to their memory-intensive architecture and fine-tuning instability. Most pathology FMs cannot be effectively fine-tuned on datasets of typical size (hundreds to a few thousand slides), forcing researchers to rely on linear probing rather than full model adaptation [19]. This limitation fundamentally undermines the core promise of foundation models as adaptable bases for diverse applications, instead reducing them to static feature extractors that cannot benefit from additional task-specific training.

Experimental Protocols: Validating CNN Performance

Protocol for Histopathological Image Classification with CNNs

The experimental methodology for achieving state-of-the-art performance with CNNs in pathology applications involves several critical phases, from data preparation through model optimization. The following protocol outlines the key steps, drawing from successful implementations in recent literature [76] [3] [64]:

Data Preparation and Annotation:

• Obtain whole-slide images (WSIs) from histopathology samples, typically stained with hematoxylin and eosin (H&E)
• Manually annotate tissue regions under supervision of experienced pathologists (e.g., tumor vs. non-tumor regions)
• Divide WSIs into non-overlapping tiles of optimal size (determined via grid search, typically 99.6-199.1μm)
• Address class imbalance through techniques such as majority class undersampling
• Apply stain normalization to mitigate variability in H&E staining across different laboratories

Data Preprocessing and Augmentation:

• Extract tiles with minimum annotated pixel threshold (e.g., ≥30% annotated pixels)
• Scale pixel intensity to 0-1 range and standardize using mean and standard deviation of training data
• Implement extensive data augmentation including rotation, mirroring, and variations in hue, saturation, brightness, and contrast
• Apply additional transformations such as blurring and additive Gaussian noise for classification tasks
• Use libraries such as Albumentations for transforming tiles and masks jointly

Model Architecture and Training:

• Select appropriate CNN backbone (efficient variants like EfficientNet-B0 often optimal)
• Modify architecture by adding global average pooling, dense layers (1024 neurons), dropout, and final classification layer
• Utilize transfer learning from ImageNet pre-trained weights
• Train with Adam optimizer with learning rate of 10^-5 for classification or 10^-4 for segmentation
• Employ appropriate loss functions (binary cross-entropy for classification, Jaccard loss for segmentation)

Validation and Interpretation:

• Implement cross-validation (e.g., 5-fold stratified) for robust performance estimation
• Apply Explainable AI techniques (Grad-CAM, HR-CAM) to visualize important features
• Validate that models rely on histopathologically relevant features aligning with pathologist expertise

Protocol for Foundation Model Adaptation

The adaptation of pathology foundation models follows a distinctly different pathway focused on leveraging pre-trained features rather than end-to-end training:

Feature Extraction Phase:

• Select appropriate foundation model (UNI, GigaPath, Virchow, CTransPath, etc.)
• Extract features from pre-trained model using frozen weights
• Process whole-slide images through model's embedding pipeline

Downstream Adaptation:

• For data-scarce settings: Use linear probing (training only final classification layer)
• For data-rich settings: Attempt full fine-tuning with caution due to instability risks
• Employ ensemble methods combining multiple FMs with uncertainty quantification

Validation and Robustness Testing:

• Test on multi-center datasets to assess cross-site generalization
• Evaluate robustness to image variations (staining, scanning differences)
• Implement out-of-distribution detection to identify atypical cases

Successful implementation of CNN-based pathology solutions requires both wet-lab and computational components. The following table details essential resources referenced in the experimental protocols:

Table 3: Essential Research Reagents and Computational Tools for Pathology AI

Category	Item/Resource	Specification/Function	Example Use Case
Wet-Lab Reagents	Hematoxylin & Eosin	Standard histological staining for tissue morphology	Routine H&E staining of biopsy specimens
Tissue Processing	Formalin Fixation & Paraffin Embedding	Tissue preservation and sectioning	Preparing FFPE tissue blocks for sectioning
Annotation Software	QuPath [64]	Open-source digital pathology analysis	Manual annotation of tumor regions in WSIs
Programming Environment	Python with TensorFlow/Keras [76] [64]	Deep learning framework for model development	Implementing and training CNN architectures
Medical Image Processing	PathML [64]	Python library for whole-slide image analysis	Extracting non-overlapping tiles from WSIs
Data Augmentation	Albumentations [64]	Image augmentation library for machine learning	Applying stain variations and transformations
Pre-trained Models	TensorFlow Hub modules [76]	Repository of pre-trained CNN architectures	Transfer learning with EfficientNet variants
Explainable AI	Grad-CAM, HR-CAM [64]	Visual explanation methods for CNN decisions	Interpreting model focus areas in tissue tiles

Critical Limitations of Foundation Models in Pathology Applications

Performance and Robustness Concerns

Despite their theoretical promise, pathology foundation models exhibit several critical limitations in practical applications. Systematic evaluations reveal fundamental weaknesses including low diagnostic accuracy, poor robustness, geometric instability, and concerning safety vulnerabilities [19]. When evaluated on 11,444 whole-slide images from 23 organs and 117 cancer subtypes, leading pathology FMs achieved only modest performance with macro-averaged F1 scores around 40-42% for top-5 retrieval, with pronounced organ-level variability [19].

A particularly concerning limitation is the lack of robustness and significant site bias exhibited by most pathology FMs. When evaluated across multiple institutions, only one model (Virchow) achieved a Robustness Index (RI) >1.2, indicating that biological structure dominated site-specific bias, while all others had RI<1, meaning embeddings grouped primarily by hospital or scanner rather than cancer type [19]. This fundamental fragility presents a major barrier to clinical deployment where models must perform consistently across diverse healthcare settings.

Geometric Fragility and Security Vulnerabilities

Foundation models based on transformer architectures demonstrate significant geometric fragility, with latent representations that fail to remain stable under basic image transformations like rotation [19]. Studies evaluating rotational invariance found that performance varied dramatically across models, with only those explicitly trained with rotation augmentation achieving acceptable invariance [19]. This sensitivity to basic image transformations raises concerns about real-world reliability.

Additionally, pathology FMs show alarming security and safety vulnerabilities. Research has demonstrated that universal and transferable adversarial perturbations (UTAP)—imperceptible noise patterns—can collapse FM embeddings across architectures, degrading accuracy from ~97% to as low as 12% [19]. These attacks represent not only security risks but also diagnostic stress tests that reveal sensitivity to minor pixel-level variations that routinely occur in laboratory workflows due to differences in H&E staining, scanner optics, compression artifacts, and slide preparation imperfections.

The evidence clearly indicates that task-specific CNNs maintain significant relevance in computational pathology despite the emergence of foundation models. For well-defined applications with sufficient training data, CNNs provide superior or equivalent performance with dramatically better computational efficiency and more straightforward implementation. Their architectural biases align well with histopathological image analysis, and their operational characteristics suit clinical deployment constraints.

However, foundation models offer complementary strengths in scenarios requiring broad generalization across multiple tissue types or adaptation to tasks with minimal labeled data. The most effective future path likely involves strategic integration of both approaches—leveraging CNNs for optimized performance on specific high-value tasks while utilizing FMs for exploratory analysis, rare conditions, and multimodal integration. As the field advances, hybrid approaches that combine the efficiency and reliability of CNNs with the adaptability of foundation models may offer the most promising direction for realizing the full potential of AI in pathology.

This nuanced understanding enables researchers and clinicians to make informed decisions about model selection based on specific use cases, resource constraints, and performance requirements, ultimately advancing the field toward more effective and deployable computational pathology solutions.

The transition of artificial intelligence (AI) models from research environments to clinical practice represents a formidable challenge in computational pathology. A model's performance on internal validation data often provides an optimistic estimate that fails to predict real-world effectiveness, as performance frequently degrades when applied to new patient populations, different imaging equipment, or varied laboratory protocols. This challenge is particularly acute when comparing traditional convolutional neural networks (CNNs) with emerging foundation models, as each employs fundamentally different approaches to achieving generalization. Traditional CNNs, typically trained on specific, limited datasets for narrow tasks, often face significant performance drops due to domain shift—variations in staining protocols, scanner differences, and population heterogeneity. In contrast, pathology foundation models (PFMs), pre-trained on massive, diverse datasets of histopathological images, theoretically offer better generalization through their learned, transferable representations of tissue morphology. However, recent evidence suggests that PFMs exhibit their own unique limitations, including site-specific bias, geometric fragility, and unexpected vulnerability to minor image perturbations that mimic real-world laboratory variations.

The clinical stakes for achieving robust generalization are substantial. In neuropathology, distinguishing between glioblastoma and primary central nervous system lymphoma (PCNSL) directly determines whether a patient undergoes surgical resection or receives chemotherapy and radiotherapy, yet these entities share overlapping pathology features that challenge accurate diagnosis [77]. Similarly, in molecular diagnostics, AI models that can predict EGFR mutation status directly from H&E-stained lung adenocarcinoma slides could potentially reduce the need for additional tissue-consuming molecular tests by up to 43%, preserving precious tissue for comprehensive genomic sequencing [78]. This technical guide provides a comprehensive framework for assessing the generalizability of both traditional and foundation models through rigorous external validation, offering detailed experimental protocols and analytical methods to evaluate clinical readiness for unseen patient populations.

Quantitative Landscape of Model Generalization

Table 1: Performance Comparison Between Traditional CNNs and Foundation Models

Model Type	Representative Model	Internal Performance (AUC)	External Performance (AUC)	Performance Drop	Clinical Context
Traditional CNN	Custom Ensemble (PICTURE)	0.989 [77]	0.924-0.996 [77]	Minimal	CNS tumor diagnosis
Foundation Model	EAGLE (Fine-tuned)	0.847 [78]	0.870 [78]	Improved	Lung cancer EGFR detection
Foundation Model	BEPH	N/A	0.994 (RCC) [79]	N/A	Multi-cancer classification
Traditional CNN	Breast Metastasis Detection	0.969 [80]	0.929 (sentinel nodes) [80]	-4.1%	Breast cancer metastasis detection
Traditional CNN	Otoscopy Model	0.95 [81]	0.76 [81]	-20%	Middle ear disease detection

Table 2: Domain Robustness Assessment of Pathology Foundation Models

Foundation Model	Robustness Index (RI)	Top-5 Retrieval F1 Score	Rotation Invariance	Energy Consumption vs. TS Models
UNI	<0.9 [19]	~40-42% [19]	Not reported	Not reported
Virchow	>1.2 [19]	~40-42% [19]	Lowest m-kNN (0.53) [19]	Not reported
GigaPath	<1 [19]	~40-42% [19]	Not reported	Not reported
PathDino	Not reported	Not reported	Highest m-kNN (0.85) [19]	Not reported
Phikon-v2	≈0.7 [19]	~40-42% [19]	Highest cosine distance (≈0.145) [19]	Not reported
General Trend	Most <1 [19]	Modest [19]	Variable [19]	Up to 35× more [19]

Methodologies for External Validation

Multi-Center Study Designs

Rigorous external validation requires testing models on completely independent datasets that capture the full spectrum of real-world variability. The PICTURE system for central nervous system tumor diagnosis exemplifies this approach, having been validated on five independent international cohorts from the Mayo Clinic, Hospital of the University of Pennsylvania, Brigham and Women's Hospital, Medical University of Vienna, and Taipei Veterans General Hospital [77]. This design intentionally incorporates variability in slide preparation protocols, staining techniques, scanner models, and patient demographics to assess true generalizability. Similarly, the EAGLE model for EGFR mutation prediction in lung cancer was validated on external test cohorts from Mount Sinai Health System, Sahlgrenska University Hospital, Technical University of Munich, and The Cancer Genome Atlas, demonstrating consistent performance with an overall AUC of 0.870 across 1,484 external slides [78].

For traditional CNNs, performance degradation often becomes apparent only through such comprehensive external validation. In otoscopy, models achieving internal AUCs of 0.95 dropped to 0.76 when applied to external images, highlighting the domain gap between different clinical environments [81]. The most significant declines occurred when models trained on sentinel lymph node data were applied to axillary dissection nodes—a seemingly minor change in surgical indication that resulted in a 40% reduction in detection performance (FROC score dropping from 0.838 to 0.503) [80].

Uncertainty Quantification and Out-of-Distribution Detection

A critical advancement in generalization assessment involves moving beyond simple performance metrics to evaluate a model's ability to quantify its own uncertainty and identify out-of-distribution samples. The PICTURE system employs Bayesian inference, deep ensemble methods, and normalizing flow techniques to estimate epistemic uncertainty in its predictions [77]. This uncertainty quantification serves as a reliability measure for pathologists, flagging cases where the model encounters unfamiliar patterns that may represent rare subtypes or technical artifacts.

This capability is particularly valuable for identifying samples belonging to categories not represented in the training data. In one study, the uncertainty-aware framework successfully identified 67 types of rare central nervous system cancers that were neither gliomas nor lymphomas, as well as normal brain tissues [77]. This approach prevents the dangerous scenario where conventional models provide overconfident but incorrect predictions for novel manifestations, potentially leading to diagnostic errors.

Domain Adaptation Techniques

When models exhibit performance degradation on external datasets, domain adaptation techniques can help bridge the generalization gap. The Adversarial fourIer-based Domain Adaptation (AIDA) framework addresses domain shift by leveraging frequency domain transformations to make models less sensitive to color variations (amplitude spectrum) and more attentive to structural features (phase-related components) [82]. This approach recognizes that while CNNs are highly sensitive to amplitude spectrum variations that often reflect staining differences rather than biological factors, human pathologists primarily rely on phase-related components for object recognition.

AIDA incorporates an FFT-Enhancer module that integrates frequency information into adversarial domain adaptation, significantly improving classification performance on multi-center datasets for ovarian, pleural, bladder, and breast cancers compared to conventional adversarial domain adaptation or color normalization techniques [82]. This frequency-based approach demonstrates that explicitly designing architectures to mimic human perceptual strengths can enhance generalization.

Experimental Protocols for Generalization Assessment

Protocol 1: Multi-Center External Validation

Objective: To evaluate model performance across independent institutions with different patient populations and technical protocols.

Materials:

• Model to be validated (traditional CNN or foundation model)
• Internal validation dataset (held-out from training)
• 3-5 external datasets from different institutions
• High-performance computing resources for inference

Procedure:

• Train model on source dataset with standard 80/20 train-validation split
• Evaluate model on internal validation set to establish baseline performance
• Apply trained model to each external dataset without any fine-tuning
• Calculate performance metrics (AUC, sensitivity, specificity, F1) for each external cohort
• Compare performance degradation across cohorts and analyze failure modes

Interpretation: Performance drops >10% in AUC indicate significant generalization issues. Consistent performance across external cohorts (<5% variation) suggests robust generalization [77] [78].

Protocol 2: Robustness Index Calculation

Objective: To quantify whether model embeddings cluster by biological class versus medical center.

Materials:

• Model embeddings from multiple institutions
• Metadata including center ID and diagnostic labels
• Computational environment for similarity calculations

Procedure:

• Extract embeddings for samples from at least 3 centers with known biological classes
• Compute within-class similarity (average similarity between samples sharing diagnosis)
• Compute within-center similarity (average similarity between samples from same center)
• Calculate Robustness Index (RI) = within-class similarity / within-center similarity
• Interpret results: RI > 1.2 indicates biological robustness, RI ≈ 1 indicates center confounding [19]

Interpretation: Models with RI < 1 are significantly influenced by site-specific biases rather than biological features, indicating poor generalization potential.

Protocol 3: Rotation Invariance Assessment

Objective: To evaluate geometric robustness of foundation models against image rotations.

Materials:

• Model to be evaluated
• Test patches from target domain (e.g., TCGA-KIRC)
• Rotation augmentation pipeline

Procedure:

• Extract representative image patches from target dataset
• Apply rotations in 15° increments from 0° to 360°
• Compute embeddings for original and rotated patches
• Calculate mean mutual k-nearest neighbors (m-kNN) between original and rotated embeddings
• Calculate mean cosine distance between original and rotated embeddings
• Compare results across models and training strategies [19]

Interpretation: Higher m-kNN scores (closer to 1) and lower cosine distances (closer to 0) indicate better rotation invariance. Models trained with explicit rotation augmentation typically outperform those without.

Visualization of Experimental Workflows

Multi-Center Validation Workflow

Domain Adaptation with AIDA Framework

The Scientist's Toolkit: Research Reagent Solutions

Table 3: Essential Research Materials for Generalization Studies

Reagent/Resource	Function	Example Implementation
Multi-Center Datasets	Assess cross-institutional performance	5+ independent cohorts with different protocols [77]
Uncertainty Quantification Framework	Measure prediction confidence	Bayesian deep ensembles, normalizing flows [77]
Adversarial Domain Adaptation	Bridge domain shifts between centers	AIDA framework with FFT-Enhancer [82]
Robustness Index (RI) Metric	Quantify biological vs. center clustering	RI = within-class similarity / within-center similarity [19]
Rotation Invariance Tests	Evaluate geometric robustness	m-kNN and cosine distance on rotated patches [19]
Stain Normalization Tools	Reduce color variation impact	Macenko method, cycleGAN-based normalization [82] [83]
Foundation Model Embeddings	Transferable feature representations	UNI, Virchow, Phikon, PathDino [19]
Computational Infrastructure	Handle large-scale validation	GPU clusters for foundation model inference [78] [79]

Discussion: Clinical Translation Pathways

The pathway to clinical implementation requires careful consideration of the trade-offs between traditional CNNs and foundation models. Traditional CNNs offer computational efficiency and can achieve excellent performance when training and deployment environments are well-matched, but they often require extensive retraining and domain adaptation when applied to new settings [80]. Foundation models provide better out-of-the-box performance on diverse datasets but come with substantial computational costs—consuming up to 35× more energy than task-specific models—and demonstrate unexpected vulnerabilities to minor image perturbations that mimic real-world laboratory variations [19].

Prospective silent trials represent a critical final step in validating clinical readiness. The EAGLE model for EGFR prediction underwent such a trial, achieving an AUC of 0.890 on prospective samples and demonstrating potential to reduce rapid molecular tests by 43% while maintaining clinical standards [78]. This approach provides the most realistic assessment of how a model will perform in actual clinical workflows before committing to formal implementation.

For both traditional and foundation models, continuous monitoring and periodic recalibration are essential for maintaining performance over time. The multi-center validation framework and uncertainty quantification methods described in this guide should not be viewed as one-time activities but as components of an ongoing quality assurance program that ensures models remain effective as clinical practices, imaging technologies, and patient populations evolve.

Robust external validation remains the critical gateway to clinical implementation for AI models in pathology. While foundation models offer theoretical advantages for generalization through their large-scale pre-training, current evidence indicates they still face significant challenges including site-specific bias, geometric fragility, and substantial computational demands. Traditional CNNs can achieve excellent performance in controlled environments but often require explicit domain adaptation strategies when deployed across multiple centers. The comprehensive validation frameworks, experimental protocols, and analytical methods presented in this guide provide researchers with the tools necessary to rigorously assess generalization to unseen patient populations, ultimately accelerating the translation of promising AI technologies from research environments to clinical practice where they can improve patient care.

Conclusion

The transition from CNNs to foundation models represents a fundamental paradigm shift in computational pathology, moving from specialized tools to versatile, general-purpose systems. While FMs demonstrate superior performance in complex tasks, low-data scenarios, and offer promising multimodal capabilities, they face significant challenges including computational cost, data hunger, and robustness issues that are less pronounced in traditional CNNs. The future of pathology AI lies not in a single approach, but in a synergistic ecosystem. This includes developing more efficient, domain-specific FMs, creating robust validation frameworks to ensure clinical reliability, and advancing towards a generalist medical AI that seamlessly integrates pathology with other data modalities to truly enable precision and personalized medicine. For researchers and drug developers, this evolution promises more powerful tools for biomarker discovery, patient stratification, and therapy response prediction.

References

• 1. Deep learning for digital pathology: A critical overview of ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12593639/]
• 2. Fusion of Vision Transformer and Convolutional Neural ... [https://link.springer.com/article/10.1007/s41314-025-00079-0]
• 3. Comparative analysis of convolutional neural networks and ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12209302/]
• 4. Transformer optimization with meta learning on pathology ... [https://www.nature.com/articles/s41746-025-01833-6]
• 5. A foundation model for clinical-grade computational ... [https://www.nature.com/articles/s41591-024-03141-0]
• 6. A multimodal whole-slide foundation model for pathology [https://www.nature.com/articles/s41591-025-03982-3]
• 7. A dual data stream hybrid neural network for classifying ... [https://www.sciencedirect.com/science/article/abs/pii/S0010482524006036]
• 8. Convolutional neural networks [https://www.pathologyoutlines.com/topic/informaticsconvnet.html]
• 9. Deep Learning Powers Cancer Diagnosis in Digital ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7902448/]
• 10. A narrative review of foundation models for medical image segmentation: zero-shot performance evaluation on diverse modalities - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC12209621/]
• 11. A clinical benchmark of public self-supervised pathology ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12003829/]
• 12. Self-supervised learning for data augmentation in ... [https://www.nature.com/articles/s41598-025-23280-9]
• 13. Current Pathology Foundation Models are unrobust to Medical Center Differences [https://arxiv.org/html/2501.18055v1]
• 14. Self-supervised Learning for Cancer Pathology Foundation ... [https://deeplearningpathology.org/]
• 15. Multiple instance learning using pathology foundation ... [https://www.nature.com/articles/s41598-025-19297-9]
• 16. Convolutional neural networks in medical image ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC7778711/]
• 17. Artificial intelligence and computational pathology [https://www.nature.com/articles/s41374-020-00514-0]
• 18. A whole-slide foundation model for digital pathology from ... [https://www.nature.com/articles/s41586-024-07441-w]
• 19. Why Foundation Models in Pathology Are Failing [https://arxiv.org/html/2510.23807v2]
• 20. Why Foundation Models in Pathology Are Failing [https://rewire.it/blog/why-foundation-models-in-pathology-are-failing-and-what-comes-next/]
• 21. State of Foundation Model Training Report 2025 [https://neptune.ai/state-of-foundation-model-training-report]
• 22. Computational pathology: A comprehensive review of ... [https://www.sciencedirect.com/science/article/pii/S2950261625000214]
• 23. A deep learning model to predict RNA-Seq expression of ... [https://www.nature.com/articles/s41467-020-17678-4]
• 24. Deep learning predicts gene expression as an intermediate ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC8138606/]
• 25. AI and Big Data in Healthcare: Opportunities and Challenges [https://ardigen.com/ai-and-big-data-in-healthcare-opportunities-and-challenges/]
• 26. Uncertainty-aware ensemble of foundation models ... [https://www.nature.com/articles/s41467-025-64249-6]
• 27. Deep learning quantifies pathologists' visual patterns for ... [https://www.nature.com/articles/s41467-025-60307-1]
• 28. SMMILe enables accurate spatial quantification in digital ... [https://www.nature.com/articles/s43018-025-01060-8]
• 29. [2504.17379] A Spatially-Aware Multiple Instance Learning ... [https://arxiv.org/abs/2504.17379]
• 30. AttriMIL: Revisiting attention-based multiple instance ... [https://www.sciencedirect.com/science/article/pii/S1361841525001781]
• 31. A quantitative approach to evaluate image quality of whole ... [https://www.sciencedirect.com/science/article/pii/S2153353922005740]
• 32. Current Trends and Future Directions of Digital Pathology ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12428701/]
• 33. Pan-Cancer Integrative Histology-Genomic Analysis via ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10397370/]
• 34. Multimodal integration of radiology, pathology and ... [https://www.nature.com/articles/s43018-022-00416-8]
• 35. Principles for valid histopathologic scoring in research - PMC [https://pmc.ncbi.nlm.nih.gov/articles/PMC3795863/]
• 36. and reproducibility of Quantitative scoring... comparison pathologist [https://diagnosticpathology.biomedcentral.com/articles/10.1186/s13000-016-0511-5]
• 37. Pathology Foundation Models [https://www.emergentmind.com/topics/pathology-foundation-models]
• 38. Computational pathology in precision oncology [https://pmc.ncbi.nlm.nih.gov/articles/PMC12634262/]
• 39. A clinical benchmark of public self-supervised pathology ... [https://www.nature.com/articles/s41467-025-58796-1]
• 40. Why Foundation Models in Pathology Are Failing [https://arxiv.org/html/2510.23807v1]
• 41. Deep learning approaches for pathological image ... [https://www.sciencedirect.com/science/article/abs/pii/S1349007925000854]
• 42. Unleashing the potential of digital pathology data ... [https://www.nature.com/articles/s41746-022-00635-4]
• 43. Computational pathology in the age of artificial intelligence [https://pmc.ncbi.nlm.nih.gov/articles/PMC12451823/]
• 44. From Microscope to Machine Learning: 3 Digital Pathology ... [https://proscia.com/from-microscope-to-machine-learning-3-digital-pathology-leaders-explore-the-path-to-precision-medicine/]
• 45. Analysis of 3D pathology samples using weakly supervised ... [https://www.sciencedirect.com/science/article/pii/S0092867424003519]
• 46. VLMs vs. CNNs: Is a New Era Dawning in Computer Vision ... [https://www.picsellia.com/post/vlms-vs-cnns-a-new-era-dawning-in-computer-vision-performance]
• 47. How to estimate carbon footprint when training deep ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC10661046/]
• 48. Measuring the Energy Consumption and Efficiency of Deep ... [https://arxiv.org/abs/2403.08151]
• 49. The Future of Foundation Model Business Models [https://www.generativevalue.com/p/the-future-of-foundation-model-business]
• 50. Domain Adaptation and Generalization Using Foundation ... [https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5345726]
• 51. Automatic data augmentation to improve generalization of ... [https://www.sciencedirect.com/science/article/pii/S0010482524001021]
• 52. Review of deep learning: concepts, CNN architectures ... [https://journalofbigdata.springeropen.com/articles/10.1186/s40537-021-00444-8]
• 53. Deep Convolutional Neural Networks Enable Discrimination ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC5828543/]
• 54. Large language models driven neural architecture search ... [https://www.nature.com/articles/s41746-025-02042-x]
• 55. A systematic review of generalization research in medical ... [https://www.sciencedirect.com/science/article/pii/S0010482524013416]
• 56. A comprehensive evaluation of histopathology foundation ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11782474/]
• 57. A Survey of Pathology Foundation Model [https://arxiv.org/html/2504.04045v1]
• 58. A visual-language foundation model for computational ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11384335/]
• 59. SimMIL: A Universal Weakly Supervised Pre-Training ... [https://arxiv.org/abs/2505.06710]
• 60. Pan-Cancer Histopathology WSI Pre-Training With Position ... [https://pubmed.ncbi.nlm.nih.gov/40030469/]
• 61. Lightweight self supervised learning framework for domain ... [https://www.nature.com/articles/s41598-025-20411-0]
• 62. Foundation Models in Digital Pathology: The Good, the Bad ... [https://beyondtheslide.substack.com/p/foundation-models-in-digital-pathology]
• 63. Digital Pathology and AI Highlights from ASCO 2025 [https://proscia.com/digital-pathology-and-ai-highlights-from-asco-2025/]
• 64. Explainable convolutional neural networks for assessing head ... [https://diagnosticpathology.biomedcentral.com/articles/10.1186/s13000-023-01407-8]
• 65. Convolutional neural networks can accurately distinguish ... [https://www.nature.com/articles/s41598-018-37638-9]
• 66. PathOrchestra: a comprehensive foundation model for ... [https://www.nature.com/articles/s41746-025-02027-w]
• 67. The STARD-AI reporting guideline for diagnostic accuracy ... [https://www.nature.com/articles/s41591-025-03953-8]
• 68. Zero-shot evaluation reveals limitations of single-cell ... [https://genomebiology.biomedcentral.com/articles/10.1186/s13059-025-03574-x]
• 69. Zero-Shot and Few-Shot Learning with Foundation Models [https://medium.com/@chaituapatil3/zero-shot-and-few-shot-learning-with-foundation-models-redefining-the-capabilities-of-ai-33b1bf3f203d]
• 70. Zero-Shot Learning vs. Few-Shot Learning vs. Fine-Tuning [https://labelbox.com/guides/zero-shot-learning-few-shot-learning-fine-tuning/]
• 71. Zero-Shot vs. One-Shot vs. Few-Shot Learning: How to Test? [https://testrigor.com/blog/zero-shot-vs-one-shot-vs-few-shot-learning-how-to-test/]
• 72. How Foundational are Foundation Models for Time Series ... [https://arxiv.org/html/2510.00742v3]
• 73. Text has enhanced contrast | ACT Rule | WAI [https://www.w3.org/WAI/standards-guidelines/act/rules/09o5cg/]
• 74. Revolutionizing Digital Pathology With the Power of ... [https://www.sciencedirect.com/science/article/pii/S0023683723001988]
• 75. Pathology Foundation Models - PMC - PubMed Central [https://pmc.ncbi.nlm.nih.gov/articles/PMC11799676/]
• 76. Performance analysis of seven Convolutional Neural ... [https://www.nature.com/articles/s41598-022-21848-3]
• 77. Uncertainty-aware ensemble of foundation models ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC12480093/]
• 78. Real-world deployment of a fine-tuned pathology ... [https://www.nature.com/articles/s41591-025-03780-x]
• 79. A foundation model for generalizable cancer diagnosis and ... [https://www.nature.com/articles/s41467-025-57587-y]
• 80. Generalization of Deep Learning in Digital Pathology [https://www.mdpi.com/2072-6694/14/21/5424]
• 81. Evaluating the generalizability of deep learning image ... [https://www.nature.com/articles/s41598-023-31921-0]
• 82. Learning generalizable AI models for multi-center ... [https://pmc.ncbi.nlm.nih.gov/articles/PMC11271637/]
• 83. Leveraging pathology foundation models for glioma ... [https://www.sciencedirect.com/science/article/pii/S001048252501340X]