Foundation Models in Computational Pathology: A Review of Challenges, Opportunities, and Impact

• Clinical translation is threatened by biased or insufficiently diverse data, opaque decisions, poor population generalization, and inadequate validation of diagnostic subtleties.The review emphasizes balancing innovation with the clinical safety and efficacy imperatives identified by Bommasani et al. 2022.Source: §1 · p. 2
• Foundation models could unify pathology, oncology, and genomic knowledge to support disease diagnosis and treatment-response prediction across diverse tasks.Their broader synthesis may exceed traditional task-specific approaches, although their ability to match specialist understanding remains uncertain.Source: §1 · pp. 1–2
• The field is shifting from task-specific and multitask systems toward generative and task-agnostic models, raising concerns about precision, reasoning safety, and fair benchmarking.The central tension is adapting models to specialized pathology tasks without sacrificing broad applicability.Source: §1 · pp. 1–2
• Gigapixel whole-slide images require multi-scale contextual integration across spatial locations, while deployment also demands substantial storage, computing, digitization, and workflow infrastructure.These requirements affect both computational cost and model performance and create practical barriers to clinical adoption.Source: §1 · p. 2
• This review extends prior surveys by critically assessing foundation models’ opportunities, risks, clinical challenges, and broader impact alongside their tools, datasets, and training methods.Chanda et al. (Chanda et al. 2024) provide a technology-focused overview, whereas Ochi et al. (Mieko, Komura, and Ishikawa 2025) emphasize applications and clinical-translation obstacles.Source: §1 · p. 2

2. Foundation Models: Definition, Importance, and Technical Foundations

• Foundation models extend deep learning’s data-driven progression by allowing many applications to adapt shared pretrained models rather than training separate systems from scratch.Transformer architectures, self-supervision, and unprecedented training-data scale enable transfer across downstream tasks, while shared foundations can propagate biases and other risks.Source: §2 · pp. 2–3

2.1. Definition and Characteristics of Foundation Models

• Foundation models are defined by scale, self-supervision, and adaptability, with emergent capabilities that exceed their simple training objectives and improve as datasets and models grow.They are trained on massive, diverse, often multimodal datasets and can perform downstream tasks not explicitly included during training.Source: §2.1 · p. 3

2.2. Pathology Foundation Models

• Pathology foundation models have progressed from image-only systems to multimodal models integrating text, RNA-seq, molecular data, and immunohistochemistry while expanding from thousands to millions of WSIs.Learning approaches have likewise advanced from classification through contrastive and self-supervised learning toward generative models that produce synthetic images, annotations, reports, and interactive outputs.Source: §2.2 · pp. 4–5
• Virchow2, Virchow2G, Phikon-v2, UNI, Virchow, H-Optimus-0, and TissueConcepts demonstrate strong performance across 37, 37, 8, 34, 33, 11, and 16 tasks, respectively.Pathology models now span tissue classification, cellular analysis, cancer subtyping, biomarker prediction, and survival prediction.Source: §2.2 · pp. 6–7
• Virchow2G was trained on 3.1 million WSIs and contains 1.85 billion trainable parameters, whereas CHIEF outperformed state-of-the-art deep learning methods by up to 36.1%.CHIEF was trained on 60,530 WSIs and validated on 19,491 WSIs from 32 independent slide sets across 24 hospitals and international cohorts.Source: §2.2 · pp. 8–9
• RudolfV outperformed UNI in 10 of 12 benchmarks and 27 of 31 datasets, and Virchow in 2 of 2 benchmarks, while generalizing across scanners, stains, and tissue types.Its pathologist-driven training used 134,000 slides and 1.2 billion patches, supporting tumor-microenvironment profiling, biomarker evaluation, and rare-disease reference search.Source: §2.2 · p. 9
• Despite rapid scaling and broader task coverage, pathology foundation models remain vulnerable to tissue, geographic, demographic, and laboratory-protocol imbalances that may limit generalizability.Common cancers dominate training data, whereas rare cancers and non-neoplastic conditions are underrepresented; systematic robustness evaluation is needed before widespread clinical deployment.Source: §2.2 · pp. 5, 9

3.1. Modeling

• Foundation models require six properties: expressivity, scalability, multimodality, memory capacity, transfer learning capacity, and systematic generalization.These properties define computational requirements for learning and generalizing from large-scale pathology data.Source: §3.1 · pp. 9–10
• Expressivity enables complex mappings from gigapixel WSIs to diagnoses across tissue types, stains, and magnification levels through sufficient depth, width, connectivity, and structural patterns.It is characterized by VC dimension and approximation bounds for Lipschitz-continuous functions (Hanin and Rolnick 2019; Montúfar et al. 2014; Raghu et al. 2017).Source: §3.1 · p. 9
• Scalability supports stable optimization as parameters and training data grow, enabling efficient WSI processing with consistent performance across laboratory preparations and scanner vendors.Scaling laws relate performance to compute budget, parameter count, and dataset size.Source: §3.1 · pp. 9–10
• Multimodality integrates WSIs with clinical metadata, molecular profiles, and structured reports through unified or shared-latent representations.Combining image and text distributions is described as important for developing intelligence (Bommasani et al. 2022).Source: §3.1 · p. 10
• Memory capacity stores and retrieves factual medical knowledge, diagnostic criteria, clinical guidelines, and procedural diagnostic workflows.It can be assessed through knowledge probing, fact-completion accuracy, and consistency across prompting strategies.Source: §3.1 · p. 10
• Transfer learning capacity enables rapid fine-tuning with minimal data and zero- or few-shot adaptation to novel pathology tasks without weight updates.Examples include adapting from common to rare cancers and applying learned diagnostic criteria to novel disease presentations.Source: §3.1 · p. 10
• Systematic generalization combines learned primitives into novel disease-pattern compositions and applies diagnostic criteria across different disease manifestations.Compositionality supports unseen-example generalization, training efficiency, interpretability, and multimodality (Bengio, Courville, and Vincent 2014; Bommasani et al. 2022; Dankers, Bruni, and Hupkes 2022).Source: §3.1 · p. 10
• These properties are implemented with multi-head self-attention, deep residual or hybrid architectures, billions of parameters, and task-agnostic pre-training objectives.The mechanisms capture long-range tissue dependencies and hierarchical features from cellular to architectural patterns.Source: §3.1 · p. 10

Training

• Self-supervised learning mines vast unlabeled text, image, and audio datasets to produce task-agnostic representations without manual annotations.Training must achieve domain completeness across diverse tasks while scaling predictably with architecture, data size, and compute resources.Source: §3.2 · p. 10
• Training objectives must balance domain generalization with computational efficiency to maximize foundation-model capabilities across tasks and domains.The central challenges are ensuring broad task coverage and designing objectives whose scaling behavior is predictable.Source: §3.2 · p. 10
• Raw inputs preserve detail but slow learning, whereas patch embeddings and tokenization improve efficiency while risking information loss.Continuous and discrete representations also offer different benefits for downstream tasks, requiring an explicit trade-off.Source: §3.2 · pp. 10–11
• Generative models provide flexibility and interactivity but are computationally intensive, while discriminative models learn faster and suit high-dimensional data such as images and audio.This architectural choice directly shapes foundation-model performance and reflects a trade-off between capability and efficiency.Source: §3.2 · p. 11

3.3. Training Pathology Foundation Models

• The review covers 40 foundation models using embedding strategies that span image-only, multimodal, and vision-language training for pathological analysis.Table 2 and Table 4 summarize data scales and training details for single-modality and multimodal models.Source: §3.3 · p. 11

3.3.1. Self-Supervised Learning and DINO-based Models

• Self-supervised and DINO-based models scale architectures and datasets while adapting training to preserve morphology, staining robustness, slide context, and efficient patch sampling.Examples include Virchow2G’s 1.9B parameters, Hibou’s dataset of over 1 million WSIs, RudolfV’s expert-curated 134,000 slides, and Prov-GigaPath’s tile-to-slide architecture with zero-shot prediction.Source: §3.3.1 · pp. 12–13

3.3.2. Vision Transformer Based Architectures

• Vision Transformer models pursue either data scale, architectural capacity, hierarchical multiscale representation, or lightweight out-of-distribution robustness for pathology analysis.Phikon-v2 used 460 million pathology tiles, BEPH used over 11 million TCGA tiles and outperformed DINO and ResNet in classification and survival prediction, while HIPT and PLUTO target multiscale representation and efficient deployment.Source: §3.3.2 · p. 13

3.3.3. Hybrid and Multi-Task Models

• Hybrid and multitask models combine convolutional, transformer, supervised, contrastive, and task-specific learning to enrich features, improve generalization, and reduce labeling or retraining demands.TissueConcepts jointly trains classification, segmentation, and detection, while REMEDIS targets out-of-distribution robustness with minimal retraining.Source: §3.3.3 · pp. 13–14

3.3.4. Multimodal and Vision-Language Integration

• Multimodal and vision-language models align stains, magnifications, images, reports, molecular data, and text to support subtyping, retrieval, question answering, report generation, and prediction.TITAN supports zero-shot classification and rare cancer retrieval, while MUSK combines 50 million pathology patches, one billion text tokens, and one million image–text pairs across diverse tasks.Source: §3.3.4 · pp. 14–16

3.3.5. Conclusion and Future Directions

• The review identifies scaling, expert domain adaptation, multimodal interaction, and efficient architectures as major trends, while calling for standardized benchmarks, interpretability, clinical validation, and richer goal-directed training.Future directions include integrating additional modalities, improving efficiency and explainability, discovering richer training signals, and enabling multitask, multiagent, and multimodal adaptation.Source: §3.3.5 · pp. 16–18

GPU Needs for Pretraining

• Pretraining large-scale single- and multimodality foundation models requires several GPU computers, creating a major barrier for academic research groups and laboratories.Table 6 presents examples of the GPU-resource needs for pretraining pathology foundation models.Source: §3.4 · p. 18, Table 6 · p. 18

3.5.1. Advancing aggregation

• Nine foundation models develop new aggregation methods, including slide-level contrastive learning, sparse or dilated attention, attention pooling, and multimodal alignment.Examples include mSTAR, Prov-GigaPath, SlideChat, CHIEF, COBRA, TITAN, and THREADS, while mSTAR consistently outperformed UNI, CONCH, PLIP, and ResNet50 across seven tasks.Source: §3.5.1 · pp. 18–19
• PathAlign represents each WSI with up to 10,240 patch embeddings, and pathologists rated its generated WSI text clinically accurate in 78% of cases.The model uses BLIP-based image-text alignment and a Q-Former image transformer.Source: §3.5.1 · p. 18
• PRISM achieved AUC scores exceeding 0.9 for cancer detection and subtyping, including rare cancers, in both linear probing and fine-tuning.It integrates Virchow tile embeddings with clinical report text through Perceiver slide encoding and BioGPT language decoding.Source: §3.5.1 · p. 19

3.5.2. Reusing Attention-based aggregation

• Seven foundation models reuse attention-based MIL aggregation, including multi-stain, multi-resolution, permutation-invariant pooling, and selective cross-attention for slide-level prediction.UNI, Virchow, and BEPH use MIL-based WSI representations, while Virchow’s Agata aggregator attends selectively to tiles contributing to label decisions.Source: §3.5.2 · pp. 19–20

3.5. Aggregation

• Most foundation models do not perform aggregation, while gigapixel WSI size makes slide-level prediction computationally intensive and limits direct computer-vision transfer.Seven models focus on ROI predictions rather than aggregating across entire slides.Source: §3.5 · p. 18, §3.5.3 · p. 20

3.5.3. Challenges and Future Directions

• Future progress requires integrating and evaluating aggregation within foundation models, improving multimodal integration, and advancing hardware and software for scalable solutions.Variation among aggregation approaches currently produces inconsistent performance across pathology tasks, leaving substantial room for growth toward clinical-grade systems.Source: §3.5.3 · p. 20

Motivation

• Robust evaluation is needed to track progress, foster understanding, and document foundation models as they are adapted across numerous applications.Source: §4 · p. 20

Evaluation framework

• Evaluation should combine intrinsic assessment of foundation-model capabilities with extrinsic assessment of adapted models’ task-specific emergent abilities.Intrinsic evaluation is task-independent, whereas extrinsic evaluation uses adapted models and meta-benchmarks, making gains difficult to attribute solely to the foundation model.Source: §4 · p. 20

Intrinsic evaluation

• Intrinsic evaluation can directly measure bias, vision, and linguistic competence independently of any specific task.Source: §4 · p. 20

Resource-aware evaluation

• Evaluations should account for all training and adaptation resources, including data used to select adaptation methods and access constraints.This broader accounting can clarify effective adaptation strategies across different contexts.Source: §4 · p. 20

Adaptation

• Adaptation adjusts broadly pretrained foundation models to improve precision and relevance on specialized tasks, often more efficiently than training from scratch (Bommasani et al. 2022; C. Shi et al. 2024).It addresses limitations in domain knowledge and contextual sensitivity while supporting real-world accuracy and efficiency (Bommasani et al. 2022; Z. Chen et al. 2024; W. Li et al. 2024).Source: §4.1 · p. 21
• Adaptation methods include prompting, parameter-updating fine-tuning, continual learning, and low-storage approaches that selectively modify parameters (Bommasani et al. 2022; Firoozi et al. 2023; J. Liu et al. 2024).Prompting changes task-specific instructions without altering internal structure, whereas fine-tuning uses domain-specific data to improve focus and accuracy.Source: §4.1 · p. 21
• Although adaptation bridges generalization and specialization, it requires additional resources and introduces efficiency, usability, misalignment, and feedback-loop concerns (Bommasani et al. 2022; Alowais et al. 2023; X. Zhang et al. 2024).No model is equally effective across all domains, making adaptation important despite its costs and risks.Source: §4.1 · p. 21
• Continual learning could reduce retraining by updating models as knowledge evolves, but catastrophic forgetting can overwrite previously learned information (Bommasani et al. 2022; Verwimp et al. 2024; T. Wu et al. 2024).Memory mechanisms and parameter updates are being explored to preserve prior knowledge while maintaining alignment with current socio-cultural expectations.Source: §4.1 · p. 21
• QPMIL-VL (Gou et al. 2024) supports incremental WSI classification by using CONCH and queryable prototypes that capture diverse visual patterns and reduce forgetting.New images are matched against stored representative features, helping retain previously learned information.Source: §4.1 · p. 21

4.2. Evaluation of pathology foundation models

• Extrinsic evaluation dominates pathology foundation-model assessment, with models adapted to specific downstream tasks across image-only, aligned, multimodal, and copilot settings.Evaluations include classification, segmentation, retrieval, survival, biomarker prediction, report generation, question answering, and interactive assistance.Source: §4.2 · p. 21, §4.2.1 · p. 21, §4.2.2 · p. 24

4.2.1. Evaluation of image only models

• 10.8% average improvement over the closest contender was achieved by RodulfV (Dippel et al. 2024) in cross-tissue disease understanding and tumor-microenvironment cell characterization.RodulfV was also evaluated for histological and molecular prediction, biomarker assessment, and image-based reference-case search.Source: §4.2.1 · pp. 22, 24
• 6% of the data and resources enabled TissueConcepts (Nicke et al. 2024) to achieve comparable performance to self-supervised foundation models, while emitting 18.91 kg versus up to 2004 kg of CO2.Its supervised multitask training was evaluated across breast, colon, lung, and prostate WSI classification; intrinsic evaluation remains rare, with Kaiko-ai using RankMe and ODCorr.Source: §4.2.1 · pp. 22–23

4.2.2. Evaluation of image and text aligned models

• Across aligned and multimodal evaluations, task-specific fine-tuning complicates measurement of generalist competence, while direct assessments of bias, standalone embedding quality, and universal zero-shot ability remain research gaps.MUSK (Xiang et al. 2025) and THREADS (Vaidya et al. 2025) may also have overestimated TCGA performance because TCGA was included in pretraining.Source: §4.2.2 · p. 24, §4.2.3 · p. 26
• MUSK outperformed seven other foundation models in zero-shot cross-modal retrieval and achieved higher performance across image retrieval, classification, molecular prediction, relapse, prognosis, response, and survival tasks.By contrast, CONCH (M. Y. Lu, Chen, Williamson, Chen, Liang, et al. 2024) and other visual-language models performed worse on challenging zero-shot problems than supervised counterparts.Source: §4.2.2 · pp. 24–25

4.2.3. Evaluation of multimodal models

• THREADS (Vaidya et al. 2025) outperformed baseline foundation models on all 54 finetuning tasks against CHIEF (Xiyue Wang et al. 2024), 40 of 54 against GIGAPATH, and 12 retrieval tasks.It was evaluated across clinical, molecular, immunohistochemistry, treatment-response, and survival tasks, and introduced molecular prompting for unsupervised transfer and generalization; TCGA inclusion may overestimate generalization.Source: §4.2.3 · p. 26

4.2.4. Evaluation of AI Copilots

• PathChat (M. Y. Lu, Chen, Williamson, Chen, Zhao, et al. 2024) was evaluated on 54 diagnostic categories and 260 open-ended questions, with seven pathologists assessing responses and interactive tests demonstrating morphology summarization, diagnostic inference, and IHC guidance.The benchmark included image-only and image-plus-clinical-context settings, while follow-up conversations examined potential human-in-the-loop differential diagnosis.Source: §4.2.4 · p. 27
• AI copilots are evaluated across diagnostic and open-ended pathology questions, interactive conversations, multimodal benchmarks, whole-slide interpretation, and specialized tasks such as cell generation and positive-cell counting.PathChat was assessed on PathQABench and follow-up multi-turn interactions; PathAsst on PathVQA and complex assisted tasks; PathGen on the PathMMU test set and dialogue data; SlideChat processes entire WSIs and supports diverse clinical reasoning tasks.Source: §4.2.4 · p. 28
• PathChat evaluation lacks assessment of when to seek contextual information, recognize unresolved morphologic ambiguity, and follow test results or institution-specific guidelines.Its inability to process entire WSIs and reliance on retrospective training can also contribute to factually inaccurate responses.Source: §4.2.4 · p. 28
• Evaluation should extend beyond accuracy to robustness, fairness, resource use, costs, environmental impact, and human-in-the-loop assessment of generative-model subtleties.Such designs aim to better capture foundation models’ true capabilities and limitations for progress and decision-making.Source: §4.2.4 · p. 28
• Foundation-model tasks can be categorized as conventional predictive analyses, advanced dense or image-text-aligned tasks, and unique capabilities demonstrated by individual models or AI copilots.Figure 6 organizes downstream computational pathology tasks according to these three categories.Source: Figure 6 · p. 29, §4.2.4 · p. 29

4.3. Computational Pathology Advance

• Computational pathology has progressed across cancer detection, triage, tissue classification, molecular prediction, mutation identification, patient outcomes, and treatment response, while foundation models pioneer additional clinical applications.The section frames foundation models as extending established downstream tasks toward broader clinical research and diagnostic workflows.Source: §4.3 · p. 29

4.3.1. Cancer Detection and Subtyping

• Foundation models set new benchmarks across cancer detection and subtyping, including rare cancers, 108 cancer types, 46-class slide-level classification, and out-of-domain evaluation.UNI classifies up to 108 OncoTree cancer types, while Virchow2 achieves state-of-the-art performance across 12 tile-level tasks and in-domain and out-of-domain benchmarks.Source: §4.3.1 · pp. 29–30

4.3.2. Tissue Classification and Image Analysis

• Foundation models advance tissue classification, nuclei and cell segmentation, whole-slide analysis, survival modeling, and genomic biomarker prediction across cellular-to-slide scales.PathAlign achieves AUROCs of 0.945 to 0.987, while OmniScreen reaches a mean AUROC of 0.89 across the 15 most common cancers.Source: §4.3.2 · p. 30

4.3.3. Molecular Pathway Prediction and Genetic Mutation Identification

• Foundation models achieve strong molecular subtyping and mutation prediction, with Prov-GigaPath reaching state-of-the-art performance in 25 out of 26 digital pathology tasks.OmniScreen simultaneously predicts 1,228 genomic biomarkers across 70 cancers and identifies biomarkers associated with phenotype-genotype correlations and treatment response.Source: §4.3.3 · p. 30

4.3.4. Report Generation and Diagnostic Assistance

• Foundation models generate comprehensive clinical reports and provide AI-copilot assistance for complex queries, zero-shot classification, multimodal interpretation, and interactive diagnosis.PathChat supports multiple-choice questions, multi-turn conversation, descriptions, guardrails, human-in-the-loop diagnosis, and open-ended queries, outperforming GPT-4V and LLaVA when images and clinical context are available.Source: §4.3.4 · pp. 30–31

4.3.5. Pathology-Specific Innovations

• Pathology-specific innovations include rare-disease classification, multiple-stain analysis, metastatic and survival modeling, and interpretation of complex pathology data.CONCH recognizes up to 30 brain-tumor categories, while other models support immunohistochemistry diagnostics, multimodal survival analysis, PD-L1 detection, cell generation, and positive-cell counting.Source: §4.3.5 · p. 31

4.4. Benchmarking Pathology Foundation Models

• Benchmarking efforts span twelve studies and assess pathology foundation models across clinical and computational applications, revealing their strengths, limitations, and potential for digital pathology.The reviewed models include self-supervised, self-distilled, and multimodal approaches evaluated across slide analysis, classification, biomarker prediction, retrieval, question answering, and report generation.Source: §4.4 · p. 31

4.4.1. Evaluating the Versatility and Generalizability of Foundation Models

• Versatility benchmarks show that diverse pretraining and online feature re-embedding improve WSI classification, while UNI consistently led adaptation assessments and PEFT or few-shot methods supported generalization.BeyondMIL-WSIModeling achieved validation AUC 0.82 versus 0.76 for attention-based MIL, using joint high-resolution end-to-end tile and slide training.Source: §4.4.1 · pp. 31–32

4.4.2. Benchmarking Across Computational Tasks

• Across computational tasks, GPFM achieved top performance on 29 of 39 tasks with average rank 1.36, while aggregation and SSL benefits depended on domain, task, and label regime.DINO exceeded 90% AUC in detection, Giga-SSL doubled predictable mutations, and no aggregation model was uniformly best across tasks.Source: §4.4.2 · pp. 32–33

4.4.3. Benchmarking Foundation Models on Clinically Relevant Tasks

• Clinically relevant benchmarks found domain-pretrained models generally outperform ImageNet baselines, while data diversity, complementary models, and task-specific refinement can matter more than model size alone.CONCH excelled in 42% of tasks, whereas an ensemble of complementary models surpassed CONCH in 66% of tasks; H-optimus-0 achieved the best ovarian subtype performance.Source: §4.4.3 · pp. 33–34

4.4.4. Critical benchmarking findings

• Foundation models show impressive but non-transformative clinical performance, often requiring retraining or fine-tuning; Giga-SSL achieved AUROC ≤0.75 for MSI prediction, underperforming supervised models.Performance also varies across biomarker prediction, and full-resolution, end-to-end aggregation remains important for meaningful slide-level improvements.Source: §4.4.4 · pp. 34–35
• Local benchmarks use differing datasets and protocols, making direct and fair comparison difficult; robust evaluation should include novel tasks, real-world corruption, multimodal safety, privacy, and explainability.Interactive clinical assistants and co-pilots require dedicated benchmarking before responsible deployment.Source: §4.4.4 · p. 35

• Foundation models could revolutionize clinical diagnostics and research, contributing to more equitable and efficient healthcare outcomes.Source: §4.5 · p. 35
• AI copilots such as Paige Alba and Judithf could democratize pathology tools, accelerate discovery, and improve precision-medicine research, but limited accessibility and evaluation difficulties constrain their use.Paige Alba integrates real-time insights for clinical teams, while Judithf supports biomedical image analysis; both are described as research-use tools.Source: §4.5 · p. 35
• AI agents and copilots can produce inequitable outcomes through intrinsic, adaptation, representational, and downstream-use harms.Intrinsic and adaptation biases arise from training and adaptation data, modeler diversity, architectures, and objectives, while extrinsic harms affect user experiences in downstream contexts.Source: §4.5 · pp. 35–36
• These harms include misrepresentation, underrepresentation, overrepresentation, and performance disparities across individuals, population groups, and subgroups (Bommasani et al. 2022).Source: §4.5 · p. 36

5.1. Adaptation of Pathology Foundation Models

• VLSA improves data-efficient WSI survival analysis by using prognostic language priors, ordinal survival prompts, and an ordinal incidence function to guide weakly supervised visual aggregation.Experiments across five datasets assessed the paradigm, with predictions interpreted using Shapley values.Source: §5.1 · p. 36
• founCTransPath with CLAM achieved the highest AURCO of 0.86 for ovarian bevacizumab response prediction, while survival models significantly stratified aggressive-subtype high- and low-risk cases.The comparison covered six foundation models and three MIL methods, with statistical significance at p < 0.05.Source: §5.1 · p. 36

5.2.1. Open AI’s CLIP adaptation for pathology tasks

• CLIP adaptations support zero-shot, few-shot, retrieval, and explainable pathology analysis through contrastive alignment, prompt learning, feature fusion, knowledge trees, and cross-modal representations.MI-Zero outperformed OpenAI’s CLIP and ARCH-trained variants but not fully supervised ABMIL on full data; the pathology knowledge tree contains 50,470 attributes for 4,718 diseases across 32 tissues.Source: §5.2.1 · pp. 36–37

5.2.2. Text-Guided foundation models for Image Classification

• CAMP reduced computational time and storage requirements by up to 94% and 85%, respectively, while GPT-4V-DPath, CITE, GPC, PathM3, FiVE, and TANGLE demonstrate multimodal adaptation gains.CITE enables training with as few as 1 to 16 slides per class, and FiVE achieved at least 9.19% higher accuracy than counterparts on TCGA Lung Cancer few-shot experiments.Source: §5.2.2 · p. 38

5.2.3. SAM-Based Adaptations in Computational Pathology

• SAM adaptations excelled in several pathology tasks, but SAM for HistoPath performed inconsistently on dense instance segmentation despite strong results for large connected tissue structures.SAM-Path improved semantic segmentation, while Cell-ViT, Cell-ViT-SAM, and fine-tuned SAM also showed strong nuclei segmentation or external generalization.Source: §5.2.3 · p. 39
• SAM-MIL outperformed pooling and MIL approaches, including CLAM, with AUCs of 96.08 and 96.01 and F1-scores of 89.36 and 91.42 on CAMELYON-16 and TCGA Lung.SAM-MIL incorporates spatial context and SAM-Guided Group Masking to address class imbalance.Source: §5.2.3 · p. 39
• SegAnyPath improved Dice scores by 29.27% over fine-tuned SAM on external datasets using stain augmentation, self-distillation, task-guided MoE decoding, and large-scale pretraining.It used over 1.5 million images and 3.5 million masks from 39 public sources spanning organs, magnifications, and staining protocols.Source: §5.2.3 · p. 39

5.2.4. Findings and Critical Remarks

• A comprehensive pathology-specific segment-anything framework remains lacking because no model is exclusively pretrained on pathology data or extensively fine-tuned on millions of histopathology slides.This limits multi-level aggregation for predictive and prognostic tasks and motivates pathology-specific foundation-model development.Source: §5.2.4 · p. 39

6.1. Prostate Cancer Detection: Paige Prostate vs. Virchow Foundation Model

• Paige Prostate achieved an AUC of 0.995 (P < 0.05), exceeding Virchow’s 0.980 AUC despite Virchow’s better generalization.Virchow was trained on 1,488,550 WSIs and fine-tuned on 35,387 slide groups, whereas Paige Prostate used 66,713 prostate blocks.Source: §6.1 · p. 40, §6.4 · p. 41

6.4. Findings and Critical Remarks

• Clinical adoption remains constrained by foundation models’ task-specific performance gaps, cohort variability, limited pre-screening evidence, and the higher financial and environmental costs of large-scale development.The section calls for comprehensive benchmarking, improved generalization, expanded metrics, and model refinement to establish clinical-grade performance across diverse medical tasks.Source: §6.1 · p. 40, §6.2 · p. 40, §6.3 · p. 41

6.2. Microsatellite Stability (MSS) Screening in Colorectal Cancer

• MSI prediction performance varied substantially across cohorts and models, with AUROCs ranging from 0.66 to 0.978 and OmniScreen reaching 0.98.External cohorts reported AUROCs of 0.96, 0.972, and 0.98, while MSIntuit CRC achieved sensitivities of 92–95% and ruled out approximately 40% of MSS patients.Source: §6.2 · p. 40, §6.4 · p. 41

6.3. Colon Biopsy Pre-Screening AI

• CAIMAN and IGUANA achieved specificities of 0.56 and 0.55, respectively, at 0.99 sensitivity, automatically ruling out more than 55% of normal biopsies.Foundation-model evaluation for this task remains limited: UNI, REMEDIS, and CTransPath reported balanced accuracies of 65%, 52%, and 39%, while PathAlign lacked quantitative diagnostic evaluation.Source: §6.3 · p. 41, §6.4 · p. 41

7.1. Data Acquisition and Quality

• Gigapixel whole-slide images, staining and scanning variability, costly expert annotation, and inconsistent pathologist reporting impede standardized, generalizable training data.These challenges slow robust model development and motivate structured, high-quality datasets and reporting standardization.Source: §7 · p. 41, §7.1 · p. 42

7.2. Data Privacy and Security

• Privacy regulations complicate cross-institutional data sharing, while proprietary, regionally limited datasets and underrepresented diseases weaken foundation-model generalizability across clinical populations.Uncertainty over data ownership further complicates collaborative development.Source: §7.2 · p. 43

7.3. Computational Infrastructure

• Processing gigapixel images requires high-performance computing that is unevenly available, and cloud infrastructure does not reliably provide the low latency required for clinical decisions.Hardware constraints and energy consumption also limit deployment across institutions, particularly in resource-limited settings.Source: §7.3 · p. 43

7.4. Model Interpretability

• Foundation-model interpretability remains a clinical-adoption barrier, although mechanistic analysis linked interpretable cell and tissue morphology features with gene expression and spatial-transcriptomics evidence.Regulatory explainability requirements and limited interoperability analysis further intensify the need for transparent models (Le et al. 2024).Source: §7.4 · p. 43

7.5. Generalization and Robustness

• Pathology-specific foundation models UNI and CONCH significantly outperformed ImageNet-pretrained ResNet models for PANDA prostate grading under domain and label shifts, even with limited training data.Standardized benchmarks modeled on GLUE, SuperGLUE, and few-shot evaluation are needed to assess robustness under real-world variability.Source: §7.5 · pp. 43–44

7.6–7.10. Integration, Ethics, Scalability, and Findings

• Clinical integration is hindered by mistrust, infrastructure costs, workflow resistance, algorithmic bias, accountability, limited accessibility, and energy-intensive scaling, despite predictive, contrastive, and generative applications improving pathology’s potential.Local training and fine-tuning can improve targeted workflow tasks, while broader access could democratize tools and support more equitable precision medicine (Hassan, Kushniruk, and Borycki 2024).Source: §7.6 · p. 44, §7.7 · p. 44, §7.9 · p. 44