A Survey of Pathology Foundation Model: Progress and Future Directions

1 Introduction

Computational Pathology (CPath), the computational analysis of patient specimens (i.e., Whole Slide Images, WSIs), is increasingly important due to the critical role of histopathology. For gigapixel WSIs, Multiple Instance Learning (MIL) is the de facto framework, involving WSI patch partitioning, feature extraction via pretrained neural networks, and feature aggregation into WSI-level features [Xiong et al., 2024b]. Therefore, MIL performance hinges on two components: the pretrained neural network (extractor) and the aggregator.

Pathology Foundation Models (PFMs), neural networks pretrained on extensive pathological data that can be directly leveraged for diverse downstream tasks without retraining, such as HIPT [Chen et al., 2022] and UNI [Chen et al., 2024], mark a paradigm shift for MIL. Conventionally, due to the lack of PFMs, ResNet-50 [He et al., 2016] pretrained on ImageNet [Deng et al., 2009] serves as the extractor [Xiong et al., 2024a], but struggles with pathology-specific characteristics like minimal color variation, rotation-agnosticism, and hierarchical tissue organization. While limited labeled WSIs prevented supervised pretraining, Self-Supervised Learning (SSL) enables PFMs that exhibit superior generalizability in morphology recognition. This overcomes natural image pretraining limitations, in which features mainly capture general visual attributes like edges and textures, enabling better performance on downstream tasks even with limited data.

Despite their potential, PFMs face multifaceted challenges: 1) most PFMs directly adopt natural image techniques, failing to cater to the discrepancy between pathology and natural images, indicating pathology-specific methodology remains underexplored; 2) MIL, as a two-stage pipeline, traps model training in local optima, while end-to-end training of WSIs requires prohibitive computational resources; 3) undefined model and data scaling bounds and resource constraints necessitate multi-institutional federated learning, demanding efficiency; and 4) the computational demands of PFMs impede deployment and maintenance, requiring continuous adaptation to evolving WSI technologies and pathological variants.

Recent surveys on PFMs have contributed significantly to the understanding of the field; however, these works either focus primarily on the impact of PFMs on the real world rather than technical investigations of them [Ochi et al., 2025], or detail the previous efforts in this field without a systematic taxonomy for technical analysis and a systematic organization of the evaluation tasks of PFMs [Chanda et al., 2024]. To address these critical gaps, we introduce a comprehensive and timely survey of the current landscape. We collected papers from high-impact journals, including Nature, Nature Medicine, Nature Biomedical Engineering, Medical Image Analysis, as well as top-tier conferences such as CVPR, ICML, and AAAI. Given the rapid evolution of the field, we also incorporated preprints from repositories such as arXiv, bioRxiv, and medRxiv, acknowledging that many influential works are still under review. In total, our survey includes 27 PFM papers, 12 of which are preprints that have not yet been accepted by peer-reviewed conferences or journals.

We present this survey with three primary contributions: 1) a hierarchical taxonomy organizing PFMs based on scope, training strategy, and design to enable holistic analysis, transferable to general vision FMs; 2) a comprehensive analysis of evaluation methodologies, examining their technical merits and limitations; and 3) a structured analysis of pathologycentric research challenges prioritizing underexplored directions. The manuscript is organized as follows: Section 2 formally formulates MIL and SSL; Section 3 introduces the proposed hierarchical taxonomy; Section 4 examines evaluation tasks for PFMs; Section 5 delineates future research directions in this field; and Section 6 concludes our survey.

2 Background and ProblemFormulation

2.1 Multiple Instance Learning

In the MIL framework, a WSI is typically represented as a bag of N unordered instances (or patches). The central objective of MIL is to predict the WSI-level label ˆ Y using only the ground truth bag label Y as supervision, without access to the ground truth instance-level labels {yi}Ni=1 . This setting reflects a common scenario in computational pathology, where obtaining slide-level annotations is feasible, but annotating individual patches is prohibitively expensive and impractical. The relationship between the bag label Y and the instance labels {yi}Ni=1 is typically defined under standard MIL assumptions such as the presence-based assumption, and can be formally expressed as [Xiong et al., 2023],

The implementation of MIL involves tessellating WSIs into non-overlapping patches X = {xi}Ni=1 ∈ RN×h×w×3 , with h,w standing for height and width, respectively. These patches undergo feature extraction through an extractor Me(·), generating corresponding features Z = {zi}Ni=1 ∈ RN×d , where each feature is computed as zi = Me(xi), and d is the hidden dimension of the extractor. Subsequently, an aggregator Mg(·) agglomerates these features to form a baglevel feature h = Mg(Z) of the WSI, which finally serves as the input for the classification layer. Throughout aggregation, the extractor Me(·) usually remains frozen and non-trainable due to GPU memory constraints, while the aggregation network Mg(·) is optimized during training. We refer readers unfamiliar with MIL to prior surveys for more details [Carbonneau et al., 2018; Waqas et al., 2024].

2.2 Self-supervised Learning

SSL leverages unlabeled data by automatically generating supervisory signals through pretext tasks [Ericsson et al., 2022].

Given an input image x, a transformation function T (·) is applied to generate a modified version ˜x = T (x) and a corresponding pseudo-label ˜ y. An extractor Me(·) extracts features from ˜x and generates a predicted label ˆ y = Me(˜x). The learning objective can be formalized as minimizing the difference between the predicted label ˆ y and the pseudolabel ˜ y. Common pretext tasks include contrastive learning, self-distillation, masked image modeling, etc., each designed to force the model to learn meaningful semantic features of the data. Through this process, the extractor can learn transferable features for downstream tasks on massive unlabeled data. We refer readers who are unfamiliar with SSL to prior surveys for more details [Ericsson et al., 2022; Shurrab and Duwairi, 2022; Gui et al., 2024].

3 Hierarchical Taxonomy

Our taxonomy systematically organizes PFMs through three interdependent dimensions and reflects a top-down design philosophy: 1) Model Scope: a categorization of the scope of the PFMs, differentiating between PFMs focused on extractors, aggregators, and both components; 2) Model Pretraining: a detailed examination of the spectrum of image-centric pretraining methods, including slide-level, patch-level, and multimodal techniques; and 3) Model Design: a rigorous analysis of architecture, categorizing PFMs according to their number of parameters and scale. This top-down structure enables systematic comparisons of PFMs, as shown in Table 1.

3.1 Model Scope

MIL consists of three parts: 1) patch partitioning, 2) feature extraction, and 3) feature aggregation. As patch partitioning has been well-established, MIL performance primarily depends on the extractor and aggregator. In addition, WSIs inherently exhibit hierarchical structures, where local histomorphological patterns captured by extractors and global hierarchical tissue organization modeled by aggregators jointly determine diagnostic accuracy. Therefore, we categorize PFMs based on their scope: extractor-centric, aggregator-centric, or hybrid-centric. The categorization of PFMs along this dimension is presented in the Model Scope column of Table 1.

Extractor-centric models constitute the predominant approach in PFM development, driven by two factors: the importance of high-quality features and the necessity to address domain shift brought by ImageNet-pretrained CNNs.

The role of the extractor aligns with established clinical practice, where pathologists emphasize cellular morphological analysis at the patch level. CTransPath [Wang et al., 2022] pioneers the extractor training with a hybrid CNNTransformer design through Semantic-Relevant Contrastive Learning (SRCL) on 15 million patches. REMEDIS [Azizi et al., 2023] demonstrates that the feature extraction capability of ResNet-50 is constrained by domain shift across different medical imaging domains. Various advancements, including Virchow [Vorontsov et al., 2024] and SINAI [Campanella et al., 2024], further stress the significance of robust extractors.

Aggregator-centric models play a vital role in slide-level tasks as they are the only trainable models under direct supervision of ground truth labels, yet they are relatively underexplored compared to the extractor. CHIEF [Wang et al., 2024b], leveraging supervised pretraining with the anatomical site to create an anatomy-aware aggregator, first demonstrates the efficacy of aggregator pretraining. More recent research like MADELEINE [Jaume et al., 2025], TITAN [Ding et al., 2024], and THREAD [Vaidya et al., 2025] utilizes multimodal data in aggregator pretraining with frozen patch features to enhance performance across downstream tasks. This paradigm shift reflects growing awareness that the aggregator critically impacts downstream task performance, particularly in low-resource clinical scenarios [Xu et al., 2024a]. This observation aligns with transfer learning principles, wherein pretraining on large-scale datasets effectively alleviates downstream data scarcity challenges. However, empirical evidence also reveals that the pretrained CHIEF aggregator occasionally performs worse than linear probing of the extractor [Ding et al., 2024], which is potentially attributable to the small model size when trained on a pretraining-scale dataset, or to the conflicts between domain bias and generic features. Consequently, further investigations are warranted to assess the advantages of pretrained larger aggregators.

Hybrid-centric models are PFMs that pretrain both the extractor and aggregator. Their advantage lies in full exploitation of the aggregator, as the aggregators can flexibly adapt to the extractor with pretraining-scale data. HIPT pioneers this approach through hierarchical pretraining of the first two layers of the extractor, excluding the last layer, which is substantiated through empirical performance. Similarly, Prov-GigaPath [Xu et al., 2024a] pretrains a ViT extractor and a LongNet [Ding et al., 2023] slide encoder; however, LongNet generates instance-level features rather than a single slide-level feature, necessitating integration of ABMIL [Ilse et al., 2018] or non-parametric pooling strategies for slidelevel tasks. TANGLE [Jaume et al., 2024] pretrains both a ViT feature extractor and a transcriptomics-guided ABMIL aggregator. Finally, mSTAR [Xu et al., 2024b] distinguishes itself as a fully-pretrained hybrid-centric model by an inverted pretraining sequence, contrasting with the conventional paradigm: first optimizing the multimodal aggregator, followed by pretraining the extractor with the aggregator.

Analysis of recent developments reveals two observations. First, research emphasis has progressively shifted from feature extractor pretraining toward aggregator pretraining, a transition potentially attributable to both the robust performance of existing extractors and the increasing awareness of aggregator significance, especially in limited-data scenarios. Second, current aggregators demonstrate a hierarchical dependency pattern, wherein each successive model builds upon the capabilities of prior models. For instance, TITAN utilizes features from CONCHv1.5, which in turn leverages UNI as its encoder, thereby forming a cascading performance dependency chain where the efficacy of TITAN is inherently contingent upon CONCHv1.5 and, by extension, UNI.

3.2 Model Pretraining

The pretraining methods can be categorized into supervised and SSL methods, with SSL prevailing due to their capabilities in capturing morphological patterns without labeled data, while only CHIEF opted for supervised pretraining for the aggregator. Based on our surveyed papers, SSL can be further divided into two main categories: vision-only and intermodal methods. Vision-only methods employ three SSL techniques: contrastive learning (SimCLR, MoCov3), masked image modeling (MIM, MAE), and self-distillation (iBOT, DINO, DINOv2). In contrast, inter-modal methods often employ multi-stage pretraining, utilizing contrastive learning methods (CLIP, CoCa) for effective cross-modal alignment before which unimodal encoders are pretrained independently. We focus on methodology contributions in this section and present the details of each method, including input modalities, magnification, and resolution of the patches, in the Model Pretraining column of Table 1.

Contrastive Learning is an SSL branch that learns representations by maximizing similarity between positive pairs while minimizing that between negative pairs. Several seminal approaches have advanced this field: 1) SimCLR [Chen et al., 2020] established foundational techniques such as aggressive data augmentation and large batch sizes; 2) MoCov3 [Chen et al., 2021] advanced self-supervised learning for ViT through stabilized training techniques; 3) CLIP [Radford et al., 2021] expanded the paradigm to multi-modal learning through large-scale image-caption pair training; and 4) CoCa [Yu et al., 2022] proposed a unified method incorporating both contrastive and captioning objectives, enabling simultaneous visual-textual alignment and text generation capabilities. In the medical domain, REMEDIS utilizes SimCLR to enhance the robustness and data efficiency in medical imaging. TANGLE adopts a revised SimCLR method with gene expression reconstruction and slide subset alignment. Pathoduet [Hua et al., 2024] enhanced MoCov3 through the integration of cross-scale positioning and crossstain transferring tasks, specifically addressing the challenges of stain transferability and tissue-level heterogeneity. CLIP is adapted for both extractors (PLIP [Huang et al., 2023]) and aggregators (Prov-GiGapath, mSTAR, MADELEINE, and THREAD), due to its versatility in aligning two or more modalities. Notably, KEEP [Zhou et al., 2024] has proposed a Knowledge-Enhanced Vision-Language (KEVL) pretraining, further adapting CLIP for the extractor by incorporating domain expertise through knowledge-graph-cleaned imagetext pairs. There are several applications of CoCa, both on the extractor and aggregator: CONCH [Lu et al., 2024] adopts this framework to pretrain an extractor on 1.17 million imagecaption pairs, enhancing both zero- and few-shot capabilities, while PRISM [Shaikovski et al., 2024] and TITAN utilize CoCa to pretrain aggregators with multimodal capabilities.

Masked Image Modeling is an SSL method that learns representations by predicting masked portions of images from their visible regions. SimMIM [Xie et al., 2022] advanced the field by simplifying existing approaches through random masking and a lightweight prediction head, and MAE [He et al., 2022] introduced an asymmetric encoder-decoder design with high masking ratios. Recent investigations have demonstrated the efficacy of MIM in pretraining extractors; notably, SINAI [Campanella et al., 2024] employs MAE to pretrain ViT models on a scale of 3.2 billion patches, establishing its scalability in pathological contexts. Similarly, MUSK [Xiang et al., 2025] and BEPH [Yang et al., 2024] further validate MIM by implementing BEiT-3 and BEiTv2 architectures, respectively. Additionally, Prov-GigaPath employs MAE to pretrain its slide encoder LongNet, demonstrating the efficacy of this method on aggregator pretraining.

Self-distillation enables model learning through its own predictions across different views, simultaneously acting as teacher and student. DINO [Caron et al., 2021] pioneered the use of self-distillation by employing a teacher-student architecture with momentum encoder and multi-crop training, while iBOT [Zhou et al., 2022] performs MIM via selfdistillation with an online tokenizer, and DINOv2 [Oquab et al., 2023] refined the DINO framework by accelerating and stabilizing the training at scale. The efficacy of selfdistillation for the extractor has been demonstrated by several investigations: Phikon [Filiot et al., 2023] implements iBOT on a corpus of 43 million patches spanning 16 distinct cancer sites; Phikon-v2 [Filiot et al., 2024] employs DINOv2 on 456 million patches derived from 30 cancer sites; RudolfV [Dippel et al., 2024] incorporates DINOv2 with pathologist knowledge on 58 tissue types and 129 stains; and Hibou [Nechaev et al., 2024] further extends DINOv2 on 1.2 billion patches. Additionally, the application of selfdistillation extends beyond the extractor, as evidenced by TITAN [Ding et al., 2024], which utilizes iBOT for generalpurpose aggregator learning. These investigations demonstrate the capacity of self-distillation in PFM pretraining. In addition, there are methodological improvements customized for pathology in this category: 1) PLUTO [Juyal et al., 2024] utilizes DINOv2 together with MAE objective and Fourier losses on 195 million patches; 2) GPFM [Ma et al., 2024] proposes Unified Knowledge Distillation (UKD), incorporating MIM, self-distillation and expert knowledge distillation together as training objectives; 3) Virchow2 [Zimmermann et al., 2024] enhances DINOv2 by applying pathology-specific augmentation and reducing tissue redundancy.

3.3 Model Design

The model design refers to the following three aspects that are vital to model performance: architecture, number of parameters (# params.), scale. The scale of a model is directly determined by its number of parameters. Through quantization of the number of parameters, we establish a hierarchical scale system, facilitating standardized cross-architectural comparisons and enabling informed model selection for practical implementations. The taxonomy of PFMs along this dimension is presented in the Model Design column of Table 1.

Architecture is a pivotal determinant of PFM capabilities. Architectures adopted by PFMs can be categorized based on scope: extractor architecture and aggregator architecture. The extractor architecture encompasses two categories: 1) CNN-based architecture such as ResNet, and 2) transformer-based architecture, including ViT [Dosovitskiy et al., 2021], CNN-integrated Swin Transformer [Wang et al., 2022], BEiTv2 [Peng et al., 2022], FlexiViT [Beyer et al., 2023], and a multimodal BEiT-3 [Wang et al., 2023]. The aggregator architecture comprises two categories: ABMIL family [Ilse et al., 2018; Ding et al., 2024] and Perceiver [Jaegle et al., 2021]. For an architectural overview of each method, we refer readers to the Architecture column in Table 1.

Scale can be derived through the number of parameters. To facilitate cross-architectural comparisons, we establish a quantization framework based on the ViT architecture, which serves as the predominant backbone across our surveyed literature. The classification includes seven categories: extra small (XS, 2.78M), Small (S, 21.7M), Base (B, 86.3M), Large (L, 307M), Huge (H, 632M), giant (g, 1.13B), and Giant (G, 1.9B). The notation ViT-B/16 indicates a ViT Base model with patch size 16. For statistics of each method, we refer readers to the # Params. and Scale columns in Table 1.

Our analysis reveals several patterns in model architectures and scaling: 1) ABMIL-derived methods demonstrate clear dominance in aggregator architectures, while the ViT family predominates in extractor architectures, which are transformer-based architectures; 2) The majority of methods utilize ViT-L as their primary backbone. Due to computational resource constraints, researchers often develop complementary smaller-scale variants (ViT-S or ViT-B) alongside their primary models, while some approaches specifically target efficiency through smaller architectures; 3) ViT-L is a popular scale for extractors, whereas ViT-XS is the primary choice for aggregators, over ViT-S. This substantial disparity in parameter counts between extractors and aggregators, despite their similar training data scale, suggests a potential data-model scale mismatch that warrants further investigation; 4) a clear trend toward larger model scales is observed: while earlier approaches frequently employed ViT-B, recent methods have increasingly standardized on ViT-L, with some extending to even larger variants such as ViT-H/g/G.

4 Evaluation Tasks for the FoundationModel

Development and evaluation constitute the two fundamental pillars of PFMs. The evaluation tasks of PFMs can be systematically categorized into four aspects: 1) slide-level tasks; 2) patch-level tasks; 3) multimodal tasks; and 4) biological tasks. A comparative analysis of these evaluation tasks is presented in Table 3, providing practitioners with comprehensive criteria for model selection based on real-world applications.

Slide-level Tasks encompass analytical tasks that utilize WSIs as primary input or output modalities. These tasks include WSI classification (Cls.), survival prediction (Surv.), WSI retrieval (Retri.), and WSI segmentation (Seg.). While survival prediction methodologically represents a classification employing specialized loss functions, its clinical application differs from standard WSI classification: the latter primarily serves for diagnosis, while the former addresses prognosis. This category is the cornerstone of CPath, enabling automated diagnosis directly from WSIs with minimal manual intervention. Consequently, the majority of methods have prioritized experimental validation in this domain.

Patch-level Tasks comprise analytical tasks on patches as inputs or outputs, including patch classification (Cls.), patch-to-patch retrieval (P2P), and patch segmentation (Seg.). These tasks effectively evaluate the efficacy of the extractor, as they operate independently of additional aggregators.

Multimodal Tasks are tasks that evaluate multimodal capabilities of PFMs. These tasks encompass cross-modal retrieval, i.e., image-to-text (I2T) and text-to-image (T2I) retrieval, report generation (RG), and visual question answering (VQA). RG in our survey includes both RG and image captioning, distinguished by their input: RG utilizes WSIs to generate clinical documentation, while captioning produces concise descriptions from patches. The increasing emphasis on these tasks reflects the clinical reality that pathologists integrate multimodal data in the decision-making process.

Biological Tasks focus on biomarker detection, including genetic alteration (GA) and molecular prediction (MP). Genetic alteration includes both mutation prediction and genetic alteration, as both predict gene mutation status. Molecular prediction targets the prediction of molecular subtypes at the gene expression level, representing a distinct biomarker from genetic alteration. While these tasks can be fundamentally categorized as classification problems at either slide or patch level, their clinical applications and biological implications warrant their classification as a separate analytical category. One recently-proposed task is molecular prompting [Vaidya et al., 2025], which aims to perform clinical tasks with canonical molecular profiles without requiring any task-specific model development in a similar manner to text prompting.

While the extensive scope of our evaluation tasks precludes exhaustive evaluation by any single model, several methods, notably CONCH, UNI, MUSK, GPFM, and TITAN, provide excellent evaluation benchmarks across multiple training paradigms, including zero-shot, few-shot, and complete supervised learning, thereby providing more holistic insights into model capabilities and generalization potential.

5 Future Directions

PFMs constitute an emerging paradigm with transformative potential. Future research directions bifurcate into two primary domains: effective PFM Development and Utilization.

5.1 FoundationModel Development

Pathology-specific Methodology design is essential for PFMs that effectively capture the unique characteristics of pathology data. Most PFMs are pretrained using algorithms originally developed for natural images, neglecting critical aspects of pathology images, as detailed in Sec. 3; therefore, there is an urgent need for algorithms designed to accommodate these challenges. This deficiency extends to multimodal pretraining as well, where CLIP and CoCa are employed without customization, resulting in the omission of inherent features of pathology and related data, including genomics and reports, that are vital for comprehensive analysis.

End-to-end Pretraining is critical to achieve optimal performance for PFMs. Current PFMs adopt a two-stage pretraining paradigm: extractors are trained independently, followed by the aggregator with the extractor frozen. Evidence suggests this complicates optimization, highlighting the need for end-to-end pretraining of PFMs, which poses significant challenges in CPath, as transitioning away fromMIL requires developing extremely sophisticated and efficient architectures and algorithms capable of simultaneously integrating local and global pathology information for gigapixel images.

Data-Model Scalability is a critical direction, as performance improvements continue to demonstrate logarithmic and sub-logarithmic scaling with model and data volume, respectively, without yet reaching a clear plateau. This domain presents four sub-directions: 1) examining the relative importance of WSI and patch quantity, particularly when considering diversity, a complex concept that is yet widely acknowledged as an indicator of high-quality data; 2) exploring efficient algorithms, due to the rapid expansion in both datasets and models, evident in transitions from CONCH (ViT-B) to CONCHv1.5 (ViT-L) and from UNI (ViT-L) to UNI2 (ViTH); 3) addressing the data-model scale mismatch problem for the aggregator, detailed in Sec. 3.1; and 4) optimizing model scale, since the giant model size poses substantial deployment challenges in both hospital and academic settings.

Federated Learning with Efficiency is essential for addressing the challenges associated with collecting massivescale datasets across multiple institutions while preserving patient privacy, as few institutions can feasibly collect WSIs at the million-scale alone. However, current research in this area remains limited; for instance, HistoFL [Lu et al., 2022] has demonstrated improved performance, yet this benefit comes at the cost of significantly increased computational overhead. As PFMs continue to grow in size, scaling federated learning further exacerbates these challenges. Consequently, there is an urgent need to develop more efficient, privacy-protected methods in such large-scale crossinstitutional collaborations.

Model Robustness addresses critical challenges in multiinstitutional data curation. The acquisition of data from various sites inevitably introduces technical heterogeneity across scanning equipment specifications, image magnification levels, and staining protocols, resulting in significant data variations that embed site information [de Jong et al., 2025]. These disparities undermine training stability and model generalizability; recent work shows that most models encode site information more strongly than biological signals [de Jong et al., 2025]. These issues will be further exacerbated in federated learning under non-IID data distributions. Consequently, developing more robust algorithms and robustness evaluation metrics for PFMs is a critical research imperative.

RAG-enhanced Pathology VLM is a trending paradigm worth investigating. Contemporary trends in Large Language Models (LLMs), such as Llama [Grattafiori et al., 2024], with the prevalence of BERT-based architectures [Devlin et al., 2019] in current multimodal PFMs suggest the potential utility of integrating LLMs with ViT architectures. Furthermore, given the demonstrated efficacy of RetrievalAugmented Generation (RAG) [Gao et al., 2023] in LLMs and the critical need for domain-specific expertise in pathology, RAG methodology offers promising directions for representation learning in pathology VLMs. This approach transcends the limitations of existing methods such as RudolfV, which relies primarily on clustering techniques for pathologist knowledge integration, providing a potentially more sophisticated framework for incorporating domain expertise.

5.2 FoundationModel Utilization

Effective Adaptation of PFMs to downstream tasks is a critical research direction in their utilization, as these models are predominantly trained on large-scale heterogeneous datasets, resulting in general-purpose features rather than task-specific ones required for optimal performance. To address this limitation, effective adaptation methodologies are essential for task-specific optimization. The significance of this domain alignment challenge parallels the established paradigm of adapting conventional architectures, such as ResNet-50, to specialized domains like pathological image analysis, albeit with varying degrees of complexity and scope.

Model Maintenance constitutes a critical research domain in the context of PFMs, given the substantial computational resources required for their initial training. The potential diminishment of model performance due to novel diseases, tissue heterogeneity, or technological advancements necessitates efficient maintenance strategies to preserve model utility. Continual learning [Wang et al., 2024a; Yu et al., 2024] represents a promising approach for maintaining PFM effectiveness, as it circumvents the necessity for model retraining by learning on newly observed instances. This approach significantly reduces the required computational overhead while ensuring the model remains current with the evolving clinical, disease, and technological developments.

6 Conclusion

This survey presents a systematic analysis of the current Pathology Foundation Models through our proposed hierarchical taxonomy and comprehensive evaluation framework. Although the PFMs demonstrate significant advances in computational pathology, critical technical challenges merit further investigation. We delineate key directions that are worth exploring and might be instrumental in advancing both the theoretical foundations and practical applications of PFMs.

Acknowledgments

The authors would like to sincerely thank Professor Irwin King from the Department of Computer Science and Engineering, the Chinese University of Hong Kong, for his active involvement in the conception and development of this paper. Due to the one-submission-per-author policy, his name could only appear in the acknowledgments. The work described in this paper was partially supported by the Research Grants Council of the Hong Kong Special Administrative Region, China (CUHK 2410072, RGC R1015-23).