Source-linked AI summary
Artificial Intelligence for Digital and Computational Pathology
Andrew H. Song, Guillaume Jaume, Drew F. K. Williamson, Ming Y. Lu, Anurag Vaidya, Tiffany R. Miller, Faisal Mahmood
TL;DR
Computational pathology still faces barriers to clinical adoption, including the lack of one-fits-all criteria and biases linked to disparities. The review highlights multiple instance learning and calls for diverse multimodal cohorts alongside improved tissue representation learning, while citing applications in prognosis and cancer detection.
Problem
Computational pathology lacks one-fits-all criteria, while prognosis combines multiple factors and algorithmic biases can reflect disparities in race and socioeconomic status.
Method
The review highlights multiple instance learning for predicting clinical endpoints and emphasizes large-scale, diverse, multimodal cohorts with improved deep learning frameworks for tissue representation learning.
Results
Computational pathology applications have been reported to predict colorectal cancer prognosis better than conventional cancer stages, grade prostate cancer on par with experienced pathologists, and detect breast-cancer lymph-node metastases.
Takeaways & Limitations
These approaches provide clinicians with more objective diagnoses and prognoses and allow discovery of novel biomarkers.
Takeaways & Limitations
Further work is required for computational pathology applications to get closer to clinical adoption, and high importance scores do not necessarily prove interpretability findings.
Abstract
from arXiv · showhide
Advances in digitizing tissue slides and the fast-paced progress in artificial intelligence, including deep learning, have boosted the field of computational pathology. This field holds tremendous potential to automate clinical diagnosis, predict patient prognosis and response to therapy, and discover new morphological biomarkers from tissue images. Some of these artificial intelligence-based systems are now getting approved to assist clinical diagnosis; however, technical barriers remain for their widespread clinical adoption and integration as a research tool. This Review consolidates recent methodological advances in computational pathology for predicting clinical end points in whole-slide images and highlights how these developments enable the automation of clinical practice and the discovery of new biomarkers. We then provide future perspectives as the field expands into a broader range of clinical and research tasks with increasingly diverse modalities of clinical data.
Key points
Computational pathology is approaching clinical-grade performance while supporting automated clinical tasks and biomarker discovery. Multiple instance learning is emerging for predicting clinical endpoints from whole-slide images, but broader adoption still requires further work.
- Computational pathology is reaching clinical-grade performance for certain tasks.
- Artificial intelligence methods in computational pathology include clinical-endpoint prediction and assistive tools for clinical or research use.
- Multiple instance learning is a rapidly growing paradigm for predicting clinical endpoints, including disease diagnosis and molecular alterations, from whole-slide images.
- Computational pathology can automate tasks performed by pathologists and discover morphological biomarkers for clinical outcomes.
- Larger, well-curated, and multimodal datasets are an important direction for advancing computational pathology.
- Further work is required for computational pathology applications to move closer to clinical adoption.
1 Introduction
Digitization, improved computing, and deep learning have transformed computational pathology from small-region analyses toward large-scale whole-slide image modeling. The field now supports clinical automation and biomarker discovery, while future progress depends on robust, generalizable representations built from diverse multimodal data.
- 1 Introduction: AI-assisted pathology can support objective diagnosis and prognosis, predict therapy response, and discover novel biomarkers.
- 1 Introduction: Reported applications include cancer-origin determination, prostate-cancer grading comparable to experienced pathologists, colorectal-cancer prognosis, and lymph-node metastasis detection.
- 1 Introduction: Large-scale digitized-slide repositories and more capable storage and processors enable studies based on thousands of samples.
- 1 Introduction: Deep learning is now the central algorithmic component of most computational pathology systems.
- 1 Introduction: Deep learning automatically identifies and extracts relevant morphological features from high-dimensional tissue-image data.
- 1 Introduction: Computational pathology can automate clinical and research workflows, including drug-exposure assessment of tissue morphology.
- 1 Introduction: Emerging modalities such as multiplex imaging, spatially resolved genomics, and 3D pathology create opportunities for multimodal integration.
- 1 Introduction: The Review consolidates technical developments for whole-slide image modeling and outlines robust, generalizable representations from large-scale, diverse, multimodal, privacy-preserving datasets.
2 Deep learning in CPath
Deep learning has broadened computational pathology from small-region analyses to WSI-level prediction and assistive tools. WSI pipelines typically preprocess and patch slides, extract representations, aggregate patch information, and address context, interpretability, and computational constraints.
- Deep-learning methods in computational pathology predict clinical endpoints from WSIs and support assistive tasks such as image segmentation and virtual staining.
- 2.1 Tissue pre-processing: WSIs are represented across magnifications, segmented to remove background, and partitioned into patches because their dimensions make direct processing computationally demanding.Patch-level outputs can subsequently be aggregated into slide-level or patient-level outcomes.
- 2.2 Multiple instance learning on WSI: Patch-level supervision requires time-consuming annotations, produces ambiguous labels for prognosis or therapy-response tasks, and becomes noisy when only a small image fraction is discriminative.Intratumoural heterogeneity further complicates annotation within the same tumour region.
- 2.2 Multiple instance learning on WSI: Multiple instance learning assigns one label to the patch set, extracts patch embeddings, aggregates them into a WSI representation, and predicts the WSI endpoint.Attention-based aggregation assigns patch importance scores that can produce interpretable heatmaps for qualitative morphological analysis.
- 2.2 Multiple instance learning on WSI: Because all WSI patches and the network cannot fit in GPU memory simultaneously, MIL often pre-extracts embeddings or uses host memory and gradient checkpointing for joint training.These strategies are described as complex and computationally demanding.
- 2.2 Emergence of context-aware approaches: Context-aware architectures model relationships among patches, while interpretability analyses can delineate morphological features in salient regions and support research biomarker discovery.Transformer self-attention assesses other patch embeddings when contextualizing each patch representation.
- Staining enhancement can improve access to high-quality sections, enhance deep-learning reliability, and reduce visual variability between samples.
3 Public datasets and open-sourced codes
Public datasets, challenges, and open-source software have supported methodological progress and reproducibility in computational pathology. However, maintaining shared software requires sustained community effort.
- Public challenges and open data banks provide datasets for benchmarking computational pathology methods across tasks including metastasis detection, grading, subtyping, and segmentation.
- Challenge scale has increased from 400 WSIs in CAMELYON17 to more than 10,000 WSIs in PANDA over four years.
- TCGA remains a major resource, containing more than 20,000 primary cancer cases across 33 cancer types with imaging, omics data, and patient information.
- Open-source libraries support WSI reading, visualization, annotation, patching, feature extraction, tissue detection, stain normalization, graph modeling, and multimodal inputs.
- Publicly released code and trained checkpoints are becoming standard for segmentation networks, attention networks, and pretrained histology image encoders.
- Because open-source libraries may not be regularly maintained by their developers, continued community-level support remains necessary.
4 Clinical impact of CPath
Computational pathology supports both automation of routine pathology work and discovery-oriented analyses from tissue data. Whole-slide, segmentation, and multimodal approaches can improve reproducibility, streamline diagnosis, predict outcomes, and reveal clinically relevant biomarkers.
- CPath for automation: CPath automation targets routine clinical tasks, including mitotic counting, tissue segmentation, grading, subtyping, and metastasis detection.These systems recapitulate or augment tasks pathologists perform during daily practice.
- CPath for automation: Segmentation outputs enable downstream tissue analysis, including tumor–stroma ratios, tumor-infiltrating lymphocyte assessment, cell-graph modeling, and biomarker quantification.Cellular and tissue-level features can also be combined, as in NAFLD and NASH grading.
- CPath for automation: Whole-slide algorithms integrate information across entire slides, supporting diagnosis and potentially reducing interobserver variability.Interobserver variability is nearly 50% for atypia detection in breast cancer.
- CPath for automation: Multi-institutional Gleason grading studies achieved performance on par with or exceeding that of pathologists.AI-assisted grading has also been applied to gliomas, colorectal carcinoma, breast cancer, and allograft rejection.
- Clinical translation: CPath tools can triage cases, reduce follow-up testing, shorten diagnostic turnaround, and identify patients who may not require further MSI testing.Examples include predicting candidate primary sites for metastatic cancer and identifying MSI with high sensitivity.
- CPath for discovery: Histology-based models predict survival, recurrence risk, metastasis risk, molecular features, and treatment response, while revealing morphological biomarkers.Combining clinical, histological, molecular, and IHC data improved survival or relapse-risk prediction in reported studies.
5 Outlook and future directions
Future computational pathology depends on larger, better-curated multimodal data and methods that improve representation, uncertainty handling, and clinical translation. Key directions include longitudinal and three-dimensional analysis, multiplex imaging, decentralized learning, and bias-aware deployment.
- Data outlook: Larger, curated, multi-institutional, multimodal cohorts are needed to advance computational pathology across prognostic prediction, biomarker discovery, and drug discovery.Existing multimodal cohorts are often only a few hundred samples, while CPath studies may include tens of thousands of WSIs.
- Data outlook: Multimodal dataset expansion is constrained by fragmented institutional data, missing modalities, high costs, and long collection timelines.Whole-genome sequencing remains prohibitively expensive for most institutions, and assembling sufficient quantities can take years.
- Data outlook: Longitudinal pathology requires methods that handle inconsistent assay timing, institutions, and sampling locations, including intra-organ heterogeneity.These issues make longitudinal data collection and integration methodologically complex.
- Emerging modalities: Multiplex imaging enables single-cell assessment of multiple biomarkers and may clarify tumour heterogeneity and treatment-response features unavailable from H&E alone.Current CPath studies primarily use instance segmentation with handcrafted features or graph neural networks.
- Translational considerations: Decentralized learning is needed when privacy, storage, and competitive concerns prevent centralizing multi-institutional data.Federated, continual, and swarm learning have been proposed, but remain challenging to implement.
- Learning better representations: Self-supervised learning can produce more disentangled, robust, and generalizable histopathological representations than supervised learning.Such representations may support few-shot learning at the whole-slide level, whereas end-to-end training risks overfitting and poor generalizability.
- Learning better representations: Uncertainty accounting is needed to address domain shifts, model miscalibration, and predictions that cannot be trusted.Pathology biases can arise from image-preparation artefacts and other factors.
6 Conclusion
Computational pathology is positioned to support both clinical workflows and pathology research by automating diagnosis-related work and discovering morphological biomarkers. Progress depends on large, diverse, multimodal cohorts and improved tissue representation learning enabled by coordinated efforts across organizations.
- Its research potential includes discovering morphological biomarkers related to molecular alterations, patient prognosis, and treatment response.
- Clinical and biomedical progress requires large-scale, diverse, multimodal cohorts alongside better deep learning frameworks for tissue representation learning.
- These goals are unlikely to be achieved within a single organization, motivating concerted multi-institutional efforts.
- Institutional data initiatives, open-source software, and continued inspiration from computer vision and AI research are identified as drivers of continued progress.
- Whole-slide images, digital pathology systems, H&E and IHC staining, deep learning, embeddings, and segmentation provide core concepts and tools for this research area.