Artificial Intelligence for Digital and Computational Pathology

TL;DR

The paper surveys the convergence of digital pathology and AI for whole-slide image analysis, focusing on predicting clinical endpoints and discovering biomarkers. It delineates end-to-end WSI prediction versus AI-assisted tools, with multiple instance learning (MIL) as a core weakly supervised framework and context-aware architectures such as graph neural networks and transformers to model tissue architecture. It reviews segmentation, virtual staining, interpretability, and the role of public datasets and open-source resources in enabling reproducibility. It outlines data, methodological, and translational challenges and provides a roadmap toward larger multimodal cohorts, self-supervised learning, uncertainty quantification, and regulatory-ready deployment.

Abstract

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.

Figures (8)

• Figure 1: Caption next page.
• Figure 1: Applications, timeline of selected milestones and trends in computational pathology.a, Overview of computational pathology (CPath) applications. b, Digital diagnostics and artificial intelligence (AI) have made considerable progress over the past decades, laying the foundations for CPath to make a clinical impact. The timelines include selected milestones that have substantially impacted CPath. c, CPath shifted from traditional machine learning (ML) models based on small cohorts of regions of interest (ROIs) to deep learning models trained on large, sometimes multimodal, multi-institutional cohorts of whole-slide images (WSIs). Higher-dimensional pathology data, such as WSIs collected longitudinally for each patient and 3D tissue images, are also expected to gain traction. The digitization of the pathology workflow, abundant computational resources, public datasets and advances in AI and computer vision have supported this transition. CPU, central processing unit; gene-seq, gene sequencing; GPU, graphics processing unit; IHC, immunohistochemistry; MIL, multiple instance learning; MSI, microsatellite instability; TCGA, The Cancer Genome Atlas; TMA, tissue microarray.
• Figure 2: Caption next page..
• Figure 2: Multiple instance learning for clinical end-point prediction on whole-slide images. All methods take as input a set of patches extracted from a whole-slide image (WSI) at a fixed magnification and learn to map it to a WSI-level clinical end point, such as cancer grade or subtype. a, Histology slides are first digitized with a scanner as WSIs, which then go through segmentation and patching. b, For patch-level supervision, each patch is assigned a label, either by using manual patch-level annotations or by assigning the slide-level label to all patches. Patches are passed through a sequence composed of a feature extractor, patch-level predictor and aggregator to produce a WSI-level prediction. c, In multiple instance learning (MIL), a feature extractor extracts embeddings for all patches, which are then aggregated (without including context) for WSI-level prediction. d,e, For context-aware MIL, the interactions between patch embeddings (extracted with the feature extractor) are explicitly encoded using either a graph representation of patches processed with a graph neural network (part d) or a sequence representation of patches processed with a transformer (part e).
• Figure 3: Caption next page.