Development of generalizable pathology foundation models for cancer prognosis using PLCO whole slide images
Principal Investigator
Name
Marvin Lerousseau
Degrees
M.Sc., Ph.D.
Institution
Spotlight Medical, SAS
Position Title
Chief Scientific Officer
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-2060
Initial CDAS Request Approval
Jul 14, 2026
Title
Development of generalizable pathology foundation models for cancer prognosis using PLCO whole slide images
Summary
Pathology foundation models are large-scale computational models trained on diverse collections of whole slide images to learn generalizable representations of histopathological patterns [1,2]. These models can capture information related to tissue architecture, cellular morphology, spatial organization, and tumor microenvironment patterns, and can subsequently be used as feature extractors for downstream clinical tasks such as prognosis prediction, patient stratification, and biomarker discovery.
At Spotlight Medical, we develop computational pathology methods to extract clinically relevant information from routine pathology whole slide images, with a focus on robust and generalizable models for cancer prognosis and patient stratification. A critical requirement for this work is access to large, diverse pathology image datasets linked to well-curated longitudinal clinical, pathological, and outcome data. Our group has previously developed and externally validated a clinicopathological assay integrating clinical variables with quantitative features extracted from H&E whole slide images, with results published in the Journal of Clinical Oncology [3].
The objective of this project is to develop and evaluate generalizable digital pathology representations using PLCO pathology whole slide images and linked clinical, pathological, treatment, mortality, and follow-up data. PLCO is a uniquely valuable resource for this objective because it includes digitized pathology images across multiple cancer sites together with standardized longitudinal phenotype and outcome data [4]. Access to the full available pathology image resource is important because foundation-model and representation-learning approaches benefit from broad morphological diversity across tissue types, cancer types, staining conditions, and disease contexts.
We request access to all available available H&E-stained images from colorectal adenoma, bladder, female breast, male breast, colorectal, lung, ovarian, and prostate tissue, as well as any available IHC-stained whole slide images, together with corresponding clinical, pathological, treatment, mortality, longitudinal follow-up, and image-linkage metadata. We will use these data to train, adapt, and evaluate computational pathology models that generate quantitative slide-level and patient-level image representations. These automatically generated features will then be assessed for their association with available cancer outcomes, including overall survival, cancer-specific mortality, and other disease-specific outcomes where available.
This is a retrospective analysis of de-identified data only. No participant contact, clinical intervention, or re-identification will be attempted. Results will be analyzed and reported only in aggregate.
[1] Vorontsov E et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature medicine. 2024 Oct;30(10):2924-35.
[2] Wang X et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature. 2024 Oct 24;634(8035):970-8.
[3] Bidard FC et al. Identifying patients with low relapse rate despite high-risk estrogen receptor–positive/human epidermal growth factor receptor 2–negative early breast cancer: development and validation of a clinicopathologic assay. Journal of Clinical Oncology. 2025 Oct 1;43(28):3090-101.
[4] Prorok PC et al. Design of the prostate, lung, colorectal and ovarian (PLCO) cancer screening trial. Controlled clinical trials. 2000 Dec 1;21(6):273S-309S.
Aims
Aim 1: Establish a computational pipeline for applying histopathology foundation models to PLCO whole slide images across available cancer sites. This will include whole slide image preprocessing, tissue-region identification, quality assessment, patch extraction, feature embedding, slide-level representation, and patient-level aggregation. The goal is to create standardized image-derived representations that can be generated reproducibly across heterogeneous archival pathology material.
Aim 2: Generate and compare foundation-model representations of PLCO pathology whole slide images using self-supervised and representation-learning methods. These models will be used to encode tissue architecture, cellular morphology, stromal organization, immune contexture, and other histopathological patterns without requiring task-specific manual annotation. Different aggregation strategies will be evaluated to determine how patch-level embeddings can best summarize slide-level and patient-level morphology.
Aim 3: Assess the relationship between foundation-model-derived pathology representations and available PLCO clinical, pathological, treatment, and longitudinal outcome data. Analyses will evaluate whether image representations capture clinically relevant variation across cancer cohorts, whether these representations complement standard clinical and pathological variables, and whether consistent morphology-outcome associations can be identified across cancer sites and outcome definitions.
Collaborators
Marvin Lerousseau Spotlight Medical, SAS
Lucie Frechin Spotlight Medical, SAS
Aymeric Vinçotte Spotlight Medical, SAS