External Validation of Foundation Model-Derived Histologic Phenotypes in Prostate Pathology Images

Principal Investigator

Name
Olmo Zavala Romero

Degrees
PH.D.

Institution
Florida State University

Position Title
Assistant Professor

Email
osz09@fsu.edu

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-2063

Initial CDAS Request Approval
Aug 24, 2026

Title
External Validation of Foundation Model-Derived Histologic Phenotypes in Prostate Pathology Images

Summary
Recent pathology foundation models enable quantitative characterization of tissue morphology directly from hematoxylin-and-eosin (H&E) whole-slide images (WSIs). In a discovery cohort of prostatectomy WSIs with spatially registered pathology annotations, we identified reproducible latent histologic phenotypes using Virchow2 foundation model embeddings followed by principal component analysis (PCA). These latent phenotypes captured biologically meaningful aspects of prostate tissue architecture, including glandular-versus-stromal tissue organization and intra-tumor architectural heterogeneity. Initial analyses demonstrated that these phenotypes were associated with tumor-enriched regions and corresponded to distinct histologic patterns.

The objective of this project is to externally validate these latent histologic phenotypes using the PLCO prostate pathology image cohort. Whole-slide H&E images will be divided into image tiles and encoded using the pretrained Virchow2 pathology foundation model without additional model training or fine-tuning. Tile-level embedding vectors will be projected into the latent morphology space established in the discovery cohort to evaluate the reproducibility of the identified histologic phenotypes across an independent population.

We will characterize the spatial distribution of these latent phenotypes within PLCO whole-slide images and quantify associated morphologic features, including glandular architecture, epithelial and stromal organization, nuclear density, and other computationally derived histologic descriptors. Where appropriate, slide-level phenotype measurements will be examined in relation to available clinicopathologic variables provided through PLCO.

The primary goal of this study is to determine whether foundation model-derived histologic phenotypes identified in the discovery cohort are reproducible and biologically interpretable in an independent prostate cancer dataset. Successful validation would support the robustness and generalizability of these computational pathology biomarkers and facilitate future studies investigating their association with clinically relevant outcomes.

Only de-identified PLCO images and associated phenotype data will be analyzed. No attempts will be made to identify individual participants, and all analyses will be conducted in accordance with the PLCO data use agreement and applicable institutional policies.

Aims

Aim 1: Extract tile-level image features from PLCO prostate H&E whole-slide images and project these features into the latent morphology space established in our discovery cohort of prostatectomy whole-slide images.
Aim 2: Validate whether the previously identified histologic phenotypes are reproducible in PLCO, including a glandular–stromal tissue compartment phenotype and a glandular architectural organization phenotype.
Aim 3: Quantify slide-level morphology measures derived from these phenotypes, including glandular content, stromal composition, nuclear density, and architectural variation, and evaluate their associations with available PLCO clinicopathologic variables, including Gleason score, tumor grade, stage, treatment information, and available outcome measures.
Aim 4: Use the validated phenotypes to support future computational pathology studies focused on interpretable prostate cancer biomarkers and potential applications in precision oncology.

Collaborators

Olmo Zavala Romero Florida State University
YIFAN WANG Florida State University