External Validation of a Deep-Learning Atypical Mitosis Based Prognostic Model Using the PLCO Pathology Image Repository
Principal Investigator
Name
William Chen
Degrees
MD
Institution
University of California, San Francisco
Position Title
Assistant Professor
Email
william.chen@ucsf.edu
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-2059
Initial CDAS Request Approval
Jul 14, 2026
Title
External Validation of a Deep-Learning Atypical Mitosis Based Prognostic Model Using the PLCO Pathology Image Repository
Summary
We propose to externally validate a previously developed deep multiple-instance learning (MIL) model that predicts patient survival from whole-slide H&E pathology images through the analysis of atypical mitotic activity and its surrounding tissue context. Whereas routine histologic grading reduces mitoses to a count, this architecture is designed to capture higher-order signals reflecting chromosomal instability and the tumor microenvironment. The pipeline comprises three stages: a mitosis detector, a model that embeds each mitosis and surrounding cellular neighborhood, and a weakly supervised attention-based MIL aggregator that produces a slide-level survival risk score. The model has demonstrated prognostic performance across 4,967 TCGA cases spanning 14 cancer types and in an independent colorectal cohort (n = 422; p = 0.018). We now seek to assess generalizability in the PLCO cohort.
PLCO offers several features well suited to this validation: 1) its slides were acquired on a matched scanner at 40x magnification, reducing a known source of domain shift, 2) cancer-specific survival endpoints are available through long-term National Death Index linkage, 3) its screening-arm cases are drawn from a prospectively enrolled population and span a broader disease spectrum than referral- or resection-based series, testing whether the model performance holds outside surgically selected cohorts.
The model will be applied to PLCO whole-slide images as a fixed inference pipeline without fine-tuning: candidate mitotic figures will be detected, embedded using the frozen backbone, and aggregated to yield a predicted risk score. We will evaluate predicted risk against PLCO cancer-specific mortality and overall survival using Cox proportional hazards regression, concordance indices, and Kaplan–Meier analyses. Pre-specified subgroup analyses will stratify by clinical stage, histologic grade, age, race/ethnicity, and screening-arm assignment. Finally, we will quantify the incremental prognostic value of the slide-level risk score relative to the atypical-mitosis fraction alone and to a combined model incorporating established PLCO clinical covariates.
Aims
Overall objective: To externally validate a deep multiple-instance learning (MIL) model that predicts cancer-specific survival from whole-slide H&E images by quantifying atypical mitotic activity and its surrounding tissue context, using the scanner-matched PLCO cohort with long-term National Death Index endpoints.
Central hypothesis: The slide-level risk score generated by this fixed inference pipeline will stratify PLCO cancer-specific mortality independently of, and incrementally to, established clinicopathologic covariates.
Aim 1: Confirm prognostic generalizability in PLCO. Apply the locked pipeline to PLCO whole-slide images as fixed inference, producing one slide-level risk score per case. Evaluate the score against cancer-specific mortality and overall survival using Cox proportional hazards regression, concordance indices, and Kaplan–Meier analyses by risk tertile. Success criterion: a statistically significant hazard ratio for cancer-specific mortality and a concordance index consistent with the range observed in discovery cohorts.
Aim 2: Characterize robustness across subgroups. Assess the stability of risk stratification through pre-specified subgroup analyses stratified by clinical stage, histologic grade, age, race/ethnicity, and screening-arm assignment. Quantify calibration within strata and test for effect-measure modification. This aim establishes whether prognostic performance holds across clinically and demographically distinct populations, and specifically interrogates differential performance by race/ethnicity to surface potential disparities in model behavior. Success criterion: consistent direction of effect across major subgroups, with calibration and discrimination metrics reported transparently for each.
Aim 3: Quantify incremental prognostic value. Determine whether the deep slide-level risk score adds information beyond simpler and established predictors. Compare three nested specifications: (i) the atypical-mitosis fraction alone, (ii) a model of established PLCO clinical covariates, and (iii) the combined model adding the deep risk score. Evaluate incremental value using the likelihood-ratio test, change in concordance index, and net reclassification metrics. Success criterion: a significant improvement in fit and discrimination when the deep score is added to clinical covariates, demonstrating value beyond mitotic counting and routine clinical data.
Expected outcome and significance: Collectively, these aims will establish whether a fixed, externally derived computational pathology biomarker generalizes from the resection-based discovery cohorts to an independent, scanner-matched screening population with rigorously collected survival endpoints. A successful validation would support the score as a scanner-robust, interpretable prognostic biomarker grounded in chromosomal-instability biology rather than mitotic count alone.
Collaborators
William Chen University of California, San Francisco