Abstract
Biochemical recurrence after radical prostatectomy is a key
endpoint for risk stratification in prostate cancer. Pathology foundation
models combined with multiple instance learning have achieved competitive
BCR prediction performance from whole-slide images, yet it remains
unclear whether these models capture recurrence associated morphology
independently of Gleason grade, or whether they largely reflect grade correlated
tissue patterns. Since Gleason grade already drives routine clinical
decisions, an imaging score that simply mirrors it adds little to what
pathological grading already provides. We propose a grade-disentangled
multi-instance learning framework that uses adversarial disentanglement
to reduce grade dependence in the learned imaging representation; the resulting
risk score is then combined with clinical variables via late-fusion
Cox modeling. On the TCGA-PRAD cohort of 487 patients with 101
BCR events, our framework achieved a C-index of 0.704, outperforming
both the clinical Cox model and imaging-only multi-instance learning
baselines. The imaging-derived risk score also showed reduced correlation
with Gleason grade, indicating the model encodes prognostic morphology
not captured by grade. Within intermediate-grade patients (ISUP 2–
3), where grade-based stratification adds little additional resolution, the
framework significantly separated BCR-free survival curves, identifying
patients at elevated recurrence risk within the same grade category.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_032.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
BibTex
@InProceedings{RajDas_What_MICCAISAT2026,
author = { Raju, Dasari Naga},
title = { { What Do Pathology Foundation Models Learn Beyond Gleason Grade in Prostate Cancer? } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}
