Interpretable Deep Learning Predicts the Molecular Endometrial Cancer Classification from H&E Images: A Combined Analysis of the Portec Randomized Clinical Trials

Background: Endometrial Cancer (EC) is molecularly classified into the POLE- mutated (POLE mut), mismatch-repair deficient (MMRd), p53 abnormal (p53abn) and no specific molecular profile (NSMP) subgroups. We aimed to develop an interpretable DL pipeline for image-based prediction of the four-class molecular EC classification (im4MEC), to identify morpho-molecular correlates and refine EC prognostication. Methods: Diagnostic hematoxylin and Eosin (H&E)-stained slides from 2028 EC patients of the combined PORTEC-1-2,-3 randomized trials and four clinical cohorts were included. im4MEC combined self-supervised learning and an attention mechanism to achieve optimal performance. Tiles with highest attention scores were reviewed to identify morpho-molecular correlates. Predictions of a nuclear classification DL model served to derive interpretable morphological features and their relative contribution in the profiling of each molecular class. Prognostic refinement was explored through morphological and Kaplan-Meier-based survival analyses. Findings: im4MEC achieved a macro-average area under the receiver-operating-characteristic curve (AUROC) on cross-validation of 0·874±0·189, and 0·876 on the independent test set PORTEC-3 with highest performance of 0·928 among p53abn EC. Overall recurrence by image-based molecular class was significantly different in PORTEC-3 (p=1·e-04). Top-attended tiles indicated a significant association between dense lymphocyte infiltrates and POLE mut and MMRd EC, high nuclear atypia and p53abn EC, and an overlap of morphological representations between POLE mut and MMRd EC. im4MEC highlighted low tumor-stroma ratio as a potential novel characteristic feature of NSMP EC. p53abn cases predicted as imMMRd showed inflammatory morphology and better prognosis; NSMP predicted as imp53abn showed nuclear atypia and worse prognosis; MMRd predicted as im POLE mut had excellent prognosis. Interpretation: We present the first interpretable DL model for H&E-based prediction of the molecular EC classification. im4MEC robustly identified morpho-molecular correlates and enables further prognostic refinement of EC patients. Trial Registration: The PORTEC clinical trials are registered at clinicaltrials.gov PORTEC-3 with identifier: NCT00411138 (this trial is used for as independent testset for our model) PORTEC-2 with identifier: NCT00376844, PORTEC-1 was published in the Lancet in 2000 (attached) and ran 1990-1997. It was not registered in clinicaltrials.gov Funding: The Hanarth Foundation. V.H. Koelzer reports grants from the Swiss Federal Institute of Technology Strategic Focus Area: Personalized Health and Related Technologies PHRT and the Promedica Foundation (F-87701-41-01) during the conduct of the study. Declaration of Interest: All authors declare that they have no conflicts of interest in relation to this paper. Ethical Approval: The protocol was approved by the Protocol Review Committee of the Dutch Cancer Society and by the medical ethics committees of the University Hospital Rotterdam/Daniel den Hoed Cancer Centre (DDHCC) and of the participating centres. Keywords: deep learning, Endometrial Cancer, Molecular classification, Morphological features, Prognostic refinement, POLEmut EC, MMRd EC, NSMP EC, p53abn EC, whole slide images, Histopathology image