Using deep learning models to predict homologous recombination deficiency (HRD) in patients with ovarian cancer based on whole slide images at 5×, 10×, 20×, and 40× magnifications

초록

Objectives Homologous recombination deficiency (HRD) status is a powerful predictive biomarker for the efficacy of poly(ADP-ribose) polymerase (PARP) inhibitors in patients with epithelial ovarian and primary peritoneal cancer. However, HRD testing is expensive and often requires several weeks to complete. We developed a deep learning model to predict HRD status by integrating a multi-instance learning (MIL) framework with UNI, a foundational model. Methods In this multicenter study, we analyzed hematoxylin and eosin (H&E) slides obtained from both primary and metastatic sites of newly diagnosed patients with ovarian or primary peritoneal cancer. Whole-slide images were scanned at 40× magnification, comprising 113 slides from 26 HRD-negative patients and 104 slides from 41 HRD-positive patients. Slides were downsampled to 20×, 10×, and 5× magnifications and segmented into small patches, yielding 1,549,550 negative and 1,369,860 positive patches. Patches were split into validation and test sets, and HRD prediction was performed using a two-layer model that combined UNI and MIL with varying bag sizes. Results At 20× magnification with a bag size of 64, the model achieved an AUC of 0.973 (95% CI, 0.948–0.972), with a sensitivity of 0.895, specificity of 0.943, and F1 score of 0.895 on the test set. At 40× magnification with a bag size of 500, the model achieved an AUC of 0.975 (95% CI, 0.961–0.979), with a sensitivity of 0.816, specificity of 0.961, and F1 score of 0.873. Conclusions The two-layer model integrating UNI and MIL accurately predicted HRD status from 40× whole-slide images of ovarian and primary peritoneal cancer. Increasing bag size improved performance, particularly by enhancing specificity.