← ResourcesPublications
Deep learning-based prediction of homologous recombination deficiency from histopathological whole-slide images in ovarian cancer
18 APR 2026
Abstract
Objective
Homologous recombination deficiency is a critical biomarker for guiding first-line treatment with poly(adenosine diphosphate-ribose) polymerase inhibitors in ovarian cancer. While genomic assays are the current standard, they remain costly, time-consuming, and often inaccessible worldwide. With the widespread adoption of whole-slide imaging in pathology, artificial intelligence offers a promising alternative for identifying homologous recombination deficiency directly from routine hematoxylin and eosin-stained slides.
Methods
We developed a fully automated deep-learning pipeline using attention-based multiple-instance learning and pre-trained vision transformers to predict homologous recombination deficiency status from whole-slide images. The model was trained on
Results
The model achieved an area under the curve of 0.73 on the internal test sets, 0.69 on The Cancer Genome Atlas cohort, and 0.70 on the PAOLA-1/ENGOT-ov25 phase III trial cohort (0.72 in BRCA wild-type patients), demonstrating strong transferability and generalization. High negative predictive values (up to 0.98) were observed, indicating strong performance in ruling out homologous recombination-proficient cases. Kaplan-Meier survival analyses in the PAOLA-1/ENGOT-ov25 phase III trial cohort demonstrated that artificial intelligence-predicted negative groups showed progression-free survival curves nearly identical to Myriad-defined negative groups, especially among BRCA wild-type patients.
Conclusions
Our results demonstrate that deep learning applied to whole-slide images can predict homologous recombination deficiency status with clinically meaningful accuracy and strong generalization across datasets. The model is particularly effective at identifying homologous recombination-proficient patients and may serve as a valuable tool to support or triage molecular testing. Integrating this artificial intelligence tool into routine pathology workflows could improve diagnostic efficiency, reduce costs, and accelerate access to targeted therapies in ovarian cancer.