Prediction Of Homologous Recombination Deficiency (HRD) From H&E Whole Slide Images Using Attention-Based Multiple Instance Learning In Ovarian Cancer

Homologous Recombination Deficiency (HRD) is a robust, but complex to calculate, predictive biomarker for stratifying patients likely to benefit from PARP inhibition therapy. Although computational pathology algorithms have shown potential in predicting HRD status from routine hematoxylin and eosin (H&E) images, this study evaluates their efficacy specifically in ovarian cancer where the few existing studies have reported suboptimal performance. Genomic instability (GI) index and HRD status were calculated from sequencing data in 484 patients (52% HRD-positive). We adapted two state-of-the-art deep learning algorithms based on classification and regression approaches. For regression we predicted the GI and used zero as threshold to define HRD status. H&E images were split into non-overlapping patches; a pre-trained RetCCL foundation model was used to extract 2048 features from each patch after background extraction and stain-normalization. An attention-based multiple instance learning algorithm was trained on the extracted features to predict GI and HRD status, and generate interpretable feature-importance maps. Additionally, we experimented by applying automatic semantic segmentation, using a Unet model with ResNet50 encoder, to drive the algorithm’s attention on cancer tissue. We used five-fold cross-validation on 380 patients and applied the best fold in a hold-out test set of 104 patients. Our deep learning models predicted HRD with moderate accuracy in the validation and test sets both for classification (AUC=0.67/0.60, F1-score=0.57/0.50 respectively) and regression (F1-score=0.74/0.57 respectively) approaches. Interestingly, despite the fact the HRD is expressed by the cancer cells, attention maps instead showed focus on ovarian stroma areas [Figure 1A]. Focusing attention on tumour areas showed improvement and increased our algorithm’s ability to generalise in the test set (classification: AUC=0.69/0.63, F1-score= 0.76/0.65; regression: F1-score=0.74/0.60 for validation/test) [Figure 1B]. Computational pathology has the potential to predict HRD from H&E images in ovarian cancer. Guiding attention-based mechanisms with cancer tissue segmentation can improve the algorithm’s stability.