Deep Learning on Histopathological Images for Risk Stratification in Indian Breast Cancer Patients

This is a single-center, retrospective observational study involving no patient intervention or direct patient contact. Existing archival hematoxylin and eosin (H&E)-stained slides from eligible hormone receptor-positive (HR+), HER2-negative breast cancer patients treated at TMC will be utilized. These slides have been previously processed as part of routine clinical management. Patients included will have available clinical data and a minimum of 5-year follow-up post-treatment. Whole-slide images will be digitized at a standardized magnification and resolution, converted into analyzable image sections (tiles), and processed using specialized image-analysis software. Imagederived features will be extracted using a self-supervised learning (SSL) model at Technion, Israel. The multimodal deep-learning model developed at Technion will integrate these features with recorded clinical variables (including patient age, tumor size, tumor grade, ER and PR status) to generate an AI-derived risk score per patient. A calibrated AI risk score will subsequently be determined. Statistical analyses will assess the association of the AI risk scores with patient outcomes, including distant recurrence-free interval (DRFI), recurrence-free interval (RFI), disease-free survival (DFS), and breast cancer-specific survival (BCSS). Patient data will be analyzed in batches, and power estimates will be refined as necessary based on interim analyses