Deep Survival Analysis from Whole Slide Images: A Multiple Instance Learning Approach

Survival analysis plays a critical role in oncology for patient care, but analyzing Whole Slide Images (WSIs) presents challenges due to their immense size and inherent variability. Traditional approaches often rely on manual Region of Interest (ROI) selection, which introduces subjectivity and limits scalability. In this paper, we propose Surv-MIL, a novel deep survival model based on Multiple Instance Learning (MIL) that processes WSIs without the need for ROI selection. Our approach divides WSIs into patches, extracts features using a pre-trained encoder, and then aggregates this information using a gated attention mechanism. This enables the model to focus on salient tumor regions while effectively handling the variability in WSI sizes. Evaluated on a real-world dataset, our method demonstrates superior performance compared to other deep survival models. This scalable framework effectively leverages the rich information contained within WSIs for survival analysis, potentially leading to improved prognosis prediction and treatment planning in oncology.