Multi-Scale Context-Aware Survival Prediction Network Based on Whole Slide Images

High-resolution Whole Slide Images (WSIs) are essential for survival prediction as they contain key multi-scale information that reflects tumor characteristics at various levels of magnification. However, the large size of WSIs presents significant challenges for existing methods, making it difficult to efficiently extract and utilize this multi-scale data. Unlike traditional WSI-level multiple instance learning (MIL), survival prediction tasks require a more sophisticated patient-level analysis, which takes into account data from multiple WSIs per patient. To address these challenges, we propose the Multi-Scale Context-Aware Survival Prediction Network (MS-CASurv), a novel approach designed to enhance the accuracy of survival predictions. MS-CASurv is built with a three-layer architecture that progressively analyzes tumor micro environments, associated tissues, and patient-specific heterogeneity. The local-level interaction layer captures detailed features, the WSI -level interaction layer integrates multi-scale data, and the patient-level interaction layer synthesizes information from multiple WSIs for a comprehensive patient profile. Experimental validation on a dataset of 3068 H&E-stained WSIs from 2422 patients across five distinct cancer types from TCGA shows that MS-CASurv outperforms existing weakly supervised methods. Our results demonstrate that MS-CASurv effectively integrates multi-scale data, offering more accurate survival risk assessments at the patient level. This framework sets a new standard for survival prediction accuracy and paves the way for advancements in personalized cancer treatment and prognosis.