CVMIL: Cluster Variance Multiple Instance Learning for Whole Slide Images Survival Prediction

Tumor survival prediction using whole slide images (WSIs) is a crucial application in pathology aimed at assisting doctors in better formulating post-surgical treatment plans. The key challenges in current WSIs survival prediction lie in the vast scale of WSIs and the scarcity of manual annotations, which hinders the extraction of effective information from WSIs. To address these issues, previous studies have mainly employed the multiple instance learning (MIL) approach. However, existing methods often fail to consider the complexity of tumors and integrate clinically relevant knowledge, leading to suboptimal outcomes in survival prediction. To capture the intricate characteristics of tumors, we propose Cluster Variance Multiple Instance Learning (CVMIL) framework capable of representing tumor heterogeneity. By leveraging the differences from cluster centers, CVMIL represents both intra-tumor and inter-tumor heterogeneity, thereby enhancing the performance of MIL methods in WSI survival prediction. Results from prognosis tasks conducted on three publicly available TCGA datasets and the in-house ARGO dataset demonstrate that our approach outperforms current state-of-the-art methods, enabling more effective prediction of patient prognosis.