CRP-MIL: A channel re-partitioning based multiple instance learning method for whole slide image classification

Whole Slide Images (WSIs) possess ultra-high resolution and contain abundant pathological information, leading to highly complex modeling processes. Current Multiple Instance Learning (MIL) models for WSI analysis suffer from two major limitations: (1) inadequate modeling of contextual relationships between tumor regions and neighboring image patches, and (2) intrinsic constraints of the MIL framework itself. WSIs must be partitioned into multiple independent image patches, making the effective modeling of global semantic information a substantial challenge in this field. To address these issues, this study proposes a WSI analysis method based on Channel Re-Partitioning (CRP) to mitigate the limitations of MIL in local-global information modeling. This method captures local relational information along the channel dimension and establishes global relational information across different channels. Unlike previous methods focusing mainly on patch-level relationships, CRP adopts a channel-centric approach to capture intra- and inter-channel dependencies, thereby enabling more effective integration of local and global information. Extensive experiments on three public datasets demonstrate that CRP-MIL achieves superior performance compared with state-of-the-art MIL methods. Moreover, CRP-MIL consistently maintains its advantage across different feature extraction backbones.