Abstract
Recently, pathological diagnosis has achieved superior performance bycombining deep learning models with the multiple instance learning (MIL)framework using whole slide images (WSIs). However, the giga-pixeled nature ofWSIs poses a great challenge for efficient MIL. Existing studies either do notconsider global dependencies among instances, or use approximations such aslinear attentions to model the pair-to-pair instance interactions, whichinevitably brings performance bottlenecks. To tackle this challenge, we proposea framework named MamMIL for WSI analysis by cooperating the selectivestructured state space model (i.e., Mamba) with MIL, enabling the modeling ofglobal instance dependencies while maintaining linear complexity. Specifically,considering the irregularity of the tissue regions in WSIs, we represent eachWSI as an undirected graph. To address the problem that Mamba can only process1D sequences, we further propose a topology-aware scanning mechanism toserialize the WSI graphs while preserving the topological relationships amongthe instances. Finally, in order to further perceive the topological structuresamong the instances and incorporate short-range feature interactions, wepropose an instance aggregation block based on graph neural networks.Experiments show that MamMIL can achieve advanced performance than thestate-of-the-art frameworks. The code can be accessed athttps://github.com/Vison307/MamMIL.