Abstract
This paper introduces MAD-MIL, a Multi-head Attention-based Deep MultipleInstance Learning model, designed for weakly supervised Whole Slide Images(WSIs) classification in digital pathology. Inspired by the multi-headattention mechanism of the Transformer, MAD-MIL simplifies model complexitywhile achieving competitive results against advanced models like CLAM andDS-MIL. Evaluated on the MNIST-BAGS and public datasets, including TUPAC16,TCGA BRCA, TCGA LUNG, and TCGA KIDNEY, MAD-MIL consistently outperforms ABMIL.This demonstrates enhanced information diversity, interpretability, andefficiency in slide representation. The model's effectiveness, coupled withfewer trainable parameters and lower computational complexity makes it apromising solution for automated pathology workflows. Our code is available athttps://github.com/tueimage/MAD-MIL.