Rethinking Overfitting of Multiple Instance Learning for Whole Slide Image Classification

Multiple instance learning(MIL) is widely used for whole slide image(WSI) classification. However, these methods suffer from severe overfitting. In this paper, we introduce two main causes of such overfitting problems by rethinking the MIL task and formulation of attention-based MIL models: (i) The model is sensitive to the proportion of positive regions, and (ii)incorrectly learns the positional relationship of patches (i.e., the order of instances). To this end, we propose recurrent random padding(RRP) module and patch shuffle(PS) module to tackle these two issues, respectively. Furthermore, we present random alignment(RA) algorithm to solve these two overfitting problems simultaneously. On CAMELYON16 and TCGA-NSCLC, the proposed plug-and-play modules improve the performance of six baselines by large margins. The significant and consistent refinement demonstrates the correctness of our theories and the effectiveness of our modules.