Abstract: Multi-instance Learning (MIL) has become the mainstream paradigm for handling super-high resolution Whole Slide Images in digital pathology. Current MIL-based methods leverage various attention mechanisms to aggregate instance features; however, they tend to concentrate attention scores on a small subset of instances, resulting in limited focus regions and inadequate recognition of diverse lesion areas with heterogeneous structures and multiple pathological patterns in histopathological images. To address this limitation, we propose MSM-MIL, a multi-stage masked multi-instance learning framework for histopathological image classification. Specifically, it first leverages a gated attention model to derive the initial bag-level embedding and initial mask, then mines the diverse pathological pattern features within a bag stage by stage through mask stacking, breaking the limitation of focusing only on a single salient feature in a single stage. Finally, it aggregates multi-stage bag-level embeddings via attention mechanisms for ultimate classification. Experimental results on two datasets demonstrate that the proposed framework outperforms existing state-of-the-art methods.

Key words: Pathology Image Classification, Multi-Instance Learning, Attention Mechanism, Masking