Region-Aware Multiple Instance Learning for Whole Slide Image Classification: Dual-Layer Attention Network Inspired by Pathologists

Computer-aided pathological diagnosis based on whole slide image (WSI) classification plays an important role in clinical practice, and it is often formulated as a weakly supervised multiple instance learning (MIL) problem. In recent years, attention-based MIL methods have yielded a promising solution for WSI classification. However, these methods usually directly generate instance-level attention scores under weak supervision signals, which often leads to inaccurate attention localization. Moreover, they fail to model the contextual relationships among patches, which are crucial for the diagnosis of WSI, given the continuum of tissue organization. To overcome these issues, we propose a region-aware dual-layer attention MIL network (RAMIL) for WSI classification. RAMIL mimics the diagnostic process of a pathologist and divides attention generation into two steps, from region refinement to instances. Specifically, the region attention module first divides the WSI into different regions based on the spatial position relationship of the patch, evaluating their importance and optimizing the instance features. Then it can generate more reasonable instance attention and aggregate instance features through the instance attention module, achieving more accurate classification and stronger model interpretability. Extensive experiments on three large public benchmarks demonstrate that RAMIL significantly outperforms state-of-the-art WSI classification methods.