MLCN: Metric Learning Constrained Network for Whole Slide Image Classification with Bilinear Gated Attention Mechanism

Whole Slide Image (WSI) classification is an important part of pathological diagnosis. Although previous approaches (such as DSMIL and CLAM) have achieved good results, the classification performance is still unsatisfactory because the learned features of WSI lack discrimination and the correlation among sub-characteristics of tumor images are ignored. In this paper, we proposed a Metric Learning Constraint Network (referred to as MLCN). Particularly, MLCN benefits from two aspects: 1) It enhances the discriminative power of features by enlarging inter-class distance and narrowing intra-class distance in both slide-level and patch-level. 2) It learns a more powerful feature aggregator by proposing the bilinear gated attention mechanism to capture relations among sub-characteristics of tumor issues. Experiments on CAMELYON16 and TCGA Kidney datasets validate the effectiveness of our approach, and we achieved state-of-the-art performance compared to other popular methods. The codes will be available soon.