Clustering-Based Multi-instance Learning Network for Whole Slide Image Classification

Automated and accurate classification of Whole Slide Image (WSI) is great significance for the early diagnosis treatment cancer, which can be realized by Multi-Instance Learning (MIL). However, current MIL method easily suffers from over-fitting due to weak supervision slide-level labels. In addition, it difficult distinguish discriminative instances in a WSI bag absence pixel-level annotations. To address these problems, we propose novel Clustering-Based (CBMIL) classification. The CBMIL constructs feature set phenotypic clusters augment data training aggregation network. Meanwhile, contrastive learning task incorporated into multi-task learning, helps regularize process. centroid each cluster updated model, weights patches are calculated their similarity centroids highlight significant patches. Our evaluated on two public datasets (CAMELYON16 TCGA-Lung) binary tumor cancer sub-types achieves better performance interpretability compared with state-of-the-art methods. code available at: https://github.com/wwu98934/CBMIL .