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