A Multiple Instance Learning framework with Instance Identification and Supervised Contrastive Learning for WSI Classification

Multiple Instance Learning (MIL) has been widely used for Whole Slide Image (WSI) classification by dividing a gigapixel WSI into lots of patches, termed bag-of-instances. In order to derive more discriminative instance features, some MIL methods introduce a pseudo-label generation tactic to attach pseudo labels to instances based on attention scores and train an instance-level classifier. However, depending on the attention mechanism completely is not very robust in the correctness of pseudo labels of instances. In addition, the imbalance problem of positive versus negative examples in WSI classification is often neglected but has some influence on the performance. To address the two issues, this paper presents a novel MIL framework with instance identification and supervised contrastive learning for WSI classification. Specifically, we adopt a two-step instance identification strategy, where top-k instances are chosen first according to attention scores and then filtered using the estimated negative-instance distribution, effectively reducing the noise on pseudo labels of instances. Secondly, we introduce supervised contrastive learning at the bag level, which enhances intra-class compactness and inter-class separability to alleviate the bag class imbalance problem. To this end, we use a very efficient random instance masking strategy to generate the augmented view for a bag and maintain two dynamic queues to construct sufficient and diversified bag pairs. Experimental results on three publicly available WSI data sets show that the proposed framework is superior to or highly comparable to state-of-the-art methods. Meanwhile, qualitative visualizations highlight the effectiveness and superiority of our framework in identifying tumor regions.