Self-supervised comparative learning based improved multiple instance learning for whole slide image classification

The huge pixels of the Whole slide image (WSI) and the lack of pixel-level annotation make WSI classification tasks such as tumor and subtype classifications face great challenges. Multiple instance learning (MIL) is introduced as a powerful learning tool for WSI classification tasks, which only requires slide-level labels to achieve good results. In this paper, we introduce an efficient transformer multiple instance learning model based on large kernel attention LKA, using self-supervised contrast learning to extract instance features and combining local and long-range attentional information to achieve state-of-the-art results on open datasets for two different tasks. Our model can cope well with various WSI classification tasks such as tumor classification, subtypes classification, etc. The AUCs for tumor classification on the CAMELYON16 datasets and cancer subtypes classification on TCGA-NSCLC datasets can reach 0.9778 and 0.9754, respectively.