Progressive Pseudo Bag Augmentation with Instance Importance Estimation for Whole Slide Image Classification
In the field of computational pathology, the classification of whole-slide images (WSI) remains a challenging task due to the vast amount of gigapixel information and the limited availability of refined manual annotations. Recently, multiple instance learning (MIL) has emerged as a promising approach to address this issue. While attention-based MIL methods utilize attention mechanisms to distill instance information for training or further fine-tuning, the current ranking of attention scores fails to accurately locate positive instances. In this study, we propose the instance importance score (IIS) based on the Shapley value to tackle this problem. This approach enables the identification and prioritization of crucial features. Building upon this foundation, we present a novel framework for the progressive assignment of pseudo bags. Through comprehensive experiments, our approach achieves state-of-the-art performance compared to other superior methods on the CAMELYON-16, BRACS, and TCGA-LUNG datasets. Furthermore, the visualization results demonstrate the enhanced interpretability provided by the IIS in the classification of WSI. Code for our framework is accessible at https://github.com/*****.