MSAA-Net: Multi-Scale Attention Assembler Network Based on Multiple Instance Learning for Pathological Image Analysis

In this paper, we present a multi-scale attention assembler network (MSAA-Net) tailored for multi-scale pathological image analysis. The proposed approach identifies essential features by examining pathological images across different resolutions (scales) and adaptively determines which scales and spatial regions predominantly influence the classification. Specifically, our approach incorporates a two-stage feature integration strategy. Initially, the network allocates the attention scores to relevant local regions of each scale and then refines the attention scores for each scale as a whole. To facilitate the training of the MSAA-Net, we employ the technique of multiple instance learning (MIL), which enables us to train the classification model using the pathologist’s daily diagnoses of whole slide images without requiring detailed annotation (i.e., pixel-level labels), thereby minimizing annotation effort. We evaluate the effectiveness of the proposed method by conducting classification experiments using two distinct sets of pathological image data. We conduct a comparative analysis of the attention maps generated by these methods. The results demonstrate that the proposed method outperforms state-of-the-art multiscale methods, confirming the effectiveness of MSAA-Net in classifying multi-scale pathological images.