CAM-TMIL: A Weakly-Supervised Segmentation Framework for Histopathology based on CAMs and MIL

Abstract

Semantic segmentation plays a significant role in histopathology by assisting pathologists in diagnosis. Although fully-supervised learning achieves excellent success on segmentation for histopathological images, it costs pathologists and experts great efforts on pixel-level annotation in the meantime. Thus, to reduce the annotation workload, we proposed a weakly-supervised learning framework called CAM-TMIL, which assembles methods based on class activation maps (CAMs) and multiple instance learning (MIL) to perform segmentation with image-level labels. By leveraging the MIL method, we effectively alleviate the influence caused by that CAMs only focus on discriminative regions. As a result, we achieved comparable performance with fully-supervised learning on Camelyon 16 only with image-level labels.