Weakly-supervised semantic segmentation in histology images using contrastive learning and self-training.

This paper presents a novel method for weakly-supervised semantic segmentation (WSSS) of histology images, where only global image-level labels are employed. We leverage an existing weakly-supervised object localization (WSOL) method to generate class activation maps (CAMs) indicating the spatial locations of relevant tissue regions. Next, we utilize a specialized encoder-decoder network to predict fine localization masks. A pixel-wise contrastive loss function is introduced to encourage the model to learn discriminative features for foreground and background regions. Additionally, a pixel-wise cross-entropy loss is incorporated for improved pixel-level supervision. An offline multi-round self-training strategy is also proposed to iteratively refine pseudo masks, enhancing segmentation performance. Our method demonstrates superior segmentation accuracy over the state-of-the-art method on the GlaS dataset (public benchmark for colon cancer). Furthermore, we investigate the efficacy of our approach in a mixed-supervision setting, achieving performance comparable to fully supervised models, indicating its practical applicability in clinical settings. Our results show that the proposed method offers an effective and practical solution for weakly-supervised semantic segmentation in histology images, potentially aiding pathologists in their diagnostic processes and facilitating the development of automated histopathological analysis systems.