Reliable Region Mining for Weakly Supervised Medical Image Segmentation

Class Activation Map (CAM) is a common tool to address image-level weakly supervised image segmentation (WSIS) tasks, which mainly extract the most discriminative regions as pseudo-labels for the segmentation networks. However, it is difficult to accurately locate the salient regions since similar image intensities and low-contrast in the medical image field. In our work, we put in a novel WSIS model to exploit visual detail features to enhance the semantic expression ability of the object target. The model consists of two branches: the classification branch and the segmented branch. The classification branch includes two subbranches: local classification and global classification. Differing from existing local-global models, we employ expansion convolution to increase the receptive field for the global classification branch, while designing an affinity-aware module in the local network to obtain the original context information of the local image. The global and local classification branches work together to provide pseudo-labels for the segmentation branch. The experimental results show that our model outperforms other state-of-the-art methods with a DSC of 46.7% on the publicly available MSD dataset.