Complementary Attention Based Dual-Task Pseudo-labeling for Medical Image Segmentation

Recent advances in semi-supervised learning (SSL) demonstrate that a dual-task consistency regularization rather than implicitly constructing data- or network-level perturbation can effectively improve image segmentation accuracy. Furthermore, a dual-task network shows a better capacity for encoding semantic features compared to a solely segmentation task. However, existing dual-task based methods ignore the inherent information loss of network. To address this problem, we develop a complementary attention module which utilizes a complementary feature to strength the spatial- and channel-wise attention. Besides, recent segmentation works show consistency regularization with well-designed pseudo-labels has great veiled potentials. However, existing pseudo-label methods mainly focus on reliable regions, most pixels may be left unused due to their unreliability. We argue that every pixel matters to model training, and we develop a dual-task pseudo-labeling method to make sufficient use of both reliable and unreliable regions. Experimental results on two benchmarks demonstrate the advantage of our approach over the current state-of-the-art methods.