Consistency learning with dynamic weighting and class-agnostic regularization for semi-supervised medical image segmentation

Recently, significant progress has been made in consistency regularization-based semi-supervised medical image segmentation. Typically, a consistency loss is applied to enforce consistent prediction of input images under different perturbations. However, most of the previous methods missed two key points: (1) Only a single weight is used to balance the supervised and unsupervised loss, which fails to distinguish the variances across samples. (2) Only forces the same pixel to have similar features in different data augmentation, yet ignoring the relationship with other pixels, which can serve as a more robust supervision signal. To address these issues, in this paper, we propose a novel framework, which contains two main components: Dynamic Weight Sampling (DWS) module and Class Agnostic Relationship (CAR) module. Specifically, our method contains one shared encoder and two slightly different decoders. Instead of a fixed weight, the DWS dynamically adjusts the weights for each sample based on the discrepancy between the predictions of the two decoders, balancing the supervised and unsupervised losses. The greater discrepancy implies that the sample is more challenging, and the prediction is less reliable. To learn from a more reliable target, a lower weight should be assigned to the challenging sample. In addition, to convey the relationship between pixels as supervision, the CAR develops relational consistency loss on class-agnostic feature regions. Extensive experiment on the Left Atrium and Pancreas-CT dataset shows that our methods have achieved state-of-the-art results.