Dynamic mask stitching-guided region consistency for semi-supervised 3D medical image segmentation

In semi-supervised medical image segmentation, consistency-based pseudo-label regularization may introduce cognitive bias, as it relies solely on models trained with limited labeled data to generate pseudo-labels. Moreover, the empirical mismatch between the limited labeled data and the abundant unlabeled data further exacerbates this issue, which is a common challenge in semi-supervised learning. To address these challenges, we propose a dynamic mask stitching-guided region consistency network that uses segmentation probability maps to generate dynamically changing masks for stitching during training. As training progresses, the mask converges to the correct segmentation target, directing the model's focus to the target's edges and mitigating cognitive bias from inaccurate predictions. Our experiments show varying prediction accuracy across different voxel source regions in the stitched image, highlighting the potential for improving segmentation through voxel-level region consistency supervision. This method uses the Ground Truth to supervise the voxel regions in the stitched image from labeled images, and the segmentation results of stitched and unlabeled images are used to construct low-entropy pseudo-labels for regional consistency regularization. We demonstrate the effectiveness of dynamic mask stitching between labeled and unlabeled data, achieving a 0.47–4.64 % improvement on the LA and Pancreas-CT datasets compared to state-of-the-art methods in semi-supervised segmentation. The code is accessible at https://github.com/Oubit1/DSRC.