Boundary-Aware Uncertainty Suppression for Semi-supervised Medical Image Segmentation

Semi-supervised learning (SSL) algorithms have received extensive attention in medical image segmentation because they can be trained with unlabeled data. However, most existing SSL methods underestimate the importance of small branches and boundary regions, resulting in unsatisfactory boundaries and nonsmooth objects. We observe that the voxels of the target boundary have relative uncertainty. When the foreground map and background map of an object have the same voxel, that voxel must be in the edge region. Therefore, in this study, we propose a novel SSL framework based on the uncertainty of bounding voxels, which we call the boundary-aware network (BoANet). Specifically, we use a dual-task network that predicts the segmentation map and background map of objects. For unlabeled data, because the geometric contour information of the target object is obtained by elementwise multiplication of the segmentation map and the background map, geometric constraints are imposed on the segmentation. Simultaneously, for labeled data, we propose a weighted cross-entropy (
wce
) loss, which can synthesize the local structural information of voxels and guide the network to mine boundary details. We evaluated our method on publicly available benchmark datasets. The experimental results show that our method can outperform the current state-of-the-art approaches.