A Boundary-Constrained Approach for 3D Medical Image Segmentation

In 3D medical images, a small number of voxels are distributed in the boundary areas with high information. These voxels usually lie near the decision boundary of the model and exhibit high predictive uncertainty. Enhancing the model's focus on challenging boundary regions improves stability and provides greater training value. However, most existing boundary loss functions have difficulty in imposing effective constraints on these voxels. Therefore, this study proposed two key technologies based on Signed distance map (SDM) to enhance the generalization performance of the model under unlabeled data. First, SDM is used to accurately locate the boundary area containing key information, and a patch-level contrastive loss function is designed to guide the model to learn discriminative features with boundary structure perception ability. Secondly, SDM is used to generate boundary deformation perturbations, and the robustness of the model to local geometric changes is improved through feature consistency constraints. Our comprehensive experimental validation on the LA and Pancreas-CT datasets demonstrate that the proposed method significantly outperforms the state-of-the-art in the boundary segmentation task.