Multi-Scale Boundary Enhanced Pseudo Label Refinement for Semi-Supervised Medical Image Segmentation

In semi-supervised medical image segmentation, current methods typically generate pseudo labels at a fixed scale and refine them without sufficient attention to boundary regions. Such approaches limit the accuracy of pseudo labels, negatively impacting the learning performance from unlabeled data. To obtain more accurate pseudo labels, we propose a method called Multi-Scale Boundary Enhanced Pseudo Label Refinement (MBPLR), which consists of two novel modules. Multi-Scale Collaboration Module (MSCM) is designed to generate pseudo labels that integrate multi-scale information, allowing for complementarity between different scales. Boundary Enhanced Module (BEM) is introduced to enhance the accuracy of boundary regions by adjusting pixel-level weights, and amplifying correct pixels while suppressing incorrect ones. Our method is evaluated on both 2D and 3D datasets and achieves the best performance compared to existing state-of-the-art methods.