BDB-Net: Boundary-Enhanced Dual Branch Network for Whole Brain Segmentation

Whole brain segmentation is of large value for scientific research and clinical practice. Because of the long processing time of multi-atlas based segmentation methods, deep learning based methods have been extensively studied in recent years. Under the framework of the state-of-the-art method spatially localized atlas network tiles (SLANT), we proposed a novel Boundary-enhanced Dual Branch Network (BDB-Net) for whole brain segmentation. It contains a semantic segmentation branch and a boundary detection branch. Through delicate coupling of these branches, both the texture and shape information could be better utilized. Besides, we also designed a dual-heads architecture for the segmentation branch to incorporate more context information and reduce the information truncation. Furthermore, in view of the symmetricity of the brain, we proposed a new splitting pattern of the volumes and reduce the number of networks. The utility of these improvements was validated through experiments conducted on two dataset. The experimental results shown the proposed method could obtain better results using less training time (97% less) and testing time (50% less) compared with the SLANT.