Accurate Neuroanatomy Segmentation Based on Voxel-wise Densely Predicted Network Using Multi-task Learning
Jiaxin Yang, Ming Zhao, Hewei Cheng · 2024
Deep learning-based methods have achieved great progress in brain segmentation based on magnetic resonance (MR) images. In this paper, we proposed a new neural network, and adopted 3D patches as input to train the model through multi-task learning for accurate segmentation of neuroanatomy, referred to as VoxelNAT. This network was trained to label the central voxel of all possible input 3D patches, generating densely voxel-wise segmentation for each one of testing magnetic resonance MRI images. Additionally, we employed an online full patch sampling strategy to train the VoxelNAT. Experimental results demonstrated that our method exhibited excellent performance in brain MRI image segmentation than alternative state-of-the-art methods. Our VoxelNAT holds significant potential for application in brain image segmentation.