Boundary-Aware Voxel-Wise Contrastive Learning for Infant Brain Tissue Segmentation

The accurate segmentation of infant brain magnetic resonance (MR) images is crucial for understanding the structure and early detection of neurodevelopmental disorders. However, it remains challenging mainly because of the inverted contrast between white matter and gray matter, and high-intensity inhomogeneity. Existing deep-learning segmentation methods do not fully learn the boundary features of brain tissues, resulting in difficulties in improving the segmentation accuracy of brain tissue boundaries. In this study, a boundary-aware voxel-wise contrastive learning method is proposed for infant brain tissue segmentation by designing a voxel-wise contrastive learning loss to represent the distance among positive samples, negative samples, and anchors in the embedding space. The proposed method was tested on a 0.35 T low-field infant brain MR dataset and the MICCAI iSeg-2019 challenge dataset. Compared with the baseline methods, the proposed method achieved an average improvement of 1.92% and 0.81% in the Dice index for different brain tissues on the respective datasets, thereby significantly enhancing the accuracy of brain tissue segmentation.