Robust 3D Convolutional Neural Network With Boundary Correction for Accurate Brain Tissue Segmentation

The morphology, symmetry, and volume of brain tissue are good indicators for measuring the central nervous system disease progression. The objective of this paper is to segment cerebrospinal fluid (CSF), gray matter (GM), and white matter (WM) automatically with multi-modality magnetic resonance scans. A novel coarse-to-fine method is proposed to segment CSF, GM, and WM using two cascade 3D convolutional neural networks. The first densely connected fully convolutional network (DC-FCN) is designed with feature reuse, which can take full advantage of the spatial information and alleviate computer memory limitation. The second 6-CNN is designed to correct boundary voxel, which can further reduce computational cost while improving the segmentation accuracy. As of today, our method ranks the 3rd on the MRBrainS13 challenge, outperforming most of the participant methods when using available input modalities (T1, T1-IR, and T2-FLAIR). In addition, we also verify the proposed framework on the IBSR dataset, which demonstrates the effectiveness of the boundary correction strategy. Through accurate segmentation of brain tissue, neuroimaging physicians can be assisted in assessing disease progression and even localizing lesions.