Neonatal Brain MRI Image Segmentation Using U-Net With Enhanced Edge Detection Layers

Due to low contrast and rapid development of immature brain tissues, MRI brain segmentation for infant brains is a more challenging task than adult brain MRI segmentation. For this research, enhanced edge detection layers were implemented in a U-Net framework by using guided filter modules (GFM). Ablation experiments were done to see the effectiveness of the modules in atlas-based 6-month infant MRI brain segmentation. This paper focused on the initial evaluation of the proposed network using the iSeg19 training and validation dataset from the Medical Image Computing and Computer Assisted Intervention (MICCAI) challenge. Dice coefficient, modified Hausdorff distance (MHD) and ASD were measured for each brain region using the variations of the model: (i) vanilla U-Net, (ii) single-edge detection layer with U-Net, (iii) U-Net with GFMs during the training and validation process. The benchmark model used for comparison is the latest model listed in the MICCAI iSeg challenge.