Automatic Brain Mask Segmentation for Mono-modal MRI

In recent years, deep learning methods have gained promising results in different kinds of image processing tasks, such as image classification, semantic segmentation, image generation and so on. This paper focuses on the research of brain masking for monomodal MRI, structural MRI, which is the most commonly used by the clinic and research. The brain mask is a basic and essential tool for brain function analysis and voxel-based structural analysis. In this paper, we present an automatic method for brain masking which would match the brain atlas for the origin image and also extract the regions of interest (ROI), like Hippocampus. Our network is developed from the U-net and a coarse mask is added into the network, which is generated by the method of region seeds growing. The combination of coarse mask and origin input speeds up the localization of the network and also increases the segmentation accuracy. In this work, two groups of experiments have been carried out, the one to do the brain mask automatically for the whole brain and the other for the region of Hippocampus extraction. Finally we have gained 0.893 dice coefficient for Hippocampus and 0.865 for the whole brain regions in average.