Brain MRI Segmentation using efficient 3D Fully Convolutional Neural Networks

Brain Magnetic Resonance Images (MRI) segmentation and analysis is a fundamental step in measuring brain anatomical structure and visualizing any change and development in the brain. Segmentation of MRI is very challenging due to low contrast of Grey Matter (GM) and White Matter (WM) tissues of the brain. We propose a 3D Fully Convolutional Neural Network (FCNN) for the brain MRI segmentation of 6 months infant into WM, GM and Cerebrospinal fluid (CSF) with the use of multimodality input of T1-weighted images and T2-weighted images. The proposed method employs careful tuning and adaptation of the architecture to drastically reduce the number of hyperparameters to increase efficiency while retaining comparable performance with the state-of-the-art. We evaluate our proposed method on the iSeg2017 infant MRI segmentation challenge, where we achieve state-of-the-art results, acquiring an average DSC score of 93%.