Brain MRI segmentation

Abstract

The automated segmentation of Brain MRI is an important initial step in the application of anomaly detection and pain localization due to neurological diseases. In this view, we aim to segment the sub structures (Caudate, Putamen, Pallidus and Accumbens) in the Basal-Ganglia region of the brain which is responsible for controlled movement and routine learning. In this paper we explore the possibility of using a convolutional neural network (SegNet) with long skip connections for BG Segmentation. Segnet allows us to upsample the lower resolution in- put features and does so with max-pooling weights only, thus reducing the training parameters significantly and long skip connections are used to skip features from the downsampling path to the expanding path in order to recover spatial information lost. Additionally we introduce the use of 3D patch along with their motion boundaries as concatenated channels in the input to the network which help in improved boundary segmentation. We use a public dataset of 83 T-1 weighted structural of healthy and disease MRIs available at [7] and produce a train and test split of 50 and 33 mixed (healthy and all diseases) samples of MRIs respectively to ensure a robust network.