Segmentation of Corpus Callosum using Attention U-Net Architecture for MRI Scan

In this paper, an attention U-net-based deep learning method for the semantic segmentation of the corpus callosum (CC) from brain Magnetic Resonance Imaging (MRI) scans is proposed and implemented. Most neurological analyses benefit greatly from the structural data that can be obtained from the segmentation of brain MRI images. The proposed technique has a deep supervised encoder-decoder architecture and a redesigned attention network. Slice by slice, the model analyzes an entire MRI image to determine the ideal mask for corpus callosum. The model was trained using the ABIDE and OASIS datasets, and its performance was analyzed for different test samples using a standard measure of dice coefficient, yielding a dice accuracy of 93.5%. Visual samples of predicted CC from brain MRI are given and contrasted with the original ground truth to help understand how well the model performs. The findings demonstrate that the suggested approach is one of the best segmentation techniques, as it achieved very competitive CC segmentation performance even with a single model.