Deep learning for corpus callosum segmentation

Abstract

Corpus callosum (CC) has been proved to be a sensitive biomarker of some neurological diseases such as multiple sclerosis (MS) and Alzheimer’s, and magnetic resonance imaging (MRI) is currently the priority technique for quantitative diagnosis and analysis of corpus callosum, in which segmentation of the CC from MRI is a vital step. During the past few years, various algorithms have been explored to fulfill the task of CC MRI segmentation. However, specific engineering knowledge and experience are still required to build such segmentation models. Furthermore, these algorithms tend to have a poor generalization ability. Therefore, a deep learning method is applied to develop a segmentation algorithm. In this dissertation, corpus callosum segmentation network (OpccNet), a CC MRI segmentation algorithm is developed, based on OpenCC, a public benchmark MRI dataset for assessing the CC segmentation algorithms. OpccNet and its extension Attention OpccNet, generated by adding attention gates to the OpccNet are trained and compared. They both produce an accurate and immediate segmentation of CC, providing an automatic segmentation of CC from MRI images with deep learning. Additionally, different augmentation methods are compared to discuss the importance of noisy data for building up a stable and reliable segmentation system in real medical applications.