Automatic Segmentation of Infant Brain Ventricles with Hydrocephalus in MRI Based on Deep Multi-path Learning

Infant Brain Ventricles with Hydrocephalus is a disease of hydrocephalus that occurs in children, in which cerebrospinal fluid accumulates in the ventricles and the ventricles expand abnormally. The ventricles have the potential to cause brain damage by compressing other brain tissues. It is crucial to extract the ventricles with less burden for early detection and postoperative follow-up. However, automatic segmentation of infant brain ventricles with hydrocephalus is a challenging task; especially for those with hydrocephalus because they have complicated and diverse shapes. Further, because preparing a large amount of annotated data is challenging, it is necessary to train with a small amount of data. Thus, achieving an accurate segmentation with conventional deep learning is challenging. We proposed a deep multi-path learning approach for the accurate segmentation of infant brain ventricles with hydrocephalus in this study. In the proposed method, we developed three deep learning models for axial, sagittal, and coronal planes, then integrated the results of the models to obtain the final segmentation result. With a minimal amount of data, our proposed method acquired massive features. The segmentation accuracy of our proposed method increased from 74.3% to 81.1%, when compared with the related method.