A Multi-View Weighted Fusion Model for Segmentation of Infant Brain Ventricles with Hydrocephalus in MRI

Hikari Maeda, Yutaro Iwamoto, Masahiro Nonaka, Yinhao Li, Yen-Wei Chen

Abstract: Brain ventricles of infants with hydrocephalus is a disease in which the accumulation of cerebrospinal fluid causes the abnormal expansion of the ventricular region of brain. The 3D MRI volumes are widely used to examine it. Segmentation of infant brain ventricles with hydrocephalus in MRI is an essential step for diagnosis. Since the infant brain ventricles with hydrocephalus have complicated and diverse shapes, the conventional 2D segmentation methods (e.g., U-Net) cannot achieve good performance because of lack of 3D information. On the other hand, the 3D models (e.g., 3D U-Net) can not be used for segmentation with limited training samples (volumes). In this paper, we propose a multi-view weighted fusion model for segmentation of infant brain ventricles with hydrocephalus. In the proposed method, we first use 2D U-Net to perform segmentation from three different views, i.e., axial, sagittal and coronal plans, respectively. Then we fused the results of the three views with different weights to form the final segmentation result. In addition, we also use z-score to normalize the MR images before segmentation. The segmentation accuracy is about 91.8% by setting higher weight on the axial view.