Abstract

Info

Title:: Automatic liver segmentation based on three-dimensional convolutional neural network

Author(s):: HE Lan; WU Qian; South-Central University for Nationalities, Wuhan 430000, China

Keywords:: ; Keywords: three-dimensional convolutional neural network; depth supervision mechanism; graph cut; prior information; liver segmentation

Abstract:: Abstract: Primary hepatic malignant tumor is an extremely harmful tumor with high incidence in China. Liver surgery (such as tumor resection, living liver transplantation, etc.) is one of the main treatments for various common benign and malignant liver diseases. The accurate segmentation of liver tissue from medical images is a fundamental and crucial step in computer-assisted liver disease diagnosis and surgical planning. Concerning the specificity and challenge of liver segmentation, an automatic segmentation algorithm model based on three-dimensional convolutional neural network (3DCNN) is proposed. 3DCNN is capable of conducting volume-to-volume learning, which can learn the plane and spatial information of liver images well. Integrating the depth supervision mechanism into 3DCNN can effectively reduce the problem of gradient disappearance or explosion, and speed up the convergence while improving the resolution. Finally, using the initial segmentation result as a priori information, the graph cut algorithm based on multi-convex constraint is used for further segmentation. Experimental results show that the segmentation model can accurately segment liver tissue from abdominal CT images.