Abstract

Info

Title:: Gastric wall segmentation based on improved V-net convolutional neural network

Author(s):: ZHAO Chenglu1; FANG Zhijun1; GAO Yongbin1; WANG Hailing1; WEI Ziran2; CAI Qingping2; 1. School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai 201620, China 2. the Second Department of General Surgery, Shanghai Changzheng Hospital, Shanghai 200003, China

Keywords:: Keywords: gastric cancer CT image image analysis convolutional neural network gastric wall segmentation

Abstract:: Abstract: The realization of gastric wall segmentation and centerline extraction in upper abdominal CT images is the premise of successful implementation of early gastric cancer screening and assisting in T staging. Gastric wall segmentation method based on improved V-net which uses the fully convolutional network framework with global average weight module can effectively solve the problem of information loss in neural network subsampling. In addition, a regularized level set loss function is proposed based on the original level set method for effectively suppressing the loss rate of gastric wall edge features and the over-fitting caused by the small amount of data, thereby improving the recognition accuracy of the gastric wall in upper abdominal CT images using neural network. The experiment shows that the Dice coefficient of the proposed method for gastric wall segmentation in upper abdominal CT image data set is up to 0.916 5, and IOU reaches 0.822 3. The segmentation accuracy of the proposed method is improved by 6% compared with 3D-Vnet method, and that is improved by 2.7% and 3.1% compared with CE-net method and Dense U-net method, respectively.