Abstract

Info

Title:: Auto-segmentation of organs-at-risk in abdominal CT after combining with synthetic-MRI information

Author(s):: MENG Xiangyin1; WU Xiangyi1; PENG Zhao1; XU Xie1; 2; PEI Xi 1; 2; 1. School of Nuclear Science and Technology, University of Science and Technology of China, Hefei 230026, China 2. Anhui Wisdom Technology Co., Ltd, Hefei 230088, China

Keywords:: Keywords: synthetic-magnetic resonance imaging CycleGAN abdomen auto-segmentation

Abstract:: Abstract: Objective To propose an abdominal multi-organ auto-segmentation model in combination with synthetic-MRI (sMRI) soft tissue information, thereby improving the performance of CT soft tissue segmentation. Methods Two independent deep neural networks were used to complete the auto-segmentations of abdominal organs-at-risk step by step. A model for converting CT images into sMRI images was constructed based on CycleGAN network, and a high-definition sMRI with consistent organ contours was obtained by an improved denoising discriminator. Then, auto-segmentation model Residual U-Net was trained by sMRI and manual segmentation information to obtain the contours of organs-at-risk in CT and sMRI. The residual module in Residual U-Net made full use of the extracted features to distinguish different organs. Dice similarity coefficient (DSC) was used as the evaluation criterion for the segmentation accuracy of the auto-segmentation model. A total of 35 patients with cervical cancer and 35 patients with prostate cancer were used for training and evaluation of the model. Results The average DSC of the auto-segmentation model for the rectum, bladder and left and right femoral heads was 0.779±0.021, 0.944±0.006, 0.834±0.006 and 0.845±0.021, respectively. Conclusion The abdominal CT auto-segmentation method which combines with sMRI information can obtain more accurate results in the rectum segmentation.