基于全局特征U-net的胰腺图像分割  

Pancreas segmentation based on U-net with global features

作　　者：向智霆 刘剑聪 魏柳 王淇锐 简丽琼[2] 肖斌[1,3] XIANG Zhiting;LIU Jiancong;WEI Liu;WANG Qirui;JIAN Liqiong;XIAO Bin(College of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,P.R.China;Ningxia Hui Autonomous Region Blood Center,Yinchuan 750001,P.R.China;Chongqing Key Laboratory of Image Cognition,Chongqing 400065,P.R.China)

机构地区：[1]重庆邮电大学计算机科学与技术学院,重庆400065 [2]宁夏回族自治区血液中心,银川750001 [3]图像认知重庆市重点实验室,重庆400065

出　　处：《重庆邮电大学学报（自然科学版）》2022年第2期216-222,共7页Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)

基　　金：国家重点研发计划基金(2016YFC1000307-3);国家自然科学基金(61976031,61806032);重庆市基础与前沿项目(cstc2018jcyjAX0117);重庆市教委科学技术研究计划重点项目(KJZD-K201800601)。

摘　　要：为了在医学图像中提高胰腺计算机断层成像(computed tomography, CT)自动分割的准确率,针对传统分割方法存在受噪声影响大、过分割、欠分割等问题,以及胰腺周围的重要结构组织关系紧密且多变、边缘界限不易确定等特点,提出了一种基于全局特征U-net(U-net with global features, GF U-net)的胰腺图像分割方法。该方法比基于传统深度卷积神经的U-net网络能够提取出更精确的形状、纹理信息,将胰腺图像区域的毛刺边缘进行平滑化,能够更好地把握胰腺的全局特征。通过对82个由美国国立卫生研究院(national institutes of health, NIH)公开的胰腺CT数据进行四折交叉验证,得到Dice相似系数(Dice similariy coefficient, DSC)的均值为87.13%±3.76%,比传统的U-net网络增长了7.43%。提出的方法不仅拥有更高的准确率,而且生成胰腺的形状边缘更加契合生物学上的胰腺形状,更容易应用在临床医学中。In order to improve the accuracy of automatic segmentation of computed tomography(CT) of pancreas in medical images, a pancreatic image segmentation method based on U-net with global features(GF U-net) is proposed to solve the problems of large noise, over-segmentation and under-segmentation in traditional segmentation methods, as well as the characteristics of close and changeable relationship between important structures and tissues around the pancreas and difficult to determine the edge boundary. Compared with the traditional U-net network with deep convolutional neural network, this method can extract more accurate shape and texture information, smooth the burr edge of the pancreatic image region, and better grasp the global features of the pancreas. Four fold cross-validation was performed on 82 CT data from the publicly available pancreatic data from the NIH. The experimental average DSC is 87.13±3.76, increasing by 7.43% compared with the original U-net network. The method in this paper not only has a higher accuracy rate, but also the shape edge of the pancreas is more consistent with the shape of the pancreas in biology, which is easier to be applied in clinical medicine.

关 键 词：U-net 胰腺分割 医学图像处理 全局特征 神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]