基于深度学习的磁共振胆胰管成像图像肝外胆管及结石的自动分割:初步研究  

Deep Learning for Segmentation of Extrahepatic Biliary Duct and Bile Duct Stones on Magnetic Resonance Cholangiopancreatography:A Pilot Study

作　　者：王可[1] 杨俊哲[1] 刘义[1] 马帅[1] 刘婧[1] 张耀峰 王祥鹏 张晓东[1] 王霄英[1] WANG Ke;YANG Junzhe;LIU Yi;MA Shuai;LIU Jing;ZHANG Yaofeng;WANG Xiangpeng;ZHANG Xiaodong;WANG Xiaoying(Department of Radiology,Peking University First Hospital,Beijing 100034,China;不详)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]北京赛迈特锐医学科技有限公司,北京100011

出　　处：《中国医学影像学杂志》2022年第7期703-709,715,共8页Chinese Journal of Medical Imaging

摘　　要：目的探索使用深度学习方法在磁共振胆胰管成像(MRCP)图像上分割肝外胆管和检出结石的可行性。资料与方法回顾性收集2019年7月5日—2020年6月30日于北京大学第一医院就诊的225例患者共230人次3D MRCP检查图像数据纳入胆管分割研究;并补充2020年7月1日—2021年2月27日通过经内镜逆行胆胰管成像或临床综合诊断证实的胆总管结石患者数据,最终合计73例存在胆总管结石患者3D MRCP检查图像纳入胆管结石分割研究。由2位影像科专家标注数据,得到267个肝外胆管和98个胆总管结石区域的标签。使用Unet3D网络分两步(coarse-fine)训练胆总管分割模型,将267个数据随机分为训练集213个、调优集27个和测试集27个。以胆总管标签为掩膜(mask),进一步训练胆总管结石的分割模型,将98个数据随机分为训练集80个、调优集9个和测试集9个。使用客观评价指标为测试集的Dice系数,并输出标注区域与模型预测区域的径线、体积等进行比较。主观评价指标包括肝外胆管分割评分、肝外胆管轴位T2WI匹配评分和结石分割评分。结果肝外胆管分割模型的测试集共27个数据,第一步(coarse)分割肝外胆管的Dice值为0.89±0.07,第二步(fine)分割肝外胆管的Dice值为0.94±0.04。胆总管结石分割模型的测试集共9个数据,分割结石的Dice值为0.83±0.06。主观评价20个临床确诊肝外胆管结石患者数据,肝外胆管分割评分均为满分10分,肝外胆管轴位T2WI匹配评分中位数为9.75分,结石分割评价中位数为8分。结论通过深度学习方法在MRCP图像上分割肝外胆管是可行的,能较准确地分割胆管结构,并用于结石和胆管梗阻的定位。Purpose To evaluate the feasibility of segmentation of extrahepatic biliary duct and detection of bile duct stones based on deep learning algorithms on magnetic resonance cholangiopancreatography(MRCP)images.Materials and Methods A total of 2303D MRCP scans(225 patients)were retrospectively collected for extrahepatic biliary duct segmentation in Peking University First Hospital from July 5,2019 to June 30,2020.Seventy-three series proved of stone in the common bileduct by endoscopic retrograde cholangiopancreatography or clinical comprehensive diagnosis were selected for the stone segmentation,including supplementary cases from July 1,2020 to February 27,2021.The images of extrahepatic biliary duct(n=267)and the stone area of the common bile duct(n=98)were annotated by two radiologists.A cascade Unet3D network was used to develop the cascade segmentation model(coarse-fine)of the extrahepatic biliary duct,and the data were further randomly divided into three groups,including training set(n=213),validate set(n=27)and test set(n=27).Taking the label of the extrahepatic biliary duct as a mask,the stone segmentation model was further trained,and the data(n=98)were randomly divided into three groups,including training set(n=80),validate set(n=9)and test set(n=9).The Dice coefficient of the test set was performed to objectively evaluate and output the performance of the model,such as diameter and volume of the annotation region and prediction region,respectively.Subjectively,evaluation indicators were as follows:(1)the segmentation performance of extrahepatic biliary tract;(2)the performance of image registration on axial T2 weighted image(T2WI);(3)the segmentation performance of stone.Results The test set of segmentation of the extrahepatic biliary tract consisted of 27 series of MRCP image,and the Dice coefficients of the two steps were 0.89±0.07(coarse)and 0.94±0.04(fine),respectively.The test set of segmentation of the stones consisted of 9 image series,and the Dice coefficient was 0.83±0.06.The subjective score

关 键 词：胆管 肝外 胆管疾病 胆管胰造影 磁共振 分割 深度学习 

分 类 号：R657.4[医药卫生—外科学] R575.7[医药卫生—临床医学] R445.2