基于深度学习在CT图像上分割胆囊的研究  

Study on automatic segmentation and measurement of gall bladder on CT images by using deep learningalgorithm

作　　者：高越 王可欣 张耀峰 孙玉梦 张晓东[1] 王霄英[1] GAO Yue;WANG Ke-xin;ZHANG Yao-feng(The Peking University First Hospital,Beijing 100034,China)

机构地区：[1]北京大学第一医院医学影像科,北京100034 [2]首都医科大学基础医学院,北京100069 [3]北京赛迈特锐医学科技有限公司,北京100011

出　　处：《放射学实践》2024年第6期755-760,共6页Radiologic Practice

摘　　要：目的:基于深度学习方法训练模型,研究其用于腹部CT图像上分割胆囊并自动测量的可行性。方法:从本院PACS系统搜集2016年1月12日至2021年5月28日行腹部CT检查的患者,从中选取1154位患者的1181次CT检查图像,共得到2559个图像序列用于训练模型。由2位影像科医师标注胆囊,将全部数据按8:1:1的比例随机分为训练集(n=2043)、调优集(n=245)和测试集(n=271),训练3D U-net模型分割胆囊并自动测量。另搜集2022年9月10-19日的腹部CT扫描图像,随机选取共141位患者的141次检查的270个图像序列作为外部验证数据集。以外部验证集的预测结果评价模型的效能。使用Dice相似系数(DSC)、体积相似度(VS)和Hausdorff距离(HD)定量评价模型分割胆囊区域的效能。使用Bland-Altman分析评价模型自动测量的胆囊体积、径线、平均CT值与医师标注测量值的一致性。结果:外部验证集的DSC、VS、HD分别为0.980(0.970,0.980)、0.990(0.990,1.000)、1.69(1.27,2.45)mm,各数据集之间DSC、VS和HD的差异均有统计学意义(P均<0.001)。外部验证集中模型预测和医师标注测量的胆囊体积、CT值、三维径线的95%一致性界限(LoA)的可信区间分别为(-2.07,3.36)、(-1.55,1.15)、(-1.28,1.47)、(-3.34,4.07)和(-1.11,2.15),分别有2.6%、3.7%、3.7%、1.1%和3.7%的点落在95%LoA以外。结论:基于深度学习模型可在腹部CT图像上自动分割胆囊区域,是将来进一步胆囊病变智能诊断的基础。Objective:To Explore the feasibility of segmentation and automatic measurement of the gall bladder on CT images by using deep learning algorithms.Methods:The abdominal CT data in PACS system were retrospectively collected from January 12,2016 to May 28,2021.A total of 2559 images in 1181 CT studies of 1154 patients were selected to develop deep learning models.Two radiologists labeled the gall bladder on the CT images.The data were randomly allocated to the training set(n=2043),validation set(n=245),and test set(n=271).A 3D U-net model was trained to segment and automatic measure the gall bladder.In addition,a total of 270 images in 141 CT studies of 141 patients from 10 to 19,September 2022,were collected as an external validation dataset.The predicted results of the test set and external validation set were compared with the manual measurements to evaluate the efficiency of the model.Dice similarity coefficient(DSC),volume similarity(VS),and Hausdorff distance(HD)were used to evaluate the prediction efficiency of the model.Bland-Altman test was used to analyze the agreement between the gall bladder volume,diameters,and average CT value measured by the model and the radiologists.Results:The DSC,VC,and HD in the external validation set was 0.980(0.970,0.980),0.990(0.990,1.000),and 1.69(1.27,2.45)mm,respectively.The 95% limits of agree(LoA)of volume,average CT value,and the diameters of the gall bladder was(-2.07,3.36),(-1.55,1.15),(-1.28,1.47),(-3.34,4.07),and(-1.11,2.15),respectively.For these measurement metrics,2.6%,3.7%,3.7%,1.1%,and 3.7% were outside of the 95% LoA.Conclusion:It is feasible to automatic segment the gall bladder on CT images by using deep learning algorithms,and to be used for further intelligent diagnosis of gall bladder disease in future.

关 键 词：深度学习 胆囊 体层摄影术 X线计算机 人工智能 图像分割 

分 类 号：R[医药卫生]