基于深度学习探索3D MRU尿路分割的初步研究  

Exploring 3D MRU urinary tract segmentation based on deep learning

作　　者：奈日乐 林子楹 额·图娅 吴鹏升 张耀峰 张晓东[1] 王霄英[1] NAI Ri-le;LIN Zi-ying;E Tu-ya(Department of Radiology,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]北京赛迈特锐医学科技有限公司,北京100011

出　　处：《放射学实践》2022年第7期865-869,共5页Radiologic Practice

摘　　要：目的:探索基于深度学习在三维磁共振尿路成像(3D MRU)图像上分割尿路的可行性。方法:回顾性收集2021年1月1日-2021年4月30日本院包含“MRU”检查项目的图像,共219例数据纳入本研究。由2名影像医生手工勾画双侧肾盂-肾盏、输尿管及膀胱区域,将219例数据随机分为训练集(175例)、调优集(22例)和测试集(22例)训练3D U-net分割模型。统计Dice相似性系数(DSC)和霍夫曼距离(HD)用于自动分割的客观评价,由2名影像医生主观评价模型自动分割勾画并应用组内相关系数(ICC)评估主观评分的一致性。结果:MRU分割模型的测试集共22个数据,DSC值均达到0.70及以上,右侧输尿管、左侧输尿管、左侧肾盂-肾盏、右侧肾盂-肾盏及膀胱的分割结果DSC值分别为0.81、0.70、0.85、0.95、0.98,HD值分别为(43.01±41.24)mm、(65.1±66.80)mm、(37.8±52.48)mm、(52.08±69.88)mm、(10.06±20.76)mm。2位影像医生对测试集的主观评分总分为29.00±1.68、28.68±1.63,ICC值为0.95(95%CI:0.89~0.98)。结论:基于深度学习的3D MRU尿路自动分割勾画在临床具备可行性,可为后续MRU的定位、定量及定性诊断提供基础。Objective:To explore the feasibility of segmenting the urinary tract on 3D MRU images based on deep learning.Methods:The MRU images were retrospectively collected in our hospital,from January 1,2021,to April 30,2021.A total of 219 cases of data were included in this study.Two radiologists manually delineated the bilateral renal pelvis-renal calyx,ureter,and bladder regions,and randomly divided 219 cases of data into training set(175 cases),validation set(22 cases),and test set(22 cases).A 3D U-net segmentation model was trained.The Dice similarity coefficient(DSC)and Hausdorff distance(HD)were used for the objective evaluation of automatic segmentation.Two radiologists subjectively evaluated the result of automatic segmentation and compared their consistency.Results:There were 22 data in the test set of the MRU segmentation model.The DSC for the right ureter,left ureter,left renal pelvis-renal calyx,right renal pelvis-renal calyx,and bladder were 0.81,0.70,0.85,0.95,and 0.98,respectively.The HD values were(43.01±41.24)mm,(65.1±66.80)mm,(37.8±52.48)mm,(52.08±69.88)mm,and(10.06±20.76)mm,respectively.The subjective score of 2 radiologists on test sets were 29.00±1.68 and 28.68±1.63.The ICC value was 0.95(95%CI:0.89~0.98).Conclusion:The automatic segmentation of 3D MRU urinary tract based on deep learning is clinically feasible and can provide a basis for the localization,quantitative and qualitative diagnosis of MRU in the future.

关 键 词：磁共振尿路成像 尿路 分割 深度学习 

分 类 号：R445.2[医药卫生—影像医学与核医学] R692[医药卫生—诊断学]