摘要
目的探讨利用U-Net算法对前列腺影像报告和数据系统(PI-RADS)5类患者扩散加权成像(DWI)图像上盆腔增大淋巴结进行分割的可行性。方法回顾性分析62例PI-RADS 5类患者的DWI图像,每例患者至少有一个增大淋巴结(短径>8 mm),共84个增大淋巴结。手工标注所有增大淋巴结,并将已标注的患者按8∶1∶1的比例随机分为训练集、调优集和测试集进行模型的训练。结果测试集中U-Net模型自动分割区域平均DICE值为0.88(95%CI:0.84~0.92),对增大淋巴结检测的精确度和召回率分别为0.90和1.00。自动分割测量的淋巴结短径和体积与金标准高度相关。结论使用U-Net算法对PI-RADS 5类患者DWI图像上盆腔增大淋巴结进行分割具有一定的可行性。
Objective To explore the feasibility of the automated segmentation of enlarged lymph nodes(LNs)on diffusion-weighted images of patients with PI-RADS 5 category using the U-Net segmentation algorithm.Methods Pelvic diffusion-weighted imaging(DWI)images of 62 patientswith PI-RADS 5 category were retrospectively collected from January 2017 to June 2018.All patients had at least one enlarged LN(short diameter>8 mm),and a total of 84 enlarged LNs in these 62 patients were manually annotated for the development of U-Net segmentation algorithm.The patients were randomly divided into training dataset(50,80%),validation dataset(6,10%)and test dataset(6,10%).In the testing dataset,the DICE coefficient was used to quantitatively evaluate U-Net segmentation performance.Precision and recall value were calculated for the qualitative assessment.The short diameter and volume of lymph nodes measured by automatic segmentation are highly correlated with the gold standard.Results There were 6 patients with 9 enlarged annotated LNs in the testing dataset.The average DICE coefficient was 0.88(95%confidence interval:0.84-0.92)and theprecision and recall value were0.90 and 1.00,respectively.TheLNs’short diameter and volume measured with manual and automated segmentationshowed a high correlation(r=0.87 and 0.98,respectively).Conclusion A U-Net algorithm is feasible for LNs segmentation on DWI of patients with PI-RADS 5 category.
作者
刘想
韩超
朱丽娜
王祥鹏
张靖远
刘伟鹏
张晓东
王霄英
LIU Xiang;HAN Chao;ZHU Lina(Department of Radiology,Peking University First Hospital,Beijing 100034,P.R.China)
出处
《临床放射学杂志》
CSCD
北大核心
2020年第12期2537-2541,共5页
Journal of Clinical Radiology
关键词
盆腔淋巴结
前列腺癌
扩散加权成像
深度学习
结构化报告
Pelvic lymph nodes
Prostate cancer
Diffusion-weighted imaging
Deep learning
Structured reporting
作者简介
通讯作者:王霄英
