用深度学习和阈值算法自动检出CT平扫图像中肾结石的可行性研究  

Automatic detection of kidney stones on plain CT images:a feasibility study with deep learning and thresholding methods

作　　者：崔应谱 孙兆男 刘想 韩超 张晓东[1] 王霄英[1] Cui Yingpu;Sun Zhaonan;Liu Xiang;Han Chao;Zhang Xiaodong;Wang Xiaoying(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,100034

出　　处：《中华放射学杂志》2020年第9期869-873,共5页Chinese Journal of Radiology

摘　　要：目的探索使用级联模型在CT平扫图像中自动检出肾结石的可行性。方法回顾性搜集2018年1月至7月在北京大学第一医院行泌尿系CT平扫并诊断为肾结石的59例患者的临床和影像资料。根据患者CT检查时间将患者图像分为两组:肾脏分割模型训练集(30例)和肾结石检出模型测试集(29例)。建模包括:(1)训练肾脏分割的深度学习模型,采用U-Net神经网络,以dice系数评估肾脏分割算法的性能;(2)在肾脏分割的基础上,用阈值算法和区域生长算法检出所得肾脏区域里的结石;(3)测量肾结石各向径线值(长径、中径和短径)和CT值,并将结果返回到结构化报告。由1名医师记录肾结石位置并手动测量肾结石各向径线值(长径、中径和短径)和CT值作为金标准,计算模型自动检出结石的灵敏度、特异度和准确度,并使用Bland-Altman法分析模型自动测量和医师手工测量结果的一致性。结果测试集29例患者,共11358层CT图像,58个肾脏纳入研究,其中38个肾脏有结石,共有56个结石,20个肾脏无结石。使用U-Net模型对测试集11358层CT图像进行分割,平均dice系数为0.96;其中10945(96.36%)层图像分割效果很好,dice系数≥0.90。以肾脏为单位,模型在测试集中检出结石的灵敏度、准确度、特异度分别为100%(38/38)、100%(38/38)、100%(20/20);以结石为单位,模型检出结石的灵敏度、准确度分别为100%(56/56)、96.6%(56/58)。结论肾结石检出级联算法可以在CT平扫图像中自动检出肾结石,并自动返回到结构化报告中,提高临床工作效率。Objective To develop and validate a cascaded deep learning algorithm for kidney stone detection on plain CT images.Methods Plain CT images of the patients with kidney stones were retrospectively archived from January 2018 to July 2018 in Peking University First Hospital.The cases were divided into two datasets according to the date of the CT scanning:training dataset(n=30)and held-out test dataset(n=29).The development of the kidney stone detection method consisted of three steps.First,a U-Net model was trained on the training dataset for kidney segmentation,and the model′s performance was estimated with the dice coefficient.Second,the thresholding and region growing methods were used to detect the stones in the renal region predicted by the trained U-Net model.Third,the stones′lengths(maximal,middle and minimal length)and CT attenuation were calculated and integrated into a structured report automatically.Using the radiologist′s labels and measurements(maximal,middle,minimal length and CT attenuation)as ground truth,the stone detection algorithm performance was evaluated with sensitivity,specificity and precision,and the stone measurement algorithm performance was evaluated with Bland-Altman plots.Results The held-out test dataset consisted of 29 cases,containing 58 kidneys and 11358 CT slices.The 38 kidneys containing 56 stones and 20 kidneys did not contain stones.The U-Net model showed good performance,with a mean dice coefficient of 0.96.And 10945 of 11358 CT slices had a dice coefficient no less than 0.90.The sensitivity,precision,and specificity of stone detection were 100%(38/38),100%(38/38)and 100%(20/20)in the organ-level.The sensitivity and precision of stone detection were 100%(56/56)and 96.6%(56/58)in the lesion-level.Conclusion A cascaded algorithm is constructed and can be used to detect kidney stones in plain CT images.The algorithm can improve efficiency with results automatically integrated into the structured report in clinical practice.

关 键 词：肾结石 自动化 体层摄影术 X线计算机 

分 类 号：R692.4[医药卫生—泌尿科学] R816.7[医药卫生—外科学]