基于低剂量胸部CT深度学习模型自动测量骨密度研究  

Bone densitometry measurement based on low-dose chest CT with deep learning model

作　　者：赵宇 张晓岚 郑超 王敏红 洪薇 周运锋 ZHAO Yu;ZHANG Xiao-lan;ZHENG Chao(Department of Radiology,the First Affiliated Hospital of Wannan Medical College,Wuhu 241001,China)

机构地区：[1]芜湖皖南医学院第一附属医院/弋矶山医院放射科,安徽241001 [2]数坤(北京)网络科技股份有限公司,北京100102

出　　处：《放射学实践》2024年第2期262-266,共5页Radiologic Practice

基　　金：科研能力“高峰”骨干KGF2019G13 XM_LHJY2022_05_13。

摘　　要：目的:评价基于低剂量胸部CT深度学习模型全自动测量与定量CT(QCT)手动测量骨密度的一致性。方法:回顾性分析2018年6月-2019年12月在弋矶山医院行低剂量胸部CT筛查联合定量CT(QCT)骨密度测量的1406例体检者的临床和影像资料。随机分为训练集(985例)和测试集(421例)。应用深度学习等方法构建骨分割模型等及内部组织校准模型,应用测试集检测其效能。以QCT结果作为参考标准,应用Spearman相关系数、组内相关系数、Bland-Altman分析两种测量方法的相关性和一致性;以QCT测量结果的骨质疏松(OP)诊断作为参考标准,应用ROC曲线评价其对OP的诊断效能。结果:训练集及测试集中深度学习模型与QCT的骨密度测量结果均呈正相关(训练集:r=0.957,P<0.001;测试集:r=0.955,P<0.001),组内相关系数为0.946(训练集)、0.945(测试集)。该模型在训练集中ROC曲线下面积(AUC)、灵敏度、特异度、准确度分别为0.986、47.5%、100%、95.7%;在测试集中分别为0.975、42.1%、100%、94.8%。结论:基于低剂量胸部CT深度学习模型和QCT的骨密度测量的一致性和相关性较好,初步实现了自动、快速的基于LDCT的骨质疏松筛查,但需扩大患者人群进一步优化和验证。Objective:To evaluate the consistency of bone mineral density(BMD)measurement between the low-dose chest CT based deep learning model and QCT.Methods:The present retrospective study was done and analyzed using the clinical and imaging data of 1406 patients who had low-dose chest CT screening and quantitative computed tomography(QCT)at Yijishan Hospital between June 28,2018 and December 31,2019.The patients were divided into a test set(985 cases)and a training set(421 cases)at random.Bone segmentation models and internal tissue calibration models were created using a deep neural network,and its efficiency was evaluated using the test set.The correlation and consistency of the two methods were examined using the QCT results as the reference standard and the Spearman correlation coefficient,interclass correlation coefficient,and Bland-Altman.Results:In both the training set and the test set,the deep learning model was positively correlated with QCT results in the training set(r=0.957,P<0.001)and test set(r=0.955,P<0.001).The intra-group correlation coefficient was 0.946 in the training set and 0.945 in the test set.In the training set,the AUC of this model was 0.986,with the sensitivity,specificity,and accuracy of 47.5%,100%,and 95.7%,respectively.In the test set,the AUC of this model was 0.975,with the sensitivity,specificity,and accuracy of 42.1%,100%,and 94.8%,respectively.Conclusion:The low-dose chest CT-based deep learning model's estimation of bone mineral density shows good consistency and correlation with the outcome determined by QCT.We have so far achieved automatic and quick osteoporosis screening using LDCT.For the further optimization and validation,we will increase the patient population.

关 键 词：人工智能 骨密度 体层摄影术 X线计算机 骨质疏松 

分 类 号：R-05[医药卫生] R336R681R814.4