基于胸部CT平扫图像的胸椎纹理分析及机器学习对骨质疏松患者骨折前风险的精准预测方法研究  

Research at risk for insufficiency fractures of thoracic vertebral body on plain chest CT in osteoporotic patients using texture analysis and machine learning

作　　者：魏璐 彭旭红[1] 雷苑麟[1] 蒋方旭 赖碧玉 WEI Lu;PENG Xuhong;LEI Yuanlin;JIANG Fangxu;LAI Biyu(Department of Imaging,The Sixth Affiliated Hospital of Guangzhou Medical University,Qingyuan,Guangzhou 511518,China)

机构地区：[1]广州医科大学附属第六医院<清远市人民医院>影像科,广东清远511518

出　　处：《影像研究与医学应用》2023年第1期36-39,43,共5页Journal of Imaging Research and Medical Applications

摘　　要：目的:基于胸部CT平扫图像胸椎的纹理分析和机器学习算法,判定患者有无骨质疏松,以及能否对脆性骨折做出精准预测。方法:回顾性分析2017年1月—2021年12月清远市人民医院PACS系统确认的48例患者为病例组,共50个稳定型胸椎椎体和50个不稳定型椎体,并随机筛选50例非骨质疏松患者胸椎椎体作为对照组。病例组骨质疏松患者均行两次以上连续扫描,第一次扫描椎体正常,若二次扫描发生骨折,为不稳定型椎体,若二次扫描未发生骨折,则为稳定型椎体。对照组的年龄、性别和椎体位置与不稳定型椎体相匹配。用传统组学方法对病例组和对照组之间、稳定型椎体和不稳定型椎体之间分别行纹理分析和机器学习。结果:共纳入150个椎体。病例组和对照组之间存在显著差异,使用支持向量机、随机森林、极度随机树、LightGBM进行分类的ROC曲线分析得出的AUC值均>0.95,其中SVM最佳,为0.99[95%置信区间(CI),0.96~1.00]。稳定型椎体和不稳定型椎体之间的组学特征没有显著差异,所有机器学习模型的准确度均较低(正确率范围为0.38~0.76),其中决策树(Decision Tree)的AUC值最高,为0.73 [95%置信区间(CI),0.49~0.97]。结论:骨纹理分析和机器学习可在胸部CT平扫中精准判定胸椎椎体骨质疏松,然而,单个椎体脆性骨折风险的预测效果欠佳。Objective To evaluate whether a patient has osteoporosis or not and to identify patients with thoracic vertebrae at risk for insufficiency fractures based on texture analysis combined with machine learning algorithms in plain chest CT images.Methods A retrospective analysis was performed on 48 patients confirmed by the PACS system of Qingyuan People’s Hospital from January2017 to December 2021,with a total of 50 stable thoracic vertebral bodies and 50 unstable vertebral bodies,and 50 non osteoporotic patients with thoracic vertebrae were randomly screened as the control group.Patients with osteoporosis in the case group underwent more than two consecutive scans,the first scan of the vertebral body was normal,if the fracture occurred on the second scan,it was an unstable vertebral body,and if the second scan did not cause a fracture,it was a stable vertebral body.The age,sex,and vertebral body position of the control group matched the unstable vertebral body.Conventional radiomics methods were used to perform texture analysis and machine learning between the case group and the control group,and between stable vertebral bodies and unstable vertebral bodies.Results 150 vertebrae were included.There were significant differences between cases and controls,and ROC curve analysis using support vector machine,random forest,extreme random trees,and light GBM for classification yielded AUC values> 0.95,with SVM being the best,at 0.99 [95% confidence interval(CI),0.96-1.00].There were no significant differences between the omics profiles of stable and unstable vertebral bodies,and all machine learning models had low accuracy(range 0.38~0.76),with the highest AUC value of 0.73 [95% confidence interval(CI),0.49~0.97] for decision tree.Conclusion Bone texture analysis and machine learning can precisely identify thoracic vertebral osteoporosis in chest CT plain scan.However,identification of single vertebra at risk remains challenging.

关 键 词：脊柱 传统影像组学 机器学习 骨质疏松症 CT 

分 类 号：R445.3[医药卫生—影像医学与核医学]