CT纹理分析联合机器学习对椎体隐匿性骨折辅助诊断价值初探  

Preliminary investigation:ancillary value of CT image texture analysis combined with machine learning in diagnosing occult vertebral fractures

作　　者：彭云[1,2] 王猛 刘畅 贾宏茹 吴晶涛 罗先富 PENG Yun;WANG Meng;LIU Chang;JIA Hongru;WU Jingtao;LUO Xianfu(Central South University,Changsha 410011,China;Clinical Medical School of Yangzhou University, Northern Jiangsu People's Hospital, Yangzhou 225000, China;Department of Radiology, National Cancer Centre/National Clinical Research Center for Cancer/Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Shenzhen 518116.)

机构地区：[1]中南大学湘雅二医院影像科,湖南长沙410011 [2]扬州大学临床医学院江苏省苏北人民医院影像科,江苏扬州225001 [3]国家癌症中心/国家肿瘤临床医学研究中心/中国医学科学院北京协和医院肿瘤医院深圳医院,广东深圳518116

出　　处：《暨南大学学报（自然科学与医学版）》2020年第3期268-275,共8页Journal of Jinan University(Natural Science & Medicine Edition)

基　　金：江苏省青年基金项目(BK20160450);江苏六大人才高峰项目(WSN-277);江苏省十三五科教强卫青年重点人才项目(QNRC2016321);扬州市十三五科教强卫重点人才项目(YZZDRC201816)。

摘　　要：目的:探究CT纹理分析联合机器学习多层感知模型对椎体隐匿性骨折早期检出的辅助诊断价值.方法:回顾性分析2017年1月至2018年12月84例经MRI证实为胸/腰椎隐匿性骨折病例的CT资料,每例患者选择一个隐匿性骨折椎体纳入实验组,选择一个相邻正常椎体纳入对照组.对CT图像进行标准化后,使用A.K.(artificial intelligent kit)软件分别提取椎体CT矢状位及轴位图像纹理特征.将实验组与对照组椎体按照7∶3的比例分为训练组与验证组,通过单因素方差分析、M-W秩和检验以及Spearman秩相关分析,依次对训练组椎体矢状位及轴位CT图像纹理参数进行筛选,采用多层感知模型对筛选出的参数进行建模,最后采用验证组验证其诊断效能.结果:椎体矢状位CT图像纹理参数筛选后获得最低密度值、能量、峰度、低密度增强值、Hara方差5个参数,对训练组参数进行学习建立多层感知模型后,训练组与验证组的曲线下面积(AUC)分别为1.00和0.84.椎体轴位CT图像纹理特征进行筛选后获得均值、能量、均质性、偏度、范围、区域百分比6个参数,对训练组参数进行学习建立多层感知模型后,训练组与验证组的AUC分别为1.00和0.85.结论:机器学习多层感知模型对椎体CT图像纹理特征进行筛选及建模,能够辅助诊断椎体隐匿性骨折,从而降低漏诊率.Objective:To investigate the added value of CT texture analysis combined with machine learning multilayer perception model detecting occult vertebral fractures.Methods:CT images of 84 patients with occult vertebral fractures of thoracic/lumbar spine confirmed by MRI between January 2017 and December 2018 were retrospectively analyzed.In each case,the vertebrae with occult fracture and a adjacent normal vertebrae were enrolled into experimental and control group,respectively.All the CT images were standardized.Texture features for the axial and mid-sagittal CT images of vertebrae were extracted using A.K.(artificial intelligent kit)software.The vertebral bodies of the experimental and control group were divided into the training and validation group according to the ratio of 7∶3.One-way analysis of variance、MW rank sum test and the Spearman correlation analysis were performed sequentially to select the sagittal and axial texture parameters in experimental group.Then,multilayer perception models were built based on the selected parameters.Finally,the model was tested by the verification group.Results:Five texture parameters including MinIntensity,Energy,Kurtosis,LowIntensity Emphasis,and HaraVariance were obtained after screening sagittal CTimages texture parameters.The area under the curve(AUC)of multilayer perception model of the training group and verification group were 1.00 and 0.84,respectively.Six texture parameters including Mean,Energy,Uniformity,Skewness,Range,and Zone Percentage were obtained by screening the axial CTtexture features.The AUC of training group and validation group of the modelwere 1.00 and 0.85,respectively.Conclusion:Combining machine learning with CT image texture analysis can assist in the diagnosis of occult vertebral body fractures and reduce the rate of missed diagnosis.

关 键 词：机器学习 纹理分析 隐匿性骨折 多层感知模型 

分 类 号：R445.3[医药卫生—影像医学与核医学]