CT影像组学结合支持向量机...体压缩性骨折的鉴别诊断价值_李邦凤.pdf

149CHINESE JOURNAL OF CT AND MRI,FEB.2023,Vol.21,No.02 Total No.160【第一作者】李邦凤，女，本科生，主要研究方向：影像组学。E-mail：【通讯作者】彭云，男，医师，主要研究方向：双能CT应用、影像组学。E-mail：论 著The Value of the CT Radiomics Combined with Support Vector Machines in Different Between Incidental Acute and Old vertebral Compression Fractures*LI Bang-feng1,2,FU Yu-ping1,2,GONG Liang-geng1,PENG Yun1,*,LIN Hua-shan3.1.Department of Radiology,The Second Affiliated Hospital of Nanchang University,Nanchang 330006,Jiangxi Province,China2.The Second Clinical College of Medicine,Nanchang University,Nanchang 330006,Jiangxi Province,China3.GE Pharmaceutical GE Healthcare,Changsha 410000,Hunan Province,ChinaABSTRACTObjective To investigate the value of vertebral texture analysis based on thoracic CT images combined with support vector machine(SVM)machine learning method in identifying acute and old vertebral compression fractures.Methods The data of 132 patients with incidental vertebral compression fractures detected on routine chest CT and confirmed by MRI from May 2018 to May 2021 were retrospectively analyzed.163 vertebrae including 98 acute fractures and 65 old were included.The Mazda software was used to extract texture features of each vertebra in axial and sagittal orientation.Then the IPMS software was used to further dimensionality reduction and model building.The old and acute fractured vertebrae were randomly divided into the training and validation samples according to the ratio of 7:3.For training sample,the T test,Wilcoxon rank sum test and Pearson correlation analysis were performed to screen the texture features in both orientations.SVM models were built based on the selected axial and sagittal parameters respectively.Then the diagnostic value was tested by the validation sample and the receiver operating characteristic curves(ROC)were obtained.Results For each vertebra,294 features were extracted from sagittal and axial imaging respectively.8 parameters were finally obtained for the sagittal orientation,the AUC of the SVM model was 0.78 and 0.68 for the training and validation sample,respectively;7 parameters were finally obtained for the axial orientation,the AUC of the SVM model was 0.80 and 0.84 for the training and validation sample,respectively.with higher model efficacy for the axial position than for the sagittal position.The axial model shown higher diagnostic value than sagittal model.Conclusion Radiomic based on chest CT combined with SVM method can differentiate acute from old vertebral compression fractures,the axial model shown better performance and may provide an auxiliary diagnosis for incidental acute vertebral fractures without a clear history of trauma thus facilitating early treatment.Keywords:Radiomics;Vertebral Compression Fracture;Support Vector Machine;Texture Analysis椎体压缩性骨折作为临床常见疾病，可引起患者生活质量下降，与病死率增高紧密相关1，该疾病按骨折时期可分为急性及陈旧性骨折，其中急性骨折具有骨质中断、骨髓水肿的病理表现，骨质结构不稳定，需早期识别并及时治疗以防止继发性骨折，引起二次损伤2。椎体压缩性骨折的大部分病例可经CT多平面重建技术明确诊断，结合外伤史，亦可对急性骨折进行诊断，但对于部分无或仅有轻微外伤史的病例，常规CT难以区别急性及陈旧性骨折。MRI的脂肪抑制序列对骨髓水肿具较高敏感性，通常被认为急性骨折的诊断标准3，但该技术具有采集时间长、费用高及禁忌症多等缺陷。双能CT可通过虚拟去钙技术识别急性压缩性骨折的骨髓水肿从而辅助诊断急性骨折，但对机器要求较高，难以普及推广4。影像组学能够高通量提取代表图像特征的多种定量纹理特征，提供肉眼无法识别的多维度信息，进而评估病变组织与正常组织的同、异质性，结合机器学习方法，即可建立病变判别模型5。研究证实基于CT图像的影像组学特征不仅可反映神经性厌食症患者椎体的骨密度，还可对椎体完整性进行评估6。本研究假设基于CT图像组学特征能够反映骨折椎体的形态变化及密度分布，可能对急性及陈旧性压缩性骨折进行鉴别。鉴于椎体CT成像通常不作为无明确外伤病史患者的首选检查方式，因此本文旨在探讨基于胸部CT图像影像组学支持向量机(support vector machine，SVM)模型在鉴别偶发性急性与陈旧性压缩性骨折中的价值，探究快速、准确判断椎体急性骨折的方法。1 资料与方法1.1 一般资料 回顾性分析南昌大学第二附属医院2018年5月至2021年5月行常规胸部CT检查并于48小时内行胸椎MRI检查证实急性或陈旧性压缩性骨折患者，急性压缩性骨折椎体纳入急性组，陈旧性压缩性骨折椎体纳入陈旧组。排除标准：具有恶性肿瘤病史可能导致骨转移或影像学检查提示骨转移；检查前接受过手术治疗(内固定/骨水泥)，影响观察的椎体；引起椎体形态变化病变(包括血管瘤、许莫氏结节)的椎体。1.2 检查方法 胸部CT：CT图像均来自于两台机器(ICT256及Brilliance16，均为飞利浦)。扫描范围自胸廓入口至肺底，管电压120kVP，自动管电流，矩阵均为512*512，CT影像组学结合支持向量机对偶发急性及陈旧性椎体压缩性骨折的鉴别诊断价值*李邦凤1,2 付玉苹1,2 龚良庚1彭 云1,*林华山31.南昌大学第二附属医院影像中心 (江西 南昌 330006)2.南昌大学第二临床医学院 (江西 南昌 330006)3.通用电气药业通用电气医疗 (湖南 长沙 410000)【摘要】目的 探究基于胸部CT图像椎体纹理分析结合支持向量机(SVM)机器学习方法在鉴别急性及陈旧性椎体压缩性骨折的价值。方法 回顾性分析132例2018年5月至2021年5月行常规胸部CT并经MRI证实为椎体压缩性骨折患者的资料，纳入急性骨折椎体98个、陈旧性骨折椎体65个，共163个椎体。采用Mazda软件提取所有椎体轴位与矢状位纹理特征，采用IPMS软件进一步筛选、降维及建模，将新旧骨折椎体按7：3的比例随机分入训练集与验证集。通过T检验、Wilcoxon秩和检验及Pearson相关分析对训练集的纹理特征进行筛选，依次建立轴位及矢状位筛选特征的SVM模型，进一步用验证集验证模型有效性，获得受试者工作特征(ROC)曲线。结果 每个椎体矢状位及轴位分别获得294个特征。矢状位最终获得8个参数，训练集及验证集SVM模型的ROC曲线下面积(AUC)分别为0.78和0.68；轴位最终获得6个参数，训练集及验证集SVM模型的AUC分别为0.80和0.84。轴位模型效能较矢状位更高。结论 胸部CT影像组学特征结合SVM可对急性及陈旧性椎体压缩性骨折进行鉴别，以轴位模型为佳，能够为无明确外伤史的偶发急性椎体骨折提供辅助诊断从而促进早期治疗。【关键词】影像组学 椎体压缩性骨折 支持向量机 纹理分析【中图分类号】R445.3【文献标识码】A【基金项目】国家自然科学基金 (81860316)基于影像学和分子生物学 探究miR-318通过激活Hippo信号通路 调控心肌梗死的机制 DOI:10.3969/j.issn.1672-5131.2023.02.050150中国CT和MRI杂志2023年02月 第21卷 第02期 总第160期每位患者扫描后均重建获得层间距1mm的薄层图像。胸椎MRI：MR扫描均在GE公司1.5T MRI机器(SIGNATM Airtist)进行，采用脊柱专用线圈，扫描序列包括矢状位T1WI、T2WI以及短时反转恢复(short time inversion recovery STIR)序列，以及轴位T2加权序列。图像解读：结合病史，CT及MR图像显示椎体高度塌陷或具有骨折线的认定为骨折椎体，其中MR图像椎体内含T1WI低信号，T2及STIR高信号的区域为骨髓水肿，诊断为急性压缩性骨折，否则诊断为陈旧性骨折。1.3 组学特征提取 将胸部CT轴位薄层图像导入工作站进行多平面重建，获得每位患者的椎体正中矢状位图像，将轴位及矢状位图像以DICOM格式导出，转换为BMP格式。由1名具有3年诊断经验的放射科医师采用Mazda软件对所有胸部CT图像纳入的骨折椎体感兴趣区(ROI)进行勾画。ROI勾画时，急性骨折椎体轴位层面选取参考MRI图像中含骨髓水肿层面选定，陈旧性骨折椎体选取椎体正中轴位层面，所有椎体矢状位层面均选取椎体正中矢状位层面进行纹理特征进行提取。ROI勾画时沿骨皮质内缘2mm勾画整个病变椎体，避开血管、椎体附件等组织7。最终每个椎体获得直方图特征、灰度共生矩阵特征、游程长矩阵特征、梯度特征、自回归模型特征以及小波转换系数6类共294个特征。1.4 统计学分析 采用SPSS19.0进行统计分析，采用卡方检验及独立样本t检验对两组间临床资料(性别、年龄)进行对比。采用GE公司IPMS软件内置的Python及R语言统计算法，对椎体CT轴位和矢状位图像组学特征依次筛选及建模。步骤如下：(1)对数据进行标准化(Z-Score)；(2)将急性组与陈旧组按7:3的比例随机分为训练集及验证集；(3)在验证集内采用独立样本t检验及Wilcoxon秩和检验筛选两组间有差异的特征，进一步行pearson相关分析去除相关性强(r0.7)的冗余参数，采用剩余特征构建SVM模型；(4)用验证集数据对模型效能进行验证。由此分别获得轴位及矢状位训练集及验证集的受试者工作特征(ROC)曲线，以ROC曲线下面积(AUC)、准确率、敏感度及特异度评估模型效能。2 结 果2.1 一般资料 最终共纳入132例患者的163个椎体，其中男性63例(43.73%)，女性69例(52.27%)，年龄64.3012.54岁；急性组98个椎体，陈旧组65个椎体。所有患者中，106例纳入1个骨折椎体，18例纳入2个椎体，6例纳入3个椎体。163个椎体内，T1、T2、T3及T4各 1个，T5 3个，T6 4个，T7、T8各 6个，T9及T10 5个，T11 24 个，T12 62个，L1 44个。两组患者的临床特征结果见表1。图1 图1A：轴位SVM模型验证集鉴别急性及陈旧性压缩性骨折椎体的ROC曲线，图1B：矢状位SVM模型验证集鉴别急性及陈旧性压缩性骨折椎体的ROC曲线。2.2 组学特征筛选结果每个椎体分别获得294个轴位纹理特征及294个矢状位纹理特征。矢状位特征：t检验与Wilcoxon秩和检验筛选出161个急性组与陈旧组间存在差异的特征，进一步经pearson相关分析筛选最终确定S(0,3)Correlat、S(0,3)DifVarnc、S(4,4)AngScMom、S(4,-4)SumVarnc、S(5,0)SumVarnc、S(5,5)SumAverg、S(5,-5)Correlat及 Vertl_RLNonUni 8个具代表性意义参数，前7个均为灰度共生矩阵参数，最后一个为游程长矩阵参数。轴位特征：t检验与Wilcoxon秩和检验筛选出108个两组间存在差异的特征，进一步经person相关分析筛选最终确定Skewness、Kurtosis、Perc.10%、S(3,0)InvDfMom、GrKurtosis、Teta2及Teta3共7个代表性意义参数，其中前3者为直方图特征，S(3,0)InvDfMom为灰度共生矩阵参数，GrKurtosis为梯度特征，最后2个特征为自回归模型特征。2.3 模型训练结果 163个椎体按7:3的比例随机分入训练集及验证集，最后训练集含69个急性骨折椎体、45个陈旧性骨折椎体，验证集含29个急性骨折椎体，20个陈旧性骨折椎体。以2.2中筛选出的矢状位特征建立SVM模型，训练集AUC为0.78，验证集AUC为0.68(准确度0.71，敏感度0.76，特异度0.65)；3 讨 论在无或仅有轻微外伤史患者的查因过程中，偶发性椎体骨折在胸部CT图像中具有较高的报告率，准确、快速识别其中的急性骨折对早期治疗很有必要8。本研究基于胸部CT图像对骨折椎体轴位及矢状位组学特征进行提取，分别筛选获得7个及8个参数构建模型，发现轴位与矢状位模型都能够鉴别急性及陈旧性骨折椎体，对比两种模型效能，证明轴位模型诊断效能高于矢状位模型。影像组学概念最早提出于2012年，其是指从CT、MRI等影像中高通量地提取特征，采用多种自动算法将ROI内影像数据转化为代表图像深层次特征的客观量化数据，可为疾病的诊断、评估及预后提供精准信息，目前已逐渐应用于骨肌系统中9。Frighetto等10通过提取腰椎压缩性骨折椎体MRI图像的形状及纹理特征，发现筛选后建立的K近邻及Bayes模型可对良性及病理性骨折进行鉴别。Kawashima等11证实基于颅脑CT图像的纹理特征可对骨质密度进行评估，从而区分骨密度正常者及骨质疏松患者。Muehlematter等12进一步发现基于影像组学的机器学习方法可识别椎体不全性骨折的潜在患者。本研究提取了新旧压缩性骨折椎体的直方图、灰度共生矩阵、游程长矩阵、梯度特征、自回归模型以及小波转换系数6类特征，最终筛选获得的特征中，矢状位以灰度共生矩阵参数最多，而轴位以直方图特征为多，由此建立的模型均可对急性及陈旧性骨折进行鉴别。在无明确外伤史的背景下，急性及陈旧性椎体骨折在CT上可出现高度塌陷及骨小梁排列的表现，仅凭肉眼难以判定；但急性骨折具有骨小梁中断及微出血(骨髓水肿)的病理特征，可导致松质骨内小梁排列及密度分布改变，陈旧性骨折椎体具有骨质修复的特征，这种病理上的差异，可通过影像组学深度挖掘从而鉴别。目前，影像组学深度学习建模已有诸如SVM、逻辑回归、K近邻、bayes等多种算法，本研究采用了SVM方法进行模型建立。SVM建立在VC维理论和结构风险最小化原理上，属于有监督机器学习方法，在解决高维数、小样本、非线性等模式识别问题中具有特定优势，能够有效解决有限数量样本的高维模型构造问题，与其他常用分类器相比具有较好分类性能13。有研究发现基于增强CT的SVM模型可对食管癌的分化程度进行预测14。本研究同样采用SVM深度学习进行建模，所得轴位及矢状位特征均在一定程度上区分新旧椎体骨折。此外，本研究所建轴位模型比矢状位SVM模型在鉴别新旧压缩性骨折方面具有更高效能，其原因可能是本研究在ROI勾画的轴位层面选择参考MRI图像，选取了急性骨折椎体具骨髓水肿的层面，而矢状位层面均选在正中层面，部分椎体的骨髓水肿并不累及整个椎体，在正中矢状位上仅累及部分区域，导致轴位ROI中具有骨髓水肿、新近骨质中断的比例较矢状位更高，即轴位特征可能在显示急性骨折及陈旧性骨折差异中更有优势。本研究存在一定的局限性：(1)由于胸腰交界区为应力区域，较其他区域椎体骨折几率更高，本研究中T11、T12及L1椎体数量占比高达79.75%，结果可能主要反映胸腰交界区骨折的情况；(2)急性组及陈旧组内各个椎体占比不完全一致，可能影响结果。综上所述，本研究通过提取胸部CT椎体组学特征，结合SVM机器学习模型，初步证实影像组学在椎体急性压缩性骨折中的诊断效能，且该效能以轴位模型为佳，期望扩大样本进一步研究，建立基于CT图像的辅助诊断系统，为放射医生快速判断无明确外伤史的偶发性急性骨折提供高效手段，为临床治疗决策提供帮助。以2.2中筛选出的轴位特征建立SVM模型，训练集AUC为0.80，验证集AUC为0.84(准确度0.76，敏感度0.83，特异度0.65)。1A1B（下转第 174 页）表1 急性组及陈旧组间临床资料对比组别 年龄(岁)性别 男 女急性组 62.8512.59 33 42陈旧组 66.2112.33 30 27t/2 1.53 0.97P值 0.13 0.38174中国CT和MRI杂志2023年02月 第21卷 第02期 总第160期（收稿日期：2022-10-25）（校对编辑：姚丽娜）pictures,they are dataJ.Radiology,2016,278(2):563-577.9J.N.Diaconis,K.C.Rao.CT in head trauma:a reviewJ.The Journal of Computed TomographyJ.1980,4(4):261-270.10L.L.Wald,Ultimate MRI.Journal of Magnetic Resonance(San Diego,Calif.:1997),2019,306:139-144.11M.T.Parisi,M.S.Bermo,A.M.Alessio,et al.Optimization of pediatric PET/CT J.Seminars in Nuclear Medicine,2017,47(3):258-274.12K.O.Almansory,F.Fraioli.Combined PET/MRI in brain glioma imagingJ.British Journal of Hospital Medicine(London,England:2005),2019,80(7):380-386.13S.Mikes,M.Haindl.Texture Segmentation Benchmark,IEEE transactions on pattern analysis and machine intelligence Pp(2021).14C.Magliaro,A.L.Callara,N.Vanello,et al.A manual segmentation tool for three-dimensional neuron datasetsJ.Frontiers in Neuroinformatics,2017,11:36.15Y.Wu,W.Yang,J.Jiang,et al.Semi-automatic segmentation of brain tumors using population and individual informationJ.Journal of Digital Imaging,2013,26(4):786-796.16M.Deng,J.Zhenhao,R.Yu,et al.The learning-based automatic segmentation algorithm of brain MR Images based on 7TJ.Current Medical Imaging,2021,17(3):342-351.17K.Gau,C.S.M.Schmidt,H.Urbach,et al.Schulze-Bonhage,C.P.Kaller,N.A.Foit,accuracy and practical aspects of semi-and fully automatic segmentation methods for resected brain areasJ.Neuroradiology,2020,62(12):1637-1648.18L.R.Mascarenhas,A.D.S.Ribeiro Jnior,R.P.Ramos.Automatic segmentation of brain tumors in magnetic resonance imaging,Einstein(Sao Paulo,Brazil),2020,18:eAO4948.19M.Patyk,J.Silicki,R.Mazur,et al.Radiomics-the value of the numbers in present and future radiologyJ.Polish Journal of Radiology,2018,83:e171-e174.20V.Parekh,M.A.Jacobs.Radiomics:a new application from established techniquesJ.Expert Review of Precision Medicine and Drug Development,2016:1(2):207-226.21L.Nanni,A.Lumini,N.Zaffonato.Ensemble based on static classifier selection for automated diagnosis of Mild Cognitive ImpairmentJ.Journal of Neuroscience Methods,2018,302:42-46.22Q.Feng,Y.Chen,Z.Liao,et al.Corpus callosum radiomics-based classification model in alzheimers disease:A case-control studyJ.Frontiers in neurology,2018,9:618.23A.Lauric,C.G.Ludwig,A.M.Malek.Enhanced radiomics for prediction of rupture status in cerebral aneurysmsJ.World Neurosurgery,2022,159:e8-e22.24S.K.Feske.Ischemic StrokeJ.The American Journal of Medicine,2021,134(12):1457-1464.25M.Olive-Gadea,C.Crespo,C.Granes,et al.Deep learning based software to identify large vessel occlusion on noncontrast computed tomographyJ.Stroke,2020,51(10):3133-3137.26T.Y.Tang,Y.Jiao,Y.Cui,et al.Penumbra-based radiomics signature as prognostic biomarkers for thrombolysis of acute ischemic stroke patients:a multicenter cohort studyJ.Journal of Neurology,2020,267(5):1454-1463.27W Qiu,H Kuang,J Nair,et al.Radiomics-based intracranial thrombus features on CT and CTA predict recanalization with intravenous alteplase in patients with acute ischemic strokeJ.Am J Neuroradiol,2019,40(1):39-44.28耿立娜.基于DWI的影像组学模型评估急性脑梗死预后的初步研究D.河北医科大学,2019.29C.Zhan,Q.Chen,M.Zhang,et al.Radiomics for intracerebral hemorrhage:are all small hematomas benign?J.The British Journal of Radiology,2021,94(1119):20201047.30H.Li,Y.Xie,H.Liu,et al.Non-contrast CT-Based Radiomics Score for Predicting Hematoma Enlargement in Spontaneous Intracerebral Hemorrhage,Clinical neuroradiology,2021.31Y.Zeng,X.Liu,N.Xiao,et al.Automatic Diagnosis Based on Spatial Information Fusion Feature for Intracranial AneurysmJ.IEEE Transactions on Medical Imaging,2020,39(5):1448-1458.32A.R.Podgorsak,R.A.Rava,M.M.Shiraz Bhurwani,et al.Automatic radiomic feature extraction using deep learning for angiographic parametric imaging of intracranial aneurysmsJ.Journal of Neurointerventional Surgery,2020,12(4):417-421.33O.Alwalid,X.Long,M.Xie,et al.CT Angiography-Based Radiomics for Classification of Intracranial Aneurysm RuptureJ.Frontiers in Neurology,2021,12:619864.34Y.Jiao,J.Z.Zhang,Q.Zhao,et al.Machine Learning-Enabled Determination of Diffuseness of Brain Arteriovenous Malformations from Magnetic Resonance Angiography,Translational stroke research,2021.35P.H.Kuo,C.C.Lee,C.F.Lu.Radiomics-based Prediction of Re-hemorrhage in Cerebral Cavernous Malformation after Gamma Knife Radiosurgery,Annual International Conference of the IEEE Engineering in Medicine and Biology Society.IEEE Engineering in Medicine and Biology Society.Annual International Conference,2021,2021:3668-3671.36Y.Liu,D.Dong,L.Zhang,et al.Radiomics in multiple sclerosis and neuromyelitis optica spectrum disorderJ.European Radiology,2019,29(9):4670-4677.37Y.Zhao,B.C.Healy,D.Rotstein,et al.Exploration of Machine Learning Techniques in Predicting Multiple Sclerosis Disease CourseJ.PloS One,2017,12(4):e0174866.38L.Tang,X.Wu,H.Liu,et al.Individualized prediction of early alzheimers disease based on magnetic resonance imaging radiomics,clinical,and laboratory examinations:a 60-month follow-up studyJ.Journal of Magnetic Resonance Imaging:JMRI,2021,54(5):1647-1657.39T.Shiiba,K.Takano,A.Takaki,et al.Dopamine transporter single-photon emission computed tomography-derived radiomics signature for detecting Parkinsons diseaseJ.EJNMMI Research,2022,12(1):39.40D.Sun,X.Wu,Y.Xia,et al.Differentiating Parkinsons disease motor subtypes:A radiomics analysis based on deep gray nuclear lesion and white matter,Neuroscience letters 760,2021:136083.41Y.Zhang,P.Yan,F.Liang,et al.Predictors of Epilepsy Presentation in Unruptured Brain Arteriovenous Malformations:A Quantitative Evaluation of Location and Radiomics Features on T2-Weighted ImagingJ.World neurosurgery,2019,125:e1008-e1015.42E.N.Cheong,J.E.Park,D.E.Jung,et al.Extrahippocampal Radiomics Analysis Can Potentially Identify Laterality in Patients With MRI-Negative Temporal Lobe EpilepsyJ.Frontiers in Neurology 12,2021:706576.（收稿日期：2021-10-25）（校对编辑：姚丽娜）参考文献1王文广,田林涛,李玉其.CT引导下PVP治疗骨质疏松性椎体骨折疗效分析J.中国CT和MRI杂志,2021,06:156-159.2Mao H,Zou J,Geng D,et al.Osteoporotic vertebral fractures without compression:key factors of diagnosis and initial outcome of treatment with cement augmentationJ.Neuroradiology,2012,54(10):1137-1143.3颜路悠,张堃,钟泽亚,等.双能CT虚拟去钙技术不同重建算法鉴别诊断急慢性椎体压缩骨折J.临床放射学杂志,2021,40(6):1176-1180.4Petritsch B,Kosmala A,Weng A M,et al.Vertebral Compression Fractures:Third-Generation Dual-Energy CT for Detection of Bone Marrow Edema at Visual and Quantitative AnalysesJ.Radiology,2017,284(1):161-168.5Yin P,Mao N,Wang S,et al.Clinical-radiomics nomograms for pre-operative differentiation of sacral chordoma and sacral giant cell tumor based on 3D computed tomography and multiparametric magnetic resonance imagingJ.Br J Radiol,2019,92(1101):20190155.6Tabari A,Torriani M,Miller K K,et al.Anorexia Nervosa:Analysis of Trabecular Texture with CTJ.Radiology,2017,283(1):178-185.7Mannil M,Eberhard M,Becker A S,et al.Normative values for CT-based texture analysis of vertebral bodies in dual X-ray absorptiometry-confirmed,normall