影像组学模型对颈动脉斑块内微钙化的诊断价值研究

Diagnostic values of radiomics models in micro-calcifications in carotid plaques

目的构建诊断颈动脉斑块内微钙化的影像组学模型并分析其诊断价值。方法纳入苏州大学附属第三医院神经内科自2017年5月至2019年11月收治的52例大动脉粥样硬化型脑梗死患者进入研究。所有患者均行颈动脉常规超声检查明确存在颈动脉斑块,同时采用Micropure■超声技术检测斑块内微钙化情况。若颈动脉斑块内存在微钙化,取微钙化密度最高的斑块横断面;若无微钙化则取斑块最大横断面。所有图像经Photoshop软件进行归一化处理后,通过MaZda 4.6软件勾画斑块为感兴趣区,自动提取斑块283个纹理特征。通过不同观察者间观察一致性分析[组内相关系数(ICC)>0.75)、两样本t检验、Lasso回归筛选出诊断效果最强的纹理特征,并通过随机森林(RF)和支持向量机(SVM)分类器构建模型,以7∶3划分训练集和验证集,通过受试者工作特征曲线(ROC)计算曲线下面积(AUC)评估模型的诊断效果,并采用Delong检验比较两种模型对验证集诊断能力的差异。结果共纳入52例患者148个斑块图像,其中斑块内有微钙化104个,无微钙化44个。最终筛选出9个纹理特征,包括5个图像灰度直方图特征:均值、方差、斜度、峰值、第99百分位数值;1个自回归模型特征:θ3;3个小波变换特征:WavEnLHs-3、WavEnLHs-4、WavEnLHs-6。RF分类器构建的模型诊断斑块微钙化的准确率为0.93,AUC为0.92;SVM分类器构建的模型诊断准确率为0.91,AUC为0.90。Delong检验显示两种模型对斑块内微钙化的诊断能力差异无统计学意义(Z=1.000,P=0.320)。结论通过RF和SVM构建的影像组学模型能准确诊断颈动脉斑块内微钙化情况,且RF和SVM两种模型诊断能力接近。

Objective To construct radiomics models of micro-calcification in carotid plaques,and compare their diagnostic values.Methods Fifty-two patients with large atherosclerotic cerebral infarction admitted to Department of Neurology,Third Affiliated Hospital of Soochow University from May 2017 to November 2019 were enrolled.All patients underwent conventional carotid artery Doppler ultrasound to detect carotid plaques and Micropure®ultrasound to detect micro-calcifications in the plaques.A cross-section image with maximum numbers of micro-calcifications was chosen when there were micro-calcifications in carotid plaques;otherwise,a cross-section image with the largest area of the plaque was chosen.After all images were normalized by Photoshop software,the plaques were delineated as regions of interest using MaZda 4.6 software and 283 texture features of the plaques were automatically extracted.The texture features with the strongest predictive value were selected through consistency analysis(intrclass correlation coefficient[ICC]>0.75),two-sample t-test,Least absolute shrinkage and selection operator(Lasso)regression.The predictive models were constructed by RandomForest(RF)and Support vector machine(SVM)classifiers.The training set and test set were divided by 7:3 to analyze the classification accuracy.Receiver operating characteristic(ROC)curves were used to calculate the area under the curve(AUC)to evaluate the diagnostic values of the models.Delong test was used to compare the difference between the diagnostic values of the 2 classifiers in test set.Results A total of 148 plaque images from 52 patients were enrolled,including 104 plaques with micro-calcification and 44 plaques without micro-calcification.Nine texture features were finally selected after ICC analysis,T test and Lasso regression:5 image gray histogram features were mean,variance,skewness,kurtosis and 99th percentile(Perc.99%);1 autoregressive model feature was Teta3,and 3 wavelet transform features were WavEnLH_s-3,WavEnLH_s-4,and WavEnLH_s-6.With RF classifier,accuracy of the diagnostic model was 0.93,enjoying AUC of 0.92;with SVM classifier,that was 0.91,enjoying AUC of 0.90;Delong test showed that the diagnostic values of the 2 classifiers in test set were significantly different(Z=1.000,P=0.320).Conclusion Radiomic models constructed by RF and SVM classifiers can identify micro-calcification in carotid plaques,and the 2 classifiers share equivalent diagnostic values.