Study on the identification of acute ischemic stroke-related carotid atherosclerotic plaques based on head and cervical CT angiography and radiomics features

Zhang Xueke; Yao Honghuan; Zhai Duchang; Cui Manman; Lin Guihan; Chen Weiyue; Wu Yuanyuan; Hu Dongliang; Zhang Xia; Cai Wu

doi:10.3969/j.issn.1672-5921.2025.12.003

Abstract：Objective To explore the value of a multimodal integration strategy combining clinical factors,CT angiography(CTA)features,and radiomics features for identifying carotid atherosclerotic(CAS)plaques related to acute ischemic stroke(AIS).Methods CAS patients who underwent head and cervical CTA and brain MRI within 1 month between January 2019 and December 2024 at three centers:Lishui Central Hospital(Center Ⅰ),the Second Affiliated Hospital of Soochow University(CenterⅡ),and the Affiliated Changshu Hospital of Soochow University(CenterⅢ)were retrospectively and consecutively enrolled.The center with the largest sample was split into a training set and an internal validation set at 7∶3 ratio;the remaining centers' patients formed the external validation set.All the patients were grouped into AIS group and non-AIS group based on the presence of AIS.Clinical characteristics including age,sex,hypertension,diabetes,smoking history,and the latest laboratory measurements(homocysteine,total cholesterol,triglycerides,low-density lipoprotein cholesterol,high-density lipoprotein cholesterol)within 30 days prior to imaging,were collected.In the AIS group,plaques corresponding to acute infarction lesions on diffusion-weighted imaging(DWI)were evaluated as AIS-related CAS plaques.In the non-AIS group,plaques from the more severely stenosed side of the extracranial carotid artery were evaluated.CTA features,including stenosis rate of the plaque-bearing arterial segment,maximum plaque thickness,plaque ulceration(defined as contrast extending≥1 mm into the plaque),and plaque length(measured on CT curved planar reformations along the vessel centerline from the plaque's proximal to distal margins),were also collected.Differences on clinical characteristics and CTA features between AIS group and non-AIS group in the training set,internal validation set,and external validation set were compared.Factors with statistically significant differences in univariable analysis between AIS group and non-AIS group in the training set were included in multivariable Logistic regression analysis to identify independent risk factors for AIS in CAS patients and to construct a clinical-CTA model.Radiomic features were extracted from the plaque regions using the Radcloud platform and standardized via Z-score.Feature reduction was performed using least absolute shrinkage and selection operator(LASSO)regression and five machine learning classifiers(support vector machine[SVM],extreme gradient boosting[XGBoost],random forest[RF],Logistic regression[LR],and naive Bayes[NB])were constructed separately in the training set,internal validation set,and external validation set to identify AIS-related CAS plaques.The receiver operating characteristic(ROC)curves were plotted for each machine learning classifier,and area under the curve(AUC)values were compared using the Delong test.Sensitivity,specificity,and accuracy were also reported to help select the optimal machine learning classifier to derive the radiomics score(Rad-score).A combined model integrating features from clinical-CTA model and Rad-score was then developed using multivariable Logistic regression,and a nomogram was generated.Performance of the clinical-CTA model,radiomics model and the combined model in the training set,internal validation set,and external validation set was evaluated using ROC analysis,decision curve analysis(DCA),and calibration(Bootstrap,B=1 000).Results A total of 231 CAS patients were included,119 patients in the training set(58 AIS,61 non-AIS),52 patients in the internal validation set(25 AIS,27 non-AIS)and 60 patients in the external validation set(26 AIS,34 non-AIS).(1)In the training set,high-density lipoprotein cholesterol level was lower in the AIS group compared to the non-AIS group(P=0.010),but no significant difference was found in the internal or external validation sets(both P＞0.05).In all three datasets,maximum plaque thickness and plaque ulceration ratio were significantly higher/more frequent in the AIS group than those in the non-AIS group(all P＜0.05).Other clinical and CTA feature differences between two groups were not statistically significant(all P＞0.05).(2)Multivariable Logistic regression analysis based on the training set revealed that plaque ulceration(OR,4.231,95%CI 1.375-13.018,P=0.012)and maximum plaque thickness(OR,1.428,95%CI 1.004-2.030,P=0.048)were independent risk factors for AIS in CAS patients and were included in the clinical-CTA model.(3)For identifying AIS-related CAS plaques,XGBoost achieved AUCs of 0.851,0.820,and 0.786 in the training,internal validation,and external validation sets,respectively.XGBoost outperformed NB in all datasets(all P＜0.05),outperformed RF in the internal validation set(P=0.025),and outperformed LR in the external validation set(P=0.043),while showing no significant difference compared to SVM in all three datasets(all P＞0.05).In the external validation set,XGBoost achieved an accuracy of 76.65%,sensitivity of 76.92%,and specificity of 76.47%,indicating a more balanced performance than SVM(accuracy 71.83%,sensitivity 61.54%,specificity 82.35%).XGBoost was selected as the optimal machine learning classifier for building the radiomics model.(4)A combined model integrating plaque thickness,plaque ulceration from the clinical-CTA model,and the Rad-score from the XGBoost radiomics model achieved AUCs of 0.896(95%CI 0.836-0.955),0.845(95%CI 0.723-0.967),and 0.807(95%CI 0.684-0.897)in the training,internal validation,and external validation sets,respectively.Delong test results showed that the combined model outperformed the clinical-CTA model in the training and internal validation sets(both P＜0.05),while no significant difference was found compared to the radiomics model(both P＞0.05).In the external validation set,AUCs were not significantly different among the three models(both P＞0.05).DCA showed that the combined model provided high net benefit within threshold probabilities of 0.37-0.92(training set),0.07-0.78(internal validation set),and 0.06-0.85(external validation set).After optimism correction using Bootstrap(B=1 000),calibration curves closely followed the ideal line,and apparent Brier scores were 0.142,0.086,and 0.095,optimism Brier scores was 0.150,0.112 and 0.114 in the training set,internal validation set and external validation set,respectively,indicating low prediction error and good calibration.Conclusion A multimodal strategy integrating head and cervical CTA with radiomics features improves the identification of AIS-related CAS plaques,demonstrating good diagnostic performance and promising clinical applicability.