Predicting symptomatic carotid artery plaques with radiomics-based carotid perivascular adipose tissue characteristics: a multicenter, multiclassifier study

Abstract

Objective: This study aims to differentiate between symptomatic and asymptomatic plaques using a computed tomography angiography (CTA)-based radiomics model of perivascular adipose tissue (PVAT).

Methods: Patients were categorized into symptomatic and asymptomatic groups based on the presence or absence of acute ischemic stroke or transient ischemic attack in the anterior cerebral circulation within two weeks prior to the CTA examination. The clinical information of all patients was collected and analyzed, and the PVAT features of CTA images were further analyzed to clarify their correlation with plaque classification. K-nearest neighbors (KNN), support vector machine (SVM), logistic regression (LR), random forest (RF), linear discriminant analysis (LDA), multinomial naive Bayes (MultinomialNB), and extreme gradient boosting (XGBoost) were trained and radiomics (Rad) score was calculated using the best classifier. A combined model was further developed based on the Rad-score and independent predictors, and the calibration, receiver operating characteristic curve, decision curve analysis, and clinical applicability were evaluated.

Results: The white blood cell count and hyperlipidemia were clinically independent predictors, and ten PVAT radiomics features showed significant correlation. The XGBoost classifier showed the best performance among different classifiers, with an average AUC of 0.797 in the validation set. The combined model integrating Rad-score and clinically independent predictors was further obtained, with AUCs of 0.942, 0.797, and 0.836 in the training, external validation sets, respectively.

Conclusion: The combined model performed excellently in predicting symptomatic carotid plaques. By early identification of high-risk patients and selecting appropriate clinical decisions, it holds significant clinical potential for improving stroke prevention.

Keywords: Carotid plaques; Computed tomography angiography; Perivascular adipose tissue; Radiomics; Stroke.

Fig. 1

The patient enrollment process. Center 1, the Fifth Affiliated Hospital of Wenzhou Medical University; Center 2, the Second Affiliated Hospital of Wenzhou Medical University; Center 3, the Lishui People’s Hospital; CTA, computed tomography angiography; MRI, magnetic resonance imaging

Fig. 2

Schematic representation of the radiomics analysis steps. CTA, computed tomography angiography; LASSO, least absolute shrinkage and selection operator; KNN, K-nearest neighbors; SVM, support vector machine; LR, logistic regression; RF, random forest; LDA, linear discriminant analysis; MultinomialNB, multinomial naive Bayes; XGBoost, extreme gradient boosting

Fig. 3

(A) The coefficients of radiomics features used to construct the Rad-score; (B) a Pearson correlation coefficient heatmap of the selected features for predicting symptomatic carotid plaque. A green color denotes a positive correlation, and a purple color represents a negative correlation. The shade of the color and the size of the square indicates the degree of correlation. Rad-score; radiomics score

Fig. 4

Receiver operating characteristic curves of the three models predicting symptomatic carotid plaque in the training set (A), external validation set 1 (B), and external validation set 2 (C). Receiver operating characteristic curves of the three models predicting symptomatic carotid plaque in the training set (D), external validation set 1 (E), and external validation set 2 (F). Decision curves of the three models in the training set (G), external validation set 1 (H), and external validation set 2 (I). If the risk threshold is less than 75.9%, the nomogram model will have a greater benefit than all or no treatment in the training set. If the risk threshold is between 45% and 60%, using the nomogram model adds more benefit than the radiomics model in external validation set 1. If the risk threshold is less than 58.6%, the nomogram model will have a greater benefit than all or no treatment in external validation set 2

Fig. 5

A nomogram based on the combination of the WBC count, hyperlipidemia, and Rad-score was developed using logistic regression analysis. Rad-score, radiomics score; WBC, white blood cell

Fig. 6

Two examples of using the nomogram to predict individual risk of symptomatic and asymptomatic carotid artery plaques