Abstract
BACKGROUND: Early detection of carotid artery stenosis is critical for effective stroke prevention. We investigated the feasibility and diagnostic efficacy of radiomics analysis based on non-contrast computed tomography (NCCT) in identifying carotid artery stenosis, and aimed to establish a radiomics-based machine-learning model as a non-invasive, cost-effective auxiliary diagnostic tool. METHODS: This retrospective study included 260 patients who underwent both non-contrast and contrast-enhanced carotid computed tomography (CT) scans at Jiaxing First Hospital between January 2021 and December 2024. The degree of carotid stenosis was confirmed by CT angiography. Patients were enrolled according to predefined inclusion and exclusion criteria. Clinical data from 260 eligible patients were collected and stratified for comparative analysis. All CT images were acquired on a single scanner. Two experienced radiologists jointly identified the stenotic segments on NCCT images and used ITK-SNAP (software tool for manual outlining of ROIs) software to manually outline the regions of interest (ROIs) in 2D along the vessel wall of the plaques themselves. First, the stenotic segments were accurately localized on computed tomography angiography (CTA) images; after rigid registration, these reference positions were mapped to the corresponding NCCT images to guide the precise outlining of ROIs. To ensure the reproducibility of annotations, a third radiologist then reviewed all ROIs: 30 randomly selected cases were independently re-segmented, and the intraclass correlation coefficient (ICC) of the extracted radiomic features was calculated. Finally, features with ICC > 0.85 were retained, and this result indicated high inter-observer consistency. Radiomic features were extracted using PyRadiomics after image preprocessing, which comprised gray-level normalization, resampling, and wavelet filtering. Feature selection was performed using Spearman correlation analysis and Lasso regression with tenfold cross-validation. The selected features were used to train multiple machine-learning classifiers, including logistic regression, support vector machine (SVM), and naïve Bayes. The dataset was randomly split into training and test sets at an 8:2 ratio. Model hyperparameters were optimized in the training cohort via fivefold cross-validation. Model performance was assessed using metrics such as accuracy, area under the curve (AUC) of the ROC curve, sensitivity, specificity, precision, F1-score, and confusion matrix. A decision curve analysis (DCA) was conducted to evaluate the clinical utility. RESULTS: Ninety-seven radiomic features were initially extracted; after selection, 12 with non-zero coefficients remained. The SVM model achieved an AUC of 0.874 in the training set and 0.774 in the test set, indicating strong discrimination and generalizability. Confusion-matrix analysis confirmed reliable separation of positive and negative cases with a low misclassification rate. Decision-curve analysis demonstrated that the model conferred superior net clinical benefit across threshold probabilities from 0.2 to 0.8 compared with both treat-all and treat-none strategies. CONCLUSIONS: Radiomics analysis of NCCT imaging provides an effective, contrast-free means of identifying carotid artery stenosis. Our model demonstrated robust diagnostic accuracy and clinical potential, especially when contrast-enhanced imaging is contraindicated. Prospective multi-center external validation and further workflow automation are nevertheless required to support broad clinical adoption.