Abstract
BACKGROUND: Carotid artery plaques (CAPs) significantly contribute to stroke. Accurate plaque characterization is crucial for predicting stroke risk. This study explored the effectiveness of a non-contrast computed tomography (NCCT)-based radiomics model in identifying and classifying CAPs. METHODS: The dataset included 600 patients with CAPs from two centers, who were divided into training (n=400), internal test (n=100), and external test sets (n=100). Radiomics features were extracted from NCCT images. Five algorithms-Gaussian processes (GP), support vector machine (SVM), decision tree (DT), logistic regression (LR), and random forest (RF)-were employed to develop a two-level binary classification model (TBCM) and four-class classification model (FCM) for predicting the four CAP subtypes. TBCM comprised three binary classifiers. Receiver operating characteristic (ROC) curve analysis was used to evaluate model performance. RESULTS: In FCM, 38 optimal features were selected. For TBCM, 14, 13, and 22 optimal features were selected from classifiers 1-3, respectively. The GP [with areas under the ROC curves (AUCs) of 0.892-1 for three classifiers] and RF models (with AUCs of 0.883-1 for three classifiers) exhibited superior performance in the internal test set. The model combining GP and RF yielded AUCs of 0.893-1. In the external test set, the GP model achieved AUCs of 0.902-1 for three classifiers, compared with 0.939-1 for the RF model. The combined model achieved AUCs of 0.939-1 for three classifiers. CONCLUSIONS: This study highlights the efficacy of the NCCT-based radiomics model in discerning the composition of CAPs.