Carotid vulnerable plaques (CVPs) represent a major cause of ischemic stroke, yet current diagnostic methods lack sufficient precision for early detection. Spectral computed tomography (CT) enables detailed plaque characterization, but its clinical utility depends on advanced analytical techniques. A Fourier Transform–enhanced U-Net model was developed using spectral CT data from 214 patients at Qiqihar Medical College. Transfer learning was applied for carotid segmentation, supported by extensive data augmentation and ten-fold cross-validation. Model performance was assessed using sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). A patient-level independent test set was reserved to evaluate model performance. The FTC-enhanced U-Net achieved a Dice similarity coefficient (DSC) of 0.89 and an intersection over union (IoU) of 0.80, accurately delineating arterial and plaque boundaries. Stable classification performance was maintained across plaque risk levels, with accuracies of 86% for low-risk, 83% for medium-risk, and 87% for high-risk plaques. Ten-fold cross-validation yielded an AUC of 0.88 ± 0.03, an accuracy of 85.3% ± 2.1%, and an F1-score of 0.84 ± 0.02, confirming strong robustness and generalizability. Artificial intelligence (AI)–based predictive models derived from spectral CT imaging show substantial potential for early diagnosis and prognostic assessment of CVPs. The FTC-enhanced U-Net provides an efficient and reliable tool for clinical decision support, facilitating broader adoption of AI in vascular imaging and improving diagnostic precision and patient management.
Ni et al. (Fri,) studied this question.