Carotid artery disease diagnosis comprehensive review and two stage artificial intelligence framework

Résumé

Abstract Carotid artery occlusion is a major risk factor for stroke and cerebrovascular disease, often developing silently without warning. Early and accurate diagnosis of carotid artery plaque is crucial for improving treatment outcomes and reducing potential complications. Thus, there is a great demand for computer-aided diagnostic techniques for the early diagnosis of carotid artery disease with improved accuracy. However, existing literature remains fragmented, typically decoupling initial classification from detailed lesion tracing. This research is organized under three primary axes to bridge this clinical and technical gap. The first axis includes a structured review-of-reviews study that evaluates previously published review papers, while the second axis presents a complete methodological overview of authenticated techniques utilized in carotid diagnosis, such as carotid lumen segmentation, plaque segmentation, carotid classification, and IMT measurement. By mapping the limitations within these domains, the third axis presents a practical two-stage framework, including classification of the carotid artery into normal and plaque-containing, and then segmentation of its plaque to mirror a radiologist's natural decision-making hierarchy. Five lightweight transfer learning networks with multiple preprocessing filters were tested for carotid classification on an independent external clinical dataset. It is followed by a YOLOv11n-based segmentation head used to delineate plaques using expert-generated annotations. Methodologically, ShuffleNet combined with local laplacian filtering produced the best classification results, achieving 98.25% accuracy, 96.17% sensitivity, 100.00% specificity, 100.00% precision, and 98.05% F1-score. Crucially, this screening gate protects the subsequent segmentation stage from processing normal scans and eliminates false positives. The YOLOv11n-seg pipeline achieves a Dice Similarity Coefficient of 90.84%, a mask recall of 77.73%, and a mAP50 of 83.21% on unseen clinical scans. By filtering out healthy frames early, this dual-stage pipeline reduces computational overhead to processing times of 6.37 ms for screening and 6 ms for segmentation per frame.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0