Automated Analysis of Carotid Plaque Components in CT Angiography: Performance and Prognosis in Symptomatic Carotid...

To develop and validate the performance and prognostic value of a deep-learning (DL) model for carotid plaque component quantification on CTA.

A multicenter retrospective study was conducted in three stages: Stage 1: Model development and concordance analysis: A DL model was developed for plaque detection and segmentation using 2164 CTA scans (Cohort 1). DL-radiologist measurement agreement was assessed using ICC and Spearman's test (Cohort 2). Stage 2: Diagnostic validation: Performance was validated against 1) 118 co-registered CTA-OCT image pairs (Cohort 3) and 2) 146 patients with paired HR-MRI and CTA (Cohort 4). Stage 3: Prognosis validation: In 610 symptomatic patients (Cohort 5), multivariable Cox regression assessed the association between lipid core burden (LCB) and recurrent cerebrovascular events, and the incremental predictive value of LCB was quantified by ΔAUC and NRI.

The DL model achieved a detection sensitivity of 0.85 with an average of 1.08 false positives per case. It demonstrated good-to-excellent agreement with radiologist assessments. In Stage 2, DL-driven lipid core component is associated with high-risk plaques identified by OCT and HR-MRI. In Stage 3, in a median 2-year follow-up, LCB independently predicted recurrent cerebrovascular events (HR: 1.08, 95% Cl: 1.03-1.13, P<0.001). The incorporation of LCB provided incremental risk stratification beyond clinical and CTA-driven anatomical factors (ΔAUC +0.09, NRI: 0.22, P=0.001).

The DL model accurately quantifies carotid plaque components on CTA, with LCB adding prognostic value for recurrent cerebrovascular events in patients with symptomatic carotid stenosis.