Deep Learning-Accelerated 3D Vessel Wall MRI: Prospective Evaluation of Sonic DL with Substantial Scan Time Reduction

While 3D vessel wall MRI (VW-MRI) is essential for characterizing vasculopathies, prolonged acquisition times limit its utility by increasing motion susceptibility. We aimed to evaluate Sonic DL, a next-generation AI-accelerated reconstruction technique, for VW-MRI by comparing it with a conventional protocol, using a comprehensive non-inferiority framework with a focus on diagnostic consistency.

In this prospective, randomized crossover study, 30 patients underwent post-contrast 3D VW-MRI using a conventional protocol (acceleration factor 3; 9:15) and a Sonic DL protocol (acceleration factor 10; 5:20). Three blinded readers evaluated image quality across five parameters (overall quality, vessel wall clarity, lumen suppression, cerebrospinal fluid suppression, and artifacts) and lesion conspicuity using a 4-point Likert scale (1: Excellent to 4: Poor). Non-inferiority was predefined by a margin of +0.5. Inter-protocol agreement for lesion detection, wall thickening patterns (eccentric vs. concentric), and contrast enhancement was also assessed. Quantitative assessment included the lumen-background contrast-to-noise ratio (CNR_LB). Inter-reader agreement was assessed using Fleiss' kappa.

Sonic DL achieved a 42% reduction in scan time while maintaining strict non-inferiority across all parameters; the upper limits of the 95% confidence intervals for the mean score differences were all within the predefined margin. A total of 79 vessel wall lesions were identified, with perfect inter-protocol agreement for lesion detection (κ = 1.00) and wall-thickening patterns (κ = 1.00; 58 eccentric, 21 concentric), indicating highly consistent lesion characterization between protocols. Inter-protocol agreement for contrast enhancement was also excellent (κ = 0.89, P = 0.50). Despite 10-fold acceleration, CNR_LB remained comparable between protocols (P = 0.95), and Sonic DL improved CSF suppression (P = 0.002). Inter-reader agreement was good to excellent for all qualitative categories (κ = 0.79-0.88).

Compared with the conventional protocol, Sonic DL-accelerated VW-MRI maintained non-inferior image quality and lesion assessment while reducing scan time to approximately 5 minutes. Excellent inter-protocol and inter-reader agreement demonstrated consistent lesion characterization across protocols, supporting the feasibility of Sonic DL as an accelerated vessel wall MRI technique.