Impact of super resolution deep learning reconstruction with a 1024 matrix on CAD-RADS classification of intermediate coronary stenosis: A bicentric retrospective study

Introduction Moderate coronary stenosis (CAD-RADS 3) is a diagnostic gray zone in coronary CT angiography (CCTA). We evaluated whether super-resolution deep learning reconstruction (SR-DLR), particularly with a 1024 matrix, modifies CAD-RADS classification and inter-reader agreement. Materials and methods: Thirty-nine CCTA examinations from two centers, each containing at least one 50–69% stenosis identified on the routine normal-resolution deep learning reconstruction (NR-DLR), were reconstructed with NR-DLR and SR-DLR (512 and 1024 matrices). Five experienced readers blindly assigned CAD-RADS categories. Agreement was assessed with Gwet's AC2 and paired reclassification with exact McNemar tests. Subgroup analyses (plaque composition, initial image quality) were exploratory. Results Forty-six lesions were analyzed (19 calcified, 16 mixed, 10 non-calcified, one in-stent). Agreement remained excellent (AC2 0.89 [0.86–0.92], 0.87 [0.83–0.90] and 0.84 [0.81–0.88] for NR-DLR, SR-DLR-512 and SR-DLR-1024). Consensus CAD-RADS 2 rose from 6.5% to 21.7% with SR-DLR-1024 (p = 0.016; 10 lesions downgraded, 2 upgraded), predominantly in lesions with non-calcified components. Per-reader CAD-RADS 2 rates ranged from 13% to 30%, increasing significantly for two of five readers (p = 0.004). Conclusion SR-DLR-1024 modified CAD-RADS classification in about one fifth of intermediate lesions while preserving excellent agreement. Without an invasive reference standard, these reclassifications reflect a change in reader interpretation rather than a proven gain in accuracy.