Super-Resolution Deep Learning Reconstruction for Improved Image Quality of Coronary CT Angiography

Purpose: To investigate image noise and edge sharpness of coronary CT angiography (CCTA) with super-resolution deep learning reconstruction (SR-DLR) compared with conventional DLR (C-DLR) and to evaluate agreement in stenosis grading using CCTA with that from invasive coronary angiography (ICA) as the reference standard.Materials and Methods: This retrospective study included 58 patients (mean age, 69.0 years +/- 12.8 [SD]; 38 men, 20 women) who underwent CCTA using 320-row CT between April and September 2022. All images were reconstructed with two different algorithms: SRDLR and C-DLR. Image noise, signal-to-noise ratio, edge sharpness, full width at half maximum (FWHM) of stent, and agreement in stenosis grading with that from ICA were compared. Stenosis was visually graded from 0 to 5, with 5 indicating occlusion.Results: SR-DLR significantly decreased image noise by 31% compared with C-DLR (12.6 HU +/- 2.3 vs 18.2 HU +/- 1.9; P < .001). Signal-to-noise ratio and edge sharpness were significantly improved by SR-DLR compared with C-DLR (signal-to-noise ratio, 38.7 +/- 8.3 vs 26.2 +/- 4.6; P < .001; edge sharpness, 560 HU/mm +/- 191 vs 463 HU/mm +/- 164; P < .001). The FWHM of stent was significantly thinner on SR-DLR (0.72 mm +/- 0.22) than on C-DLR (1.01 mm +/- 0.21; P < .001). Agreement in stenosis grading between CCTA and ICA was improved on SR-DLR compared with C-DLR (weighted kappa = 0.83 vs 0.77).Conclusion: SR-DLR improved vessel sharpness, image noise, and accuracy of coronary stenosis grading compared with the C-DLR technique.