"Super-Resolution" Deep Learning Image Reconstruction in Dynamic Myocardial Perfusion: A Prospective Evaluation of...

To evaluate the impact of Super-Resolution Deep Learning Reconstruction (SR-DLR) (Canon Medical Systems Corporation) on image quality and myocardial hemodynamic parameters in dynamic myocardial computed tomography (CT) perfusion compared with filtered-back projection (FBP), hybrid iterative reconstruction (IR), and normal-resolution deep learning reconstruction (NR-DLR).

This prospective single-center study included 25 patients (mean age ± SD, 65 ± 10; 21 men) who underwent dynamic myocardial CT perfusion. For qualitative analysis, image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) were assessed, while qualitative analysis included overall image quality and lesion visibility. Myocardial blood flow (MBF) at rest and stress, as well as coronary flow reserve (CFR) were analyzed. Image quality and hemodynamic parameters were compared across 4 reconstruction methods.

SR-DLR achieved the lowest image noise (20.33 ± 2.45 HU), significantly lower than FBP (145.20 ± 74.81 HU), hybrid IR (47.19 ± 10.02 HU), and NR-DLR (22.92 ± 2.63 HU) ( P < 0.001). In rest imaging, SR-DLR showed significantly higher SNR (6.71 ± 1.88) and CNR (15.41 ± 5.48) compared with other reconstruction methods ( P < 0.001). Similar improvements were observed in stress imaging, with SR-DLR providing significantly enhanced SNR and CNR compared with all other methods. The mean CFR was 2.75 ± 1.88 for SR-DLR, 2.75 ± 1.99 for NR-DLR, 2.74 ± 2.44 for hybrid IR, and 2.56 ± 3.17 for FBP, with no statistically significant differences observed in any pairwise comparisons. Qualitative analysis showed that SR-DLR achieved the highest overall image quality and lesion visibility, significantly outperforming FBP and comparable to hybrid IR and NR-DLR.

SR-DLR and NR-DLR significantly enhanced image quality by reducing noise and improving SNR and CNR while maintaining hemodynamic quantification.