Super-resolution deep learning reconstruction achieves low-dose high-resolution imaging in pediatric cardiac computed tomography.

Abstract

Background: Cardiac computed tomography (CT) is widely utilized in pediatric cardiology, but minimizing radiation exposure is essential. Recently, super-resolution deep learning reconstruction (SR-DLR) has emerged as a potential advancement over conventional deep learning reconstruction (C-DLR).

Objective: To evaluate the potential of SR-DLR for further radiation dose reduction in pediatric cardiac CT through a comparison with C-DLR.

Materials and methods: The phantom-based study assessed noise power spectrum and spatial resolution. For the clinical study, we compared nine C-DLR images (conventional radiation dose) and 11 SR-DLR images (lower radiation dose). Evaluation metrics included noise levels, contrast-to-noise ratio, and edge rise distance and edge rise slope.

Results: In the phantom-based study, SR-DLR (250 mA) displayed equivalent noise characteristics as C-DLR (800 mA) while achieving superior spatial resolution. In clinical images, preliminary comparisons within the high-dose group revealed that SR-DLR significantly outperformed C-DLR in spatial resolution metrics, specifically edge rise distance and edge rise slope (P<0.001). Furthermore, despite the lower radiation dose used for SR-DLR than for C-DLR (volume CT dose index, 4.25±0.54 mGy vs. 13.3±6.43 mGy), there were no significant differences in noise, contrast resolution, and spatial resolution between the groups.

Conclusion: SR-DLR maintains non-inferior image quality compared with C-DLR even at reduced radiation doses. These findings substantiate the potential for further dose reduction in pediatric cardiac CT without compromising image quality.