Arbitrary Factor Super-Resolution for 3D Whole-Heart MRI Using a Frequency-Domain Informed Neural Network

Motivation: Three-dimensional (3D) whole-heart MRI is important in diagnosing congenital heart disease. However, it requires long acquisition times, which can lead to patient discomfort and irregular motion. Goal(s): The goal of this study is to develop a deep learning super-resolution (SR) method to decrease acquisition time, without degrading image quality. Approach: The proposed method implements a frequency-domain regularization to inform training, with the framework also enabling arbitrary factor SR. Results: The image quality metrics PSNR and SSIM show that the proposed method outperforms basic and state-of-the-art methods. Qualitative comparisons demonstrate that the proposed method better maintains diagnostically important, small anatomical structures. Impact: By implementing frequency-domain regularization inform network training and arbitrary factor super-resolution, the proposed method offers the potential to decrease acquisition time in 3D whole-heart MRI, while maintaining fine image detail important for diagnostic utility.