Applications of generative adversarial networks for super-resolution of cerebrovascular 4D Flow MRI

Motivation: 4D Flow MRI provides full-field mapping of blood flow. However, challenges remain concerning resolution and noise, particularly in the intracranial space. While convolutional neural networks (CNNs) have demonstrated potential to provide super-resolved 4D Flow MRI, enhancing near-wall flows remain difficult. Goal(s): To evaluate various Generative Adversarial Network (GAN) setups for enhanced super-resolution and denoising of 4D Flow MRI data, focusing on vessel-boundary performance. Approach: We trained and validated multiple GAN setups on anatomically and sequence-realistic synthetic 4D Flow MRI data. Results: Results indicate that Wasserstein GAN outperforms established CNN approaches for near-wall velocity enhancement, suggesting improvements in boundary voxel recovery with generative networks. Impact: This study highlights the potential of generative adversarial networks to enhance super-resolution in 4D Flow MRI, enabling more accurate intracranial flow assessments near vessel walls.