Spatiotemporal 4D-UNet for Physics Consistent Super-Resolution and Denoising of 4DFlow-MRI

Motivation: To improve the quality of 4DFlow-MRI allowing for an increase in the accuracy of key measured hemodynamic parameters. Goal(s): We aim to reduce noise and artifacts in 4DFlow-MRI velocity images while adhering to the physics of blood flow. Approach: We implement a spatiotemporal 4D-UNet network that can take an entire sequence and denoise and super-resolve in-vitro MRI data of a phantom model with 87% arterial stenosis. Our model is trained to reduce a physics divergence and vorticity residual loss. Results: We observe reductions in divergence and vorticity residual errors that demonstrate our network's capabilities to improve 4DFlow in-vitro data. Impact: Our model can improve 4D-Flow MRI data that has been hindered by noise, artifacts, and lower resolution. As a result, hemodynamic parameters that are critical for diagnosing various disease can be more accurately measured