Residual Convolutional Neural Network for Temporal Super-resolution 4D Flow MRI

Motivation: 4D Flow MRI can assess full volumetric flow and enable hemodynamic parameter calculation. However, clinical application is limited as high-resolution, low-noise data require long acquisition times. Goal(s): We aimed to develop a post-processing method for denoising and temporal super-resolution of 4D Flow MRI. Approach: We propose a convolutional neural network, trained on cardiovascular in-silico models and validated on in-vivo datasets (N=3). Results: The network achieved high alignment on the in-silico test set and the in-vivo datasets, specifically at peak flow timepoints. The network can generalize to unseen domains, identify and enhance fluid regions, without boundary segmentations or retraining. Temporal resolution is increased twofold. Impact: By showcasing the feasibility of temporal super-resolution networks, we aim to enhance 4D Flow MRI data and increase its clinical applicability. Including flow and hemodynamic parameter analysis in broader clinical applications has the potential to improve cardiovascular disease diagnosis.