TransVort: A Temporally Coherent Physics-Guided Neural Network for Super-Resolution and Denoising 4D flow MRI of CSF

Cerebrospinal fluid (CSF) plays diverse roles in cushioning the brain, regulating intracranial pressure, and clearing metabolic wastes. While alterations in CSF flow are implicated in diseases such as hydrocephalus, in vivo assessment of the underlying flow structures is limited. 4D flow MRI can measure CSF velocities within the cerebral ventricles but with low spatiotemporal resolution and velocity-to-noise ratios. We developed TransVort, a temporally coherent physics-guided neural network (PGNN), to denoise and super-resolve 4D flow MRI. This network uses a novel loss function based on the vorticity transport equation to improve temporal coherence. We compare this network performance to our previously proposed divergence-constrained PGNN. Our preliminary results document the TransVort network, when applied to synthetic 4D flow MRI in the ventricles, outperforms our prior PGNN by reducing the root mean square error in core voxels by 17.9% vs 10.5% and in edge voxels by 39.7% vs 34.5% relative to linear interpolation. We assess performance using in vitro 4D flow MRI acquired at two resolutions in 3D printed phantoms and in vivo 4D flow MRI in a healthy subject. By improving in vivo imaging of CSF flow, we hope to further inform investigations of cerebroventricular diseases.