Neurala implicita representationer för brusreducering och superupplösning av cerebrovaskulär 4D Flow MRI

Haemodynamic assessment is key to diagnosing and managing cardiovascular disease. 4D Flow MRI is a non-invasive imaging technique that enables mapping of the blood flow, however, its clinical utility is limited by low spatial resolution, long acquisition times and effective imaging noise. The objective of this thesis is to study the super-resolving potential of neural implicit representations in cerebrovascular 4D Flow MRI. For this purpose, two networks, sinusoidal representation network (SIREN) and Fourier Feature Networks (FFNs), were implemented and evaluated. The networks were trained under different physics-informed and spectral bias conditions, using in-silico 4D Flow MRI data. The generalisation capabilities of the networks were studied across different geometries from three generated \textit{in-silico} CFD simulations, originating from acquired patient-specific 4D Flow MRI data. The results showcased the capability of the networks to accurately reconstruct the cerebrovascular flow. Using physics-informed configurations proved positive for the prediction of physically feasible flows. Furthermore, the inherent spectral bias values were observed to have a fundamental impact on super-resolution performance. Moreover, optimal ranges for spectral bias were observed to coincide across the three studied cases, highlighting the generalisation capabilities of this approach. These findings highlight the applicability of the neural implicit representations for super-resolution of the cerebrovascular flow, which could set a precedent for future studies aiming to improve the clinical applicability of 4D Flow MRI.