Integrating Flow Imaging and Deep Learning into Patient-Specific Models Following Myocardial Infarction

4D-flow magnetic resonance imaging (MRI) provides non-intrusive blood flow reconstructions in the left ventricle (LV) and other cardiac chambers and has the potential to become a key tool in both research and clinic. However, low spatio-temporal resolution and the presence of significant noise artifacts hamper the accuracy of derived haemodynamic quantities and thus limit the effectiveness of the modality to establish links between haemodynamic abnormalities and pathologies. Furthermore, models that are constrained by boundary conditions are impacted by additional uncertainty which arises due to the low spatial resolution of the structural cine-MRI.

4D-flow MRI data corruption, introduced by low resolution and noise artefacts, may be alleviated through super-resolution and de-noising methods, which have been explored in the literature for haemodynamic flow in the vasculature. In this thesis, a physics-informed neural network (PINN) model is introduced to provide super-resolution and de-noising, specifically of cardiac 4D-flow MRI. The model is constrained through weak enforcement using the low-resolution 4D-flow MRI data, the no-slip boundary condition on the endocardium and the governing physical equations. Model components are compared and incorporated to address specific challenges introduced by modelling haemodynamic flow in the heart chambers, such as flow across a range of length and time scales within a heavily deforming domain. Validation of the model is performed across synthetic and in vivo studies, evaluating the robustness of the model to uncertainty in both the 4D-flow MRI data and the position of the deforming endocardium. Following this, the model is applied to a small cohort of LV remodelling patients.

It is demonstrated that the PINN model is able to effectively upsample and de-noise the velocity field across a range of spatio-temporal resolutions and signal-to-noise ratios (SNR), and is robust to positional uncertainty in the deforming endocardium for flow variables measured away from the domain boundaries. Further, variables that are not directly measured, such as relative pressure and flow derivatives, are reconstructed to an acceptable degree of accuracy. In the dual-resolution in vivo validation study, it is shown the model is generally independent to the spatial resolution and SNR of the input 4D-flow MRI data.

Through synthetic and in vivo validation studies, it is demonstrated that the PINN model introduced in this thesis is effective. It is concluded that the use of this type of model is feasible for super-resolution of cardiac 4D-flow MRI data, although certain limitations should be addressed and the model should be further validated using in vivo or in vitro data.