Super-resolution of Magnetic Resonance Images with Physics-Informed Neural Networks

Niveau : Master 2

Période : 2024

Keywords: Physics-Informed Neural Network, Medical Imaging, Super-resolution, 4D Flow MRI, Fluid Mechanics

Master Thesis Subject:
Deep neural networks has proven its efficiency to successfully manage complex tasks and particularly when the database is large. In case of small database, some strategies can be set such as data augmentation, fine tuning, loss regularization, etc. Consequently, the lack of data is balanced with a priori information resulting of transformation, another database or a loss. In the latter strategies, a new type of network emerged with the Physics-Informed Neural Networks (PINN) [RPK17, RPK19, FPRB + 20]. The consideration of physics equations in the loss regularization makes the network more robust to the lack of data and reduce the need of a big neural network. The main objective of this Master thesis is to develop new strategies to apply efficiently PINN on medical images.
In the context of a collaboration between the cardiac and vascular diagnostic imaging unit of the University Hospital of Nantes and the IPI team of the LS2N laboratory, the candidate will work on medical imaging with specific interest in super-resolution. In clinical routine, Phase-Contrast MRI is used to image the anatomy and velocity across a 2D acquisition plane and along the cardiac cycle. In the last decades, 4D flow MRI interest increased with the possibility to image a 3D region of interest [MFK + 12]. Specifically, the anatomy and three velocity components are measured along the cardiac cycle. Acquisition time being constrained, the image signal-to-noise ratio and resolution are degraded. A former study [LCI + 20] demonstrated the impact of the image noise and resolution on hemodynamic biomarkers. In terms of methodology, 4D flow MRI super-resolution literature can be distinguished in three categories : computer vision approaches [dHvPJV14, CG17], inverse problem solving [RNNC15, FPRB + 20], and machine learning solutions [BBV + 17, FBB + 18, FSD + 20]. In this work, we expect to focus on neural network based solutions. A specific focus is set on the solution relative to Physics informed neural networks.

Thus, the Master thesis objectives are:

1. Elaborate a bibliography on Physics-Informed Neural Networks and its variant solutions;
2. Reproduce the solution of [FPRB + 20] and assess its performance;
3. Propose hybrid solutions with machine learning to reduce computation load.

Sébastien LEVILLY sebastien.levilly@ls2n.fr

Simon PERRIN simon.perrin@ls2n.fr