Applicant
Prof. Dr. Moritz Zaiss
Institute of Neuroradiology
University Hospital Erlangen
Friedrich-Alexander-Universität Erlangen-Nürnberg
Project Overview
We recently discovered a new realistic and efficient way of computing magnetic resonance imaging signals by a Fourier Domain approach with Phase Distributed Graphs (PDG) [1]. This approach itself leads to two order of magnitude faster simulations, and has unique acceleration possibilities, which we want to unleash within this project.
Objectives:
Extend the framework of MR-zero [2], a novel self-supervised learning framework to automatically optimize Magnetic Resonance (MR) sequences, for the new Fourier Domain approach as massively Distributed and Data parallel simulation on HPC
systems.
Improve frame for memory efficiency, numerical precision, scalability, performance optimization of computational time. Develop simulation as Application Programming Interface (API) to handle ultra realistic, high resolution, multi-slice, longer 3D sequences
for Application specific MR Imaging.
Extend API compatibility with for generating synthetic MRI data for Machine learning models, assist with model-based reconstruction, MRI fingerprinting and other quantitation method.
Accelerate the sequence generation for real-time optimization by re-implement sequence as neural network layers in PyTorch. All the above objectives can be achieved by the following 5 step work strategy.
References
[1] E. Jonathan, “Phase distribution graphs for fast, differentiable, and spatially encoded Bloch simulations of arbitrary MRI sequences,” Magnetic Resonance in Medicine, 2024.
[2] A. Loktyushin, “MRzero – Automated discovery of MRI sequences using supervised learning,” Magnetic Resonance in Medicine, 2021.