MRzero - Automated discovery of MRI sequences using supervised learning.

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Machine Learning
• Medical Physics

Background:

• Conventional MRI sequence design is complex and time-consuming.
• Optimizing MR sequences for specific contrasts requires extensive expertise.
• Exploration of novel sequence strategies is limited by traditional methods.

Purpose of the Study:

• To develop a supervised learning framework for automated MR sequence generation and reconstruction.
• To enable efficient exploration of novel MR sequence strategies based on target contrast.
• To integrate a flexible, task-driven cost function for optimization.

Main Methods:

• Simulated end-to-end scanning and reconstruction using differentiable Bloch equation simulations.
• Supervised learning framework trained on target contrasts (e.g., conventional MR images, T1 maps).
• Optimization from scratch using a loss function incorporating data fidelity, SAR penalty, and scan time.

Main Results:

• MRzero successfully learned gradient and RF events to generate target images from scratch.
• A neural network in the reconstruction module enabled learning of arbitrary targets.
• Experiments were validated on a 3T Siemens PRISMA system using phantoms and in vivo human brain imaging.

Conclusions:

• Automated MR sequence generation is achievable using differentiable Bloch equation simulations and supervised learning.
• This framework facilitates the discovery of novel MR sequence strategies.
• The approach holds promise for advancing MRI acquisition and reconstruction techniques.