Spring School Hackathon: Optimizing MRI Coil Configurations with Machine Learning
As part of the Spring School on Physics Informed Machine Learning for Medical Sciences, I participated in a hackathon focused on optimizing MRI coil configurations. The event brought together participants from diverse backgrounds to tackle a real-world challenge at the intersection of physics, machine learning, and medical imaging.
Project Overview
The hackathon tasked teams with maximizing the homogeneity of the B1+ field or minimizing the peak Specific Absorption Rate (SAR) in a simulated MRI environment. Using precomputed electric and magnetic field data, we adjusted the phase and amplitude of individual dipoles to achieve optimal coil configurations for improved imaging and safety.
Technical Details
Each team received data representing the contributions of individual dipoles. The challenge was to combine these dipoles by tuning their phase (φ) and amplitude (A), optimizing for two cost functions:
- B1+ Homogeneity: Maximize mean field strength and minimize variability.
- B1+ Homogeneity with SAR Constraint: Add a penalty to reduce peak SAR while maintaining homogeneity.
Teams explored various optimization techniques, including global search algorithms, gradient-based methods, and parallel processing, to balance performance and safety.
Results
Seven teams participated, each employing unique optimization strategies such as gradient descent and genetic algorithms. All teams improved the cost function values, with the best achieving up to 65% improvement for B1+ homogeneity and up to 430% for the SAR-constrained task. Our team was awarded second place in the competition.