Deep reinforcement learning based magnet design for arm MRI system

Throwing elbow is a common sports injury that can be diagnosed more promptly by portable magnetic resonance imaging (MRI) than by conventional superconducting MRI. The main magnet of portable MRI systems typically consists of permanent magnets. However, the limited length of the main magnet can lead to poor magnetic field homogeneity, resulting in image distortion. Therefore, it is essential to optimize the permanent magnets arrangement. The traditional genetic algorithm (GA) lacks a timely feedback mechanism during the search process, and there is no gradual interaction with the magnetic field map in a single iteration, which has potential for improvement. To address this problem, a deep reinforcement learning (DRL) based magnet design algorithm for an arm MRI system is proposed. Based on the magnetic field map, the method is performed to design a high homogeneity magnet under the weight constraint on the main magnet, significantly better than the multi-objective method NSGA-II. The results indicate that the proposed method achieves a 26.7% gain in homogeneity at a higher average magnetic field strength compared to the GA. In a scenario where the volume of the main magnet is uniform and without weight constraint, an adaptive search mechanism is proposed that enables the method to achieve a 62.90% improvement in homogeneity compared to the GA.