Function-based Deep Learning Autoencoder for Passive Shimming in a Portable MRI Magnet with Enhanced Field Homogeneity.

High-field magnetic resonance imaging (MRI) offers excellent sensitivity and spatial resolution, yet its substantial cost and limited mobility restrict deployment in decentralized or resource-constrained environments. Portable MRI systems offer a complementary imaging modality to high-field MRI, but their performance depends critically on the uniformity of the main magnetic field (B0). Passive shimming, achieved by placing ferromagnetic materials within the magnet bore, is widely used to correct spatial field variations. Existing optimization strategies- including linear programming and evolutionary algorithms-can improve B homogeneity, but often struggle to achieve an effective trade-off among field strength, uniformity, and shim mass. To address this challenge, a function-based deep learning autoencoder (FDLAE) is proposed to optimize the spatial distribution of passive shims. The framework operates without a precompiled dataset and performs online self-supervised learning to iteratively refine the shim configuration. This approach improves B0 homogeneity from 4873 ppm to 547 ppm (parts per million) and increases the average field strength from 80.402 mT to 80.948 mT. Compared with linear programming (LP), it achieved lower inhomogeneity, reduced the shim mass by 0.29 kg, and shortened the optimization time from approximately 1.2 to 0.5 h. MRiLab simulations also showed a reduction in image NRMSE from 0.889 to 0.543. Relative to genetic-algorithm-based shimming, the resulting field uniformity is improved by nearly a factor of five. These results demonstrate that FDLAE may provide an efficient and comparatively robust passive-shimming approach for portable MRI magnets.