Magnetic Particle Imaging (MPI) faces a major bottleneck in the lengthy calibration required for system matrix acquisition. To address this, a complex-valued neural network-based super-resolution framework, CVU-SM, is proposed. Built on a U-shaped encoder-decoder architecture with complex-valued residual-in-residual dense blocks, CVU-SM preserves both magnitude and phase information to accurately reconstruct high-resolution system matrices from low-resolution inputs. Evaluated on the open source dataset, it outperforms existing deep learning methods in system matrix quality and downstream image reconstruction, demonstrating strong generalization, thereby enabling fast, high-quality MPI calibration.
Zhong et al. (Mon,) studied this question.