SFNet: Spatial-Frequency Collaborative Network for Fast System Matrix Calibration in Magnetic Particle Imaging

Magnetic particle imaging (MPI) is a promising high-resolution molecular imaging technology that reconstructs images using system matrix (SM) describing the nonlinear response of magnetic nanoparticles excited along the field-free region (FFR) trajectory. Given SM is required whenever scan parameters or tracer changes, fast SM calibration is important in practical deployment. While sparse sampling enables fast reconstruction without a full-size SM, the predefined sampling grid often misaligns with FFR trajectory. This misalignment leads to incomplete and spatially inconsistent signal acquisition, resulting in SM distortion in both spatial-domain and frequency-domain and consequently introducing structural degradation and artifacts in reconstructed images. Here, we propose a spatial-frequency collaborative network (SFNet), which comprises a frequency-domain feature extraction module and a spatial-frequency feature interaction module. These modules are iteratively optimized to enhance cross-domain representation learning, thereby correcting SM distortion and improving reconstruction fidelity. Experimental results show that the outstanding performance of SFNet against state-of-the-art methods.