Degradation Consistent Conditional Diffusion Model with Frequency Alignment for 3D Mag-netic Particle Image Enhancement

Abstract

Magnetic Particle Imaging (MPI) provides quantitative visualization of magnetic nanoparticle distributions but suffers from limited spatial resolution and anisotropic blurring due to system nonlinearities and hardware imperfections. To address the low-resolution issues caused by various noise sources, we propose a Degradation-Consistent Conditional Diffusion Model (DCCDM) for three-dimensional (3D) MPI image super-resolution. The proposed model introduces two physically interpretable constraints: a degradation-consistency loss, utilizing a degradation operator to ensure that the reconstructed high-resolution volume remains consistent with the observed low-resolution data, without requiring a known system matrix, and a frequency-alignment regularization, enforcing structural fidelity in the low-frequency band and detail enhancement in the high-frequency band. Experiments on simulated 3D MPI datasets demonstrate that DCCDM achieves superior performance compared with existing CNN-, GAN-, and diffusion-based models.