Multimodal structure-guided diffusion model for Magnetic Particle Imaging reconstruction.

Magnetic Particle Imaging (MPI) is a promising biomedical imaging modality renowned for its high sensitivity. However, MPI reconstruction poses a fundamental yet severely ill-posed inverse problem, where accurate quantitative reconstruction is essential for its practical applications. Recently, diffusion models have shown great potential in addressing such inverse problems due to their generative flexibility and robustness. Nevertheless, existing diffusion-based methods often overlook structural information from auxiliary modalities, which can provide valuable anatomical priors to enhance reconstruction fidelity. In this work, we propose DiMAGnet, a multimodal diffusion-based framework for MPI reconstruction. To ensure consistency with physical measurements, we incorporate a range-null space decomposition that integrates the system matrix and measured signal directly into the diffusion process. Additionally, the proposed Structure-guided Temporal Enhancement Processing (STEP) module adaptively fuses anatomical information from external modalities across diffusion steps. To overcome the inefficiency of iterative sampling and remove dependence on structural images during inference, we further develop a knowledge distillation strategy, resulting in a lightweight student model (LiteMAGnet) capable of single-step, structure-free reconstruction. Extensive experiments on both simulated and public OpenMPI datasets demonstrate that DiMAGnet outperforms state-of-the-art methods in terms of image quality and concentration accuracy. Moreover, LiteMAGnet provides a practical and efficient solution for scenarios lacking structural data, achieving nearly 50× faster reconstruction compared to DiMAGnet. The code will be available at DiMAGnet.