End-to-end MPI image reconstruction with dual-task generative adversarial network

Magnetic Particle Imaging (MPI) is an emerging technology utilized for mapping the spatial distribution of magnetic nanoparticles within biological tissues. The complexity of the physical behavior of nanoparticles poses a challenge to directly reconstructing images from signals received by MPI devices. High-quality image reconstruction is crucial in MPI applications. Here, we propose a new dual-task generative network to achieve high-quality MPI image reconstruction. In this model, the generator simultaneously undertakes MPI image segmentation and end-to-end MPI image reconstruction tasks. The image reconstruction is the main task of the generator, which generates images that closely approximate ground truth through adversarial strategies to deceive the discriminator. The image segmentation serves as the auxiliary task, guiding the main task to focus on key areas and boundary regions during the generation process. To the best of our knowledge, it is the first attempt to perform feature sharing on the synthesis and the segmentation of MPI images in medical image analysis. Our method outperforms existing calibration methods on OpenMPI dataset and simulation dataset. Our study verified that the generalization ability of the proposed dual-task model outperforms the previous MPI reconstruction methods using single-task model as well as traditional MPI reconstruction methods. Our experimental results also demonstrated that MPI reconstruction task and the segmentation task significantly promote each other.