Deep Generative Adversarial Networks for Direct Super-resolution Magnetic Particle Imaging Reconstruction

Abstract

Due to the complex physical behavior of the nanoparticles, it is challenging to reconstruct the image from the magnetic particle signal. Since the system matrix reconstruction is time-consuming and the x-space reconstruction ignores the relaxation effect of the particles, we proposed to reconstruct the high-resolution 2-D image directly from the 1-D magnetic particle voltage signal by using the machine learning method of generative adversarial network (GAN). We first built a large simulation image dataset, which includes 291,597 binary images and each image’s corresponding MPI voltage signal simulated with our developed MPI simulation software MPIRF. By using the large simulation dataset, we trained a conditional-GAN model, which we termed “MPIGAN”, that can successfully convert the 1-D MPI voltage signal to the high-resolution MPI image directly and efficiently. Experiment results showed that, compared to the traditional methods, our proposed MPIGAN could better retrieve the fine-scale structure of the patterns of images from the 1-D voltage signals, and achieved better reconstruction performance in both visual effects and quantitative assessments, e.g., SSIM, MSE, PSNR. Our study provides a promising end-to-end AI solution for the efficient and high-resolution magnetic particle imaging reconstruction.