Deep Learned Super Resolution of System Matrices for Magnetic Particle Imaging.

Abstract

Magnetic Particle Imaging (MPI) is a new imaging technique that allows high resolution & high frame-rate imaging of magnetic nanoparticles (MNP). It relies on the nonlinear response of MNPs under a magnetic field. The imaging process can be modeled linearly, and then image reconstruction can be case as an inverse problem using a measured system matrix (SM). However, this calibration measurement is time consuming so it reduces practicality. In this study, we proposed a novel method for accelerating the SM calibration based on joint super-resolution (SR) and denoising of sensitivty maps (i.e., rows of SM). The proposed method is based on a deep convolutional neural network (CNN) architecture with residual-dense blocks. Model training was performed using noisy SM measurements simulated for varying MNP size and gradient strengths. Comparisons were performed against conventional low-resolution SM calibration, noisy high-resolution SM calibration, and bicubic upsampling of low-resolution SM. We show that the proposed method improves high-resolution SM recovery, and in turn leads to improved resolution and quality in subsequently reconstructed MPI images.