A CNN Based Super-Resolution Technique for Magnetic Particle Imaging System Matrix

Magnetic Particle Imaging (MPI) is a new imaging technique that allows high resolution & high frame-rate imaging of super-paramagnetic iron-oxide (SPIOs) particles. The imaging process can be modeled linearly. However, due to non-idealities and variations of the experimental system in time, many MPI systems first measure the forward model matrix (i.e. calibrate the system), then reconstruct frames from the measurements using these matrices. The image resolution and size is directly affected by the size of the system matrix. However, the calibration process may consume a lot of time depending on the field of view and resolution. In this study, we propose applying superresolution techniques on measured low-resolution system matrices to get to high resolution system matrices. Here, we propose a convolutional neural network (CNN) based super-resolution technique tailored for MPI, and show its effectiveness against linear interpolation. The methods are compared in a noiseless simulation environment, and for a super-resolution factor of 4×4, our proposed technique resulted in %2.92 normalized Root Mean Squared Error (nRMSE), while bicubic interpolation techniques resulted in %12.47 nRMSE.