Deconvolution of direct reconstructions for MPI using Convolutional Neural Network

Abstract

Recently, an approach was presented that allows a direct and fast image reconstruction without the use of a system
matrix for Lissajous trajectories of the so called field free point. The method is based on weighting frequency
components of the measured voltage signals and additional factors with Chebychev polynomials of the second
kind, resulting in reconstructions of the convolved spatial distribution of magnetic nanoparticles. In order to
obtain meaningful images, these reconstructions have to be deconvolved afterwards. For this purpose, different
methods have already been proposed. In this work, a U-shaped neural network is used for the deconvolution. The
network was trained and tested on simulated data of blood vessel like structures. The proposed model outperforms
conventional methods and improves the image quality of the reconstructions.