Image reconstruction for a shift variant magnetic particle imaging system using deep learning

Magnetic Particle Imaging (MPI) is currently a preclinical tracer-based biomedical imaging modality, which detects superparamagnetic nanoparticles to form an image. To date, many different scanner designs have been built with different image reconstruction techniques such as system function or projection imaging. Our group is developing a single-sided scanner, which allows for imaging of arbitrarily large subjects, with finite imaging depth, making it promising for cancer diagnostics. One major challenge with single-sided topology comes from the inhomogeneous magnetic fields used for the encoding gradient and the excitation, which make the system inherently shift variant. This leads to difficulties in image reconstruction because standard techniques require the system to be shift invariant. To address this challenge, our group has implemented a convolutional neural network (CNN) that has learned to correct features in the measured signal that originate from the inhomogeneous magnetic fields of our system. We use U-Net, a CNN initially developed for biomedical image segmentation, and we train it on custom MATLAB simulations of randomly generated MPI targets. We show that images reconstructed using our network have increased field of view, improved spatial resolution, and reduced imaging artifacts that result from inhomogeneity in our magnetic fields, making this approach promising for image reconstruction in shift variant imaging modalities.