Data augmentation for training a neural network for image reconstruction in MPI

Abstract

Neural networks need to be trained with immense datasets for successful image reconstruction. Acquiring these datasets may be a difficult task, especially in medical imaging. Data augmentation techniques are used to enlarge an available dataset by synthesizing new data. In this work, it is proposed to use the single measurements of a system matrix measurement in magnetic particle imaging for training a neural network for image reconstruction. Before training, mixup augmentation is used to create linear combinations of the single measurements and thus, enlarging the training dataset. Image reconstruction results using neural networks trained with an augmented system matrix are compared to images that have been reconstructed using the conventional system-matrix-based approach.