A Large Dataset for Model-Based Image Reconstruction and Operator Correction in MPI

Abstract

Within the context of MPI, the accuracy of model-based system matrices has witnessed remarkable recent improvements. However, model-based reconstruction approaches still require substantial enhancement to compete with measured system matrix approaches, particularly including multi-dimensional Lissajous-type trajectories of the field-free point. The latter approach has demonstrated its ability to produce satisfactory reconstructions under the application of standard techniques. This makes it one of the preferable state-of-the-art methods for image reconstruction in MPI, despite its extended calibration time. Determining the source of reconstruction artifacts turns out as a challenging aspect for model-based reconstruction approaches. It is not straightforward to distinguish between errors caused by background noise and errors resulting from inaccurate assumptions on the chosen physical model, leading to deviations of the forward operator. We provide a dataset tailored for operator correction in the context of model-based reconstruction approaches in MPI, where we use most recent particle magnetization models for the simulation of system matrices. The integration of precisely calibrated components from the Bruker MPI system contributes to the enhancement of model accuracy. Our dataset comprises measurements from simulated system matrices, combined with corresponding ground truth phantoms. It covers a variation of physical models and diverse noise scenarios, using realistic noise data captured by the Bruker MPI system. Accessing the set of data tuples facilitates an analysis of model deviations, potentially serving as valuable prior information for model corrections. Moreover, we provide validations of standard reconstruction methods when applied to our model-based simulations of the system matrix.