Multi-purpose deep learning framework for MPI based on contrastive learning

Abstract

Deep learning can be used in many tasks for MPI, prominently to reduce the calibration time of system matrix (SM)-based reconstruction by recovering undersampled SMs or directly reconstructing measurements without an SM. The success of supervised machine learning methods depends on the used training data, which should have high quality, match the distribution of the desired test cases and be cleanly labeled. For MPI, such data rarely exists. To find robust features in complex input data, the unsupervised method contrastive learning can be used. In this work, we show its applicability to MPI voltage signals and that the learned features improve the performance of tasks like SM recovery and direct reconstruction of real MPI data in 2D. SM recovery is performed by predicting voltage signals of samples placed in the MPI FOV, which could also provide an alternative to classic simulation frameworks that cannot match real MPI measurements.