Magnetic Particle Imaging (MPI) produces voltage signals that are periodic and highly correlated across time. Classical deep-learning reconstruction methods generally treat these signals as static feature vectors. In this work, we explore a purely data-driven sequence-to-image approach using a bidirectional long short-term memory (BiLSTM) network followed by a lightweight convolutional neural network (CNN). The model directly maps 1D MPI-signals to 2D images. We evaluate performance using a synthetically generated, trajectory-encoded augmented MNIST dataset, and analyse robustness under different signal-to-noise ratios (SNR). Results demonstrate that the BiLSTM architecture successfully reconstructs image structure and remains robust even at low SNR levels, highlighting the potential of sequence modelling for future MPI reconstruction.
Gladiß et al. (Sun,) studied this question.