TSMPI-Net: a time-series generative adversarial network for short-frame-interval magnetic particle imaging.

Objective.Magnetic particle imaging (MPI) is an emerging imaging modality that offers high sensitivity and the potential for high-speed imaging. In many MPI applications, the imaged object moves over time, requiring reconstruction of temporally coherent image sequences at short frame intervals. However, conventional MPI reconstruction methods, such as the system matrix method and X-space method, typically reconstruct frames in isolation and fail to fully exploit the temporal correlations essential for dynamic imaging, resulting in degraded temporal reconstruction quality. Recently developed deep learning-based MPI reconstruction methods address only portions of the reconstruction pipeline, failing to generate images directly from raw signals while overlooking temporal dependencies.Approach.To address this gap, we propose an end-to-end deep-learning framework that directly reconstructs high-quality temporally consecutive MPI image sequences at a short frame interval from one-dimensional voltage signals. The proposed time series MPI Net (TSMPI-Net) integrates temporal and spatial correlation modules into a generative adversarial network. We trained and evaluated the model on a large simulated time-series dataset and an in-house semi-measured dataset. Additionally, we applied the network to experimentally acquired MPI sequences from moving phantoms after uniform signal segmentation.Main results.On the simulated dataset, TSMPI-Net accurately reconstructed time-series MPI image sequences. On the in-house dataset, it outperformed the system-matrix method and several representative deep-learning baselines by more faithfully reconstructing fine-scale spatial distributions of super paramagnetic iron-oxide nanoparticles and improving temporal consistency across frames, yielding improved quantitative metrics. For real phantom data acquired at a 40 ms frame interval, TSMPI-Net enabled robust end-to-end reconstruction of time-series MPI image sequences and improves temporal reconstruction quality compared with the system matrix baseline.Significance.These results validate TSMPI-Net as a practical sequence-aware, end-to-end reconstruction approach for time-series MPI, which enhances overall spatiotemporal reconstruction quality without requiring modifications to MPI hardware or acquisition protocols.