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

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Magnetic Particle Imaging (MPI) is an emerging high-sensitivity imaging technique.
• Dynamic MPI applications require reconstruction of temporally coherent image sequences.
• Conventional methods struggle with temporal correlations, degrading dynamic imaging quality.

Purpose of the Study:

• To develop an end-to-end deep learning framework for reconstructing time-series MPI image sequences.
• To address limitations of existing methods in exploiting temporal correlations for dynamic MPI.
• To improve spatiotemporal reconstruction quality in dynamic MPI imaging.

Main Methods:

• Proposed Time Series Magnetic Particle Imaging Net (TSMPI-Net), an end-to-end deep learning framework.
• Integrated temporal and spatial correlation modules into a generative adversarial network (GAN).
• Trained and evaluated on simulated and in-house datasets, including experimental data from moving phantoms.

Main Results:

• TSMPI-Net accurately reconstructed time-series MPI image sequences from raw 1D voltage signals.
• Outperformed system matrix and deep learning baselines in reconstructing nanoparticle distributions and temporal consistency.
• Achieved robust end-to-end reconstruction and improved temporal quality for real MPI data.

Conclusions:

• TSMPI-Net is a practical, sequence-aware, end-to-end reconstruction approach for time-series MPI.
• Enhances spatiotemporal reconstruction quality without altering MPI hardware or protocols.
• Validates the potential of deep learning for advanced dynamic MPI reconstruction.