1. Dataset Introduction

This dataset contains system matrix and phantom data acquired using a self-developed device. The device is based on a field-free line (FFL) magnetic particle imaging (MPI) system. Detailed information about the device can be found in:

Zechen Wei et al., A compact low-power magnetic particle imaging scanner based on a permanent-magnet field-free-line generator with high gradient, Review of Scientific Instruments, 2026, 97 (6).

The data were acquired using the FFL projection imaging method described in the above article (see Fig. 5). During acquisition, the FFL was scanned along the y direction, while the phantom was mechanically displaced along the x-axis to achieve two-dimensional scanning in the x-y plane. The basic scanning parameters are listed below.

1) System Matrix Data

The system matrix data are one-dimensional system matrix signals acquired along the y direction. The scanning field of view (FOV) is ±9 mm with a step size of 0.5 mm. The corresponding files are SM_signal.mat and SM_background.mat. Each data file contains a matrix with dimensions of 2,500,000 × N, where N represents the number of discrete spatial points within the FOV. Therefore, each 2,500,000-dimensional signal corresponds to the raw time-domain signal measured at one discrete spatial position within the FOV.

The SM_signal.mat file stores particle signals from the delta phantom at different discrete positions within the FOV (with background subtraction). The SM_background.mat file stores the corresponding background signals (without particle signals). The two signals are matched one-to-one along the N dimension.

2) Phantom Data

The phantom dataset contains particle signals from three phantoms: C, A, and S, corresponding to the files Phantom_C.mat, Phantom_A.mat, and Phantom_S.mat. Using the scanning method described above, each phantom was scanned line-by-line along the y direction using one-dimensional FFL projection scanning. The phantom was moved along the x direction using a translation stage to achieve two-dimensional scanning. The one-dimensional scanning FOV is consistent with the system matrix scanning range (parameters shown above), while the x-direction scanning range is ±8 mm with a step size of 0.5 mm.

Each phantom data file contains a matrix with dimensions of 2,500,000 × M, where M represents the number of scanned lines along the x direction. Therefore, each 2,500,000-dimensional signal corresponds to the raw time-domain signal acquired from one row of the phantom.

2. Data Usage

1) Basic Reconstruction Workflow

• Fourier Transform. The original signals are time-domain signals. Fourier transform should first be applied to both the system matrix signals and phantom signals to obtain the corresponding frequency-domain signals.
• Frequency Selection. Since particle information mainly exists in harmonic frequencies, frequency selection can be performed on the frequency-domain signals (e.g., selecting the 2^nd-10^th harmonics for reconstruction). Alternatively, based on the provided system matrix background signals, frequency components with SNR above a certain threshold can be calculated and selected for reconstruction.
• Row-wise Reconstruction and Generation of 2D Images. Each row signal of the phantom (2,500,000-dimensional) is reconstructed to obtain a one-dimensional image at position x=xi. The final two-dimensional image is obtained by stacking the reconstructed M rows.

3. Dataset Applications

This dataset can be used for:

1) Validation of MPI Image Reconstruction Algorithms. Evaluation of different reconstruction algorithms, including Kaczmarz, Tikhonov regularization, deep learning-based reconstruction, etc.

2) System Matrix Processing Research. Frequency-domain analysis, frequency selection, system matrix compression, system matrix quality improvement and super resolution, etc.

3) Signal Enhancement and Denoising Experiments. Denoising of raw MPI time-domain signals, signal enhancement, deep learning-based signal recovery, etc.

Citation:

Zechen Wei et al., Harmonic Autoencoding Framework for Multiple Tasks in Magnetic Particle Imaging Reconstruction, IEEE Transactions on Neural Networks and Learning Systems, 2026.