Abstract
Magnetic Particle Imaging (MPI) is a high-contrast, radiation-free medical imaging modality that relies on a System Matrix (SM) for image reconstruction. However, the acquisition of a high-resolution SM is a significant bottleneck, often requiring hours of calibration. This paper presents the Position-Prior-Guided Network (PPGnet), a deep learning framework that incorporates spatial position information as a prior to enhance SM super-resolution. By leveraging the inherent mathematical symmetries of magnetic fields, PPGnet achieves significant improvements in reconstruction accuracy, reducing Normalized Root Mean Square Error (NRMSE) by approximately 33% compared to existing state-of-the-art methods.
Introduction
Magnetic Particle Imaging (MPI) offers high-contrast imaging of magnetic nanoparticles without ionizing radiation. The reconstruction process typically utilizes a System Matrix (SM), which maps the spatial distribution of nanoparticles to the received signal. Traditionally, constructing this matrix involves a point-by-point calibration process that can take upwards of 15 hours for even small volumes. While super-resolution (SR) techniques have been proposed to estimate high-resolution SMs from sparse measurements, many existing models fail to account for the underlying physics of the magnetic field, leading to artifacts and reduced fidelity.
Methods
The proposed Position-Prior-Guided Network (PPGnet) integrates physical priors directly into the neural network architecture to exploit the structured nature of the SM.
The architecture consists of three primary components:
- Position Information Embedding: The network receives voxel coordinates ($x, y, z$) as additional input channels. These coordinates are encoded to reflect the odd and even symmetries characteristic of magnetic field-free point (FFP) trajectories.
- RRDBnet Backbone: PPGnet utilizes Residual Dense Blocks (RDBs) with 3D convolutions. This structure allows the model to capture complex volumetric features while maintaining low-level signal details often lost in deep networks.
- Symmetry-Aware Upsampling: The model employs linear interpolation followed by 3D convolutional layers to reconstruct high-resolution SMs from undersampled data (e.g., 2x and 4x ratios).
The theoretical foundation of PPGnet lies in the observation that SM components can be modeled using Chebyshev polynomials, making spatial coordinates a powerful prior for exploiting field symmetries.
Results and Discussion
The performance of PPGnet was evaluated using the OpenMPI dataset, specifically testing its ability to generalize across different nanoparticle types (Synomag-D for training, Perimag for testing).
As shown in the quantitative results, PPGnet significantly outperforms existing models:
- 3D Reconstruction: At a 2x downsampling ratio, PPGnet achieved a 3.44% NRMSE, compared to 5.14% for 3dSMRnet.
- 2D Reconstruction: PPGnet reached an NRMSE of 3.14%, surpassing the Transformer-based TranSMS (4.08%).
- Visual Fidelity: Qualitative analysis via error maps indicated that PPGnet preserves the structural integrity of the system response better than competing methods, which often exhibit distortions at higher undersampling rates (4x).
- Phantom Imaging: Reconstructed SMs were used to image physical phantoms, where PPGnet-derived matrices yielded higher PSNR and SSIM scores, indicating superior image quality in practical applications.
Conclusion
PPGnet demonstrates that embedding physical position priors into deep learning models can drastically improve the efficiency of MPI calibration. By exploiting the mathematical symmetries of the imaging system, the network achieves state-of-the-art accuracy and robust generalization across different tracers. Future work will focus on improving robustness against real-world hardware imperfections and noise to move closer to instantaneous MPI calibration.
Summary updated