Deep learning-enhanced X-space reconstruction for magnetic particle imaging: a physics-consistent approach.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Artificial Intelligence

Background:

• Magnetic Particle Imaging (MPI) is a sensitive, high-temporal-resolution imaging technique.
• X-space reconstruction in MPI is limited by point spread function (PSF) blur, impacting spatial resolution.

Purpose of the Study:

• To enhance X-space MPI reconstruction quality.
• To preserve physical consistency during reconstruction.
• To reduce blur and improve spatial resolution in MPI images.

Main Methods:

• Developed X-Space-PC-Restore, a physics-consistent deep learning framework.
• Utilized a hybrid U-Net encoder-decoder architecture with Transformer attention.
• Incorporated PSF-guided loss functions for training.

Main Results:

• Achieved a PSNR of 15.48 dB and NRMSE of 0.175, a 23.0% PSNR improvement and 45.8% NRMSE reduction over Richardson-Lucy.
• Demonstrated consistent superiority across SNR levels (5-40 dB).
• Achieved near-ground-truth resolution, outperforming classical methods in blur reduction and noise robustness.

Conclusions:

• Physics-informed deep learning offers a promising strategy for MPI image reconstruction.
• The proposed method effectively reduces blur, enhances image fidelity, and improves reliability.
• X-Space-PC-Restore represents a significant advancement in MPI image processing.