Multiscale frequency attention transformer for resolution enhancement in magnetic particle imaging.

BACKGROUND

Magnetic particle imaging (MPI) is an emerging functional imaging modality that enables high-resolution (HR) visualization of superparamagnetic iron oxide nanoparticles (SPIONs). It offers significant advantages, including high penetration capability, absence of ionizing radiation, high contrast, and exceptional temporal resolution and sensitivity, making it highly promising for a broad spectrum of biomedical and clinical applications.

PURPOSE: Although MPI holds great promise, fundamental spatial resolution is inherently constrained by the gradient field strength and the saturation magnetization physics of the SPIONs. To address the image blurring characterized by the system's point spread function (PSF) and the dependency on high-gradient fields for HR imaging, this study aims to develop a robust deep learning framework to reduce reliance on high gradient hardware. We aim to achieve high-fidelity resolution enhancement computationally, recovering fine structural details from low-resolution inputs, thereby bypassing the need for costly hardware upgrades.

METHODS

We propose a Multiscale Frequency Attention Transformer (MFAT) model, which integrates a frequency-domain discriminative feedforward network (DFFN) with a multiscale attention block (MAB) to form an end-to-end framework for recovering HR MPI images from low-resolution reconstructions. The MFAT architecture combines multiscale feature aggregation with a novel frequency attention mechanism, allowing it to capture both global context and fine local details across multiple scales while enhancing fidelity and sharpness in the frequency domain.

RESULTS

Extensive evaluations using both simulated and real-world MPI data demonstrate that the proposed MFAT model outperforms existing state-of-the-art deep learning methods, achieving an improvement of 7.51% in peak signal-to-noise ratio (PSNR) in the MNIST dataset, corresponding to a gain of 2.48 dB over the second strongest baseline. Furthermore, in terms of spatial resolution, MFAT significantly reduced the full width at half maximum (FWHM) by 17.86% on the complex FIVES vascular structures compared to the second-performing baseline, verifying its superior capability in recovering high-frequency details.

CONCLUSIONS

Thus, MFAT provides an effective solution for enhancing spatial resolution in MPI image reconstruction and significantly improves image quality. This advancement holds strong potential to facilitate the translation of MPI technology into more refined clinical applications.