Synergistic Spatial-Frequency Feature Guided Transformer Network for Denoising X-Space Reconstruction in Magnetic...

Magnetic Particle Imaging (MPI), a non-invasive imaging technique, leverages the detection of superparamagnetic nanoparticles to visualize biological processes with high sensitivity and temporal-spatial resolution. However, low tracer concentrations can diminish the signal-to-noise ratio of the magnetization signal, resulting in significant noise artifacts in MPI images reconstructed using the X-space method. Hardware improvements have high complexity, while traditional methods lack robustness to different noise levels, making it difficult to improve the quality of low concentration MPI images. In this paper, we introduced a synergistic spatial-frequency feature-guided approach for MPI image denoising and quality enhancement, based on a sparse lightweight transformer model. Specifically, the Fast Fourier Adjustment (FFA) module employs the Fast Fourier Transform and a series of convolutional operations to generate frequency-domain feature maps, which co-guide the network reconstruction with spatial-domain features. The Spatial and Frequency Guidance (SAG) module integrates pixel-level and implicit frequency detail features by learning dual-domain spatial-frequency representations. It consists of a Layer Normalization (LN), a Spatial and Frequency Fusion Attention (SFA) block, and a Multilayer Perceptron (MLP) module. Simulation experiments demonstrate that our method outperforms existing deep learning techniques in MPI image denoising and generates high-quality MPI images even at low concentrations.