An exploratory investigation of dual-domain denoising network for field-free-line magnetic particle imaging

Magnetic particle imaging (MPI) is a functional imaging modality that enables highly sensitive tracking of magnetic nanoparticles. Field-free-line (FFL) MPI provides higher signal-to-noise ratio (SNR) than field-free-point MPI, however, noise in the sinogram domain and its propagation during reconstruction can introduce artifacts that degrade image quality.

This study aims to develop a dual-domain denoising framework to improve SNR in both the sinogram and image domains for FFL-MPI tomographic reconstruction.

We propose DudoMTD, a dual-domain cascaded network consisting of two components: (1) a Sinogram Denoising Transformer (SDT) integrates convolutional layers with a Vision Transformer to capture both local and long-range angular dependencies in the sinogram domain; and (2) an edge guiding autoencoder (EGA) operates in the image domain using convolutional filtering and an adaptive Canny operator to preserve structural boundaries. Simulated sinograms were generated using a standard FFL-MPI forward model based on the Langevin magnetization equation and system matrix formulation. The dataset consisted of 13 039 simulated images, with 80% used for training and validation and 20% for testing. In addition, FFL-MPI phantom imaging data were used to evaluate the model under realistic measurement noise. Performance was compared with three benchmarks methods-DuDoNet, RED-CNN, and DnCNN-using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and root mean square error (RMSE). Statistical significance was assessed using the Wilcoxon signed-rank test, with Benjamini-Hochberg correction for multiple comparisons. Effect sizes (Cohen's d) were reported to quantify the magnitude of improvements.

Across four noise conditions (fixed SNRs of 15, 20, 30 dB; 30 dB with additional Poisson noise), DudoMTD showed statistically significant improvements (p < 0.05) over all benchmark methods, with medium-to-large effect sizes. Specifically, the PSNR gains corresponded to Cohen's d values ranging from 0.378 to 1.311, while SSIM improvements yielded d values of 0.323-1.305. Performance gains were pronounced in the 20 dB, 30 dB, and 30 dB + Poisson noise scenarios, where DudoMTD exceeded competing methods by 3%-10% in SSIM and demonstrated consistently superior structural preservation. In phantom experiments acquired on an in-house FFL-MPI system, DudoMTD achieved the highest average SNR (13.809) and effectively suppressed measured noise.

DudoMTD mitigates dual-domain noise in FFL-MPI and improves tomographic image quality across diverse noise conditions. These improvements may facilitate downstream quantitative MPI applications, particularly in low-dose imaging scenarios.