FSC-loss: A Frequency-domain Structure Consistency Learning Approach for Signal Data Recovery and Reconstruction

Liwen Zhang, Zhaoji Miao, Fan Yang, Gen Shi, Jie He, Yu An, Hui Hui, Jie TianarXiv 2025 · DOI · arXiv

Introduction

In biomedical imaging, Magnetic Particle Imaging (MPI) is gaining prominence for its high sensitivity and lack of ionizing radiation. However, the modality faces a significant bottleneck: the acquisition of the Signal Matrix (SM). The SM describes the response of magnetic nanoparticles at every point in the scanner, and measuring a high-resolution version can take upwards of 32 hours for a standard 3D scan.

While researchers have explored "Super-Resolution" (SR) to predict high-resolution matrices from sparse, low-resolution samples (reducing time to roughly 30 minutes), standard algorithms like Bicubic interpolation or traditional CNNs often fail. These methods tend to smooth out high-frequency signals, resulting in blurry reconstructions that lose critical structural details. To address this, researchers have developed FSC-loss, a deep-learning approach that reconstructs high-resolution signal data up to 60 times faster than physical measurements while preserving sharp, high-frequency details.

Methodology

The proposed framework utilizes a three-pronged strategy to bridge the gap between acquisition speed and image resolution:

  1. RIM-embedding: Instead of using raw complex numbers, the network receives a three-channel input consisting of the Real part, Imaginary part, and Magnitude. The magnitude serves as a "physical prior," helping the model understand signal strength and energy distribution.
  2. FSC-loss (Frequency-domain Structure Consistency): Traditional loss functions like Mean Squared Error ($L_2$) focus on pixel-wise differences, which often leads to blurring. FSC-loss adapts the Structural Similarity Index (SSIM) to focus on local structural patterns specifically in the frequency domain, ensuring robustness against extreme signal values.
  3. SMSW Transformer: The architecture employs a self-adaptive multi-scale shifted window Transformer. Unlike standard CNNs with limited receptive fields, this Transformer uses attention mechanisms to capture both global context and fine-grained local details across the signal matrix.
Overview of the FSC-loss framework, including SM measurement and the learning function.

Experimental Setup

The study validated the FSC-loss approach across several data sources:

  • Simulation Datasets: Created using Lissajous trajectories to test theoretical limits under various noise conditions.
  • Open MPI Data: A public benchmark utilizing Perimag and Synomag-D tracers for comparison against state-of-the-art (SOTA) methods.
  • In-house Systems: Deployment on three custom MPI scanners (Single-angle, Streamlined, and Sequential) to demonstrate real-world utility.

Key Results

The FSC-loss approach demonstrated significant improvements in both numerical accuracy and visual fidelity:

  • Superior Accuracy: On the Open MPI benchmark, the method achieved a normalized Root Mean Square Error (nRMSE) of 0.041, outperforming previous SOTA methods like TranSMS.
  • High-Frequency Recovery: Visual error maps indicate that while competing methods struggle at higher frequencies (e.g., 379 kHz), FSC-loss maintains sharp structural integrity and low error rates.
  • Generalization: The model showed strong cross-tracer performance, successfully recovering data for Perimag tracers even when trained on Synomag-D data.
  • Efficiency: The model can recover a high-resolution SM from a 16x down-sampled input in less than 15 seconds, effectively reducing a day-long process to a matter of seconds.
Visual comparison of signal recovery across different frequency channels and methods.

Conclusion

FSC-loss represents a major step forward in MPI signal processing by solving the trade-off between acquisition speed and data quality. By focusing on frequency-domain structural consistency and leveraging Transformer-based architectures, the method enables high-resolution imaging without the prohibitive time costs of traditional calibration. Future work will focus on higher down-sampling factors and integrating this recovery process into real-time dynamic imaging workflows.

Summary updated

BibTeX
@misc{zhang2025fsc,
  author = {Liwen Zhang and Zhaoji Miao and Fan Yang and Gen Shi and Jie He and Yu An and Hui Hui and Jie Tian},
  title = {{FSC-loss: A Frequency-domain Structure Consistency Learning Approach for Signal Data Recovery and Reconstruction}},
  year = {2025},
  eprint = {2501.04308},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.04308},
}
Figure