Bridging Synthetic-to-Real Gaps: Frequency-Aware Perturbation and Selection for Single-shot Multi-Parametric Mapping Reconstruction

Yue HuangarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces Frequency-Aware Perturbation and Selection (FPS), a novel unsupervised domain adaptation framework for multi-parametric MRI reconstruction. FPS effectively bridges the gap between synthetic and real-world data by intelligently perturbing and selecting frequency components, leading to more accurate and robust MR parameter maps crucial for clinical diagnosis.

What Problem Does This Paper Address?

Medical imaging, particularly multi-parametric quantitative MRI (mqMRI), has seen significant advancements with AI. However, a major hurdle is the scarcity of real-world, ground-truth data, often necessitating the use of synthetic data. This leads to a "synthetic-to-real gap," where models trained on synthetic data perform poorly on real scans. This gap is especially problematic for low-level tasks like mqMRI reconstruction, where preserving structural integrity and ensuring mapping accuracy are critical for clinical translation. Existing methods struggle with domain adaptation, structural loss, and inefficient computation, hindering their widespread adoption.

Methods

The proposed FPS framework consists of two core components:

  • Wasserstein Distance-modulated Frequency-Aware Perturbation (WDFP): This module addresses the domain gap by adaptively perturbing synthetic and real-world data in the frequency domain. It leverages the 1-Wasserstein distance to identify frequency components that are most susceptible to domain variations. By modulating these components, WDFP encourages the learning of domain-invariant features.
  • Hierarchical Frequency-Aware Selection Network (HFSNet): This acts as the backbone for both the teacher and student models. HFSNet integrates several novel techniques:
  • Feature-Aware Architecture Integration (FAI): Combines CNNs and Swin Transformers to capture multi-scale global and local information.
  • Frequency-Aware Adaptive Selection (FAS) and Compact FAS (cFAS): These mechanisms dynamically select and emphasize informative frequency components, refining optimal solutions within the perturbed data. FAS uses multiple branches with varying filter sizes, while cFAS offers a more lightweight approach.

These components work together in a closed-loop learning pathway, where perturbation activates domain-invariant learning, and selection refines the results.

Datasets

The paper utilizes a combination of datasets for training and evaluation:

  • Numerical brain data: Used for initial testing and comparison.
  • Human brain data:
    • 5 healthy volunteers
    • 94 ischemic stroke patients
    • 46 meningioma patients
  • Synthetic training data: 10,000 pairs of synthetic data were generated, with abnormal templates introduced to expand distributions and mimic real-world variations.
  • Diffusion Tensor Imaging (DTI) data: From the Human Connectome Project WU-Minn-Ox Consortium public database, used to demonstrate the versatility of FPS.

Key Insights and Ideas

  • Frequency Domain is Key: The paper strongly emphasizes the importance of the frequency domain (k-space) for bridging synthetic-to-real gaps in MRI. Low-frequency components are identified as particularly sensitive to domain variations.
  • Perturbation as a Regularizer: Instead of just trying to match distributions, WDFP uses adaptive perturbations to force the model to learn features that are robust to these variations.
  • Selection for Refinement: HFSNet's FAS and cFAS modules act as intelligent filters, ensuring that the most relevant frequency information is used for reconstruction, leading to improved accuracy and reduced artifacts.
  • Synergy of Components: The paper demonstrates that the combination of WDFP and HFSNet (with FAI, FAS, and cFAS) is crucial for achieving superior performance, with ablation studies highlighting the individual contributions of each module.

Key Results

  • Numerical Brain: FPS achieved the best SSIM and MAE, with regression lines closely matching identity lines and superior R² values compared to ResNet, UNet, and DAFrame.
  • Healthy Human Brain: Visually, FPS produced higher structural fidelity and more accurate textures. Bland-Altman plots showed narrower error distributions and fewer outliers, indicating superior robustness.
  • Ischemic Stroke Patients: FPS demonstrated superior performance in classifying acute vs. non-acute patients, achieving the highest AUC (0.87) in ROC curves.
  • Meningioma Patients: FPS accurately captured lesion characteristics and showed distinct ADC distributions for hard and soft meningiomas, supporting clinical assessment.
  • DTI Task: FPS achieved optimal metrics (PSNR, SSIM, NRMSE) across various DTI parameter maps, outperforming DIFFNet and SuperDTI.

Limitations and Future Work

  • Limitations: The paper acknowledges that while FPS significantly bridges the synthetic-to-real gap, challenges remain in areas like fat signal contamination, flow artifacts, and multi-center datasets.
  • Future Work: The authors suggest exploring the feasibility and reliability of FPS in other application areas and addressing more challenging scenarios. Further research will focus on expanding its applicability to diverse imaging modalities and clinical needs.

Summary updated

Abstract

Data-centric artificial intelligence (AI) has remarkably advanced medical imaging, with emerging methods using synthetic data to address data scarcity while introducing synthetic-to-real gaps. Unsupervised domain adaptation (UDA) shows promise in ground truth-scarce tasks, but its application in reconstruction remains underexplored. Although multiple overlapping-echo detachment (MOLED) achieves ultra-fast multi-parametric reconstruction, extending its application to various clinical scenarios, the quality suffers from deficiency in mitigating the domain gap, difficulty in maintaining structural integrity, and inadequacy in ensuring mapping accuracy. To resolve these issues, we proposed frequency-aware perturbation and selection (FPS), comprising Wasserstein distance-modulated frequency-aware perturbation (WDFP) and hierarchical frequency-aware selection network (HFSNet), which integrates frequency-aware adaptive selection (FAS), compact FAS (cFAS) and feature-aware architecture integration (FAI). Specifically, perturbation activates domain-invariant feature learning within uncertainty, while selection refines optimal solutions within perturbation, establishing a robust and closed-loop learning pathway. Extensive experiments on synthetic data, along with diverse real clinical cases from 5 healthy volunteers, 94 ischemic stroke patients, and 46 meningioma patients, demonstrate the superiority and clinical applicability of FPS. Furthermore, FPS is applied to diffusion tensor imaging (DTI), underscoring its versatility and potential for broader medical applications. The code is available at https://github.com/flyannie/FPS.
BibTeX
@misc{huang2025bridging,
  author = {Yue Huang},
  title = {{Bridging Synthetic-to-Real Gaps: Frequency-Aware Perturbation and Selection for Single-shot Multi-Parametric Mapping Reconstruction}},
  year = {2025},
  eprint = {2503.03475},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.03475},
}
Figure