FgC2F-UDiff: Frequency-guided and Coarse-to-fine Unified Diffusion Model for Multi-modality Missing MRI Synthesis

Xiaojiao Xiao, Qinmin Vivian Hu, Guanghui Wang

Abstract

FgC2F-UDiff is a unified diffusion-based framework designed to synthesize missing MRI modalities (such as T2 or FLAIR) from any available combination of existing scans. By guiding the denoising process with frequency-domain information and a coarse-to-fine strategy, it achieves state-of-the-art image quality while significantly reducing the inference time typically associated with diffusion models.

Introduction

In clinical neurology, multi-modality MRI is the gold standard for diagnosis and treatment planning. Different sequences—T1, T2, FLAIR, and T1 contrast-enhanced (T1ce)—provide complementary views of brain anatomy and pathology. However, it is common for one or more of these sequences to be missing due to scan time constraints, patient motion, or contrast agent intolerance.

Multi-modality MRI sequences (T1, FLAIR, T2, T1ce) and their corresponding intensity histograms.

While researchers have previously utilized Generative Adversarial Networks (GANs) to "fill in the blanks," these models often suffer from training instability and "mode collapse," failing to capture the full diversity of medical images. Denoising Diffusion Probabilistic Models (DDPMs) offer a more stable alternative but are notoriously slow and usually designed for specific one-to-one mappings. FgC2F-UDiff addresses these limitations by creating a unified model capable of handling any combination of inputs and outputs (many-to-many) efficiently and accurately.

Methodology

The FgC2F-UDiff architecture rests on three primary pillars:

  1. Coarse-to-fine Unified Network (CUN): This exploits the natural iterative property of diffusion models. The denoising process is split into two stages: a "coarse" stage focusing on global anatomical structures and a "fine" stage that recovers intricate textures and edges.
  2. Frequency-guided Collaborative Strategy (FCS): The model uses Gaussian filters to decompose images into low-frequency (LF) and high-frequency (HF) components. During the coarse stage, LF information provides a "structural roadmap." In the fine stage, HF information guides the model to synthesize sharp details.
  3. Specific-acceleration Hybrid Mechanism (SHM): To ensure clinical viability, the authors implemented Curriculum Learning (CL)—training the model on simpler tasks (one missing modality) before moving to complex ones (three missing modalities). They also optimized the sampling process to use only 200 timesteps, reducing inference time to approximately 1.5–2 seconds per image.

Key Results and Insights

The model was evaluated on the BraTS 2021 (1,251 brain tumor cases) and IXI (600 healthy subjects) datasets. FgC2F-UDiff outperformed several state-of-the-art methods, including pix2pix, pGAN, and standard Latent Diffusion Models (LDM).

  • Quantitative Performance: On the BraTS dataset, the model achieved a PSNR of 29.43 dB and an SSIM of 0.951 for many-to-many synthesis, significantly surpassing GAN-based competitors.
  • Visual Fidelity: Qualitative results demonstrate superior delineation of tumor boundaries and edema areas—critical regions where GANs often produce blurry or hallucinated artifacts.
  • Efficiency: The model achieved a 4x to 5x speedup in inference compared to standard LDM approaches, bringing generation time down to a few seconds per slice.

Conclusion and Future Work

FgC2F-UDiff demonstrates that aligning frequency-domain priors with the diffusion timeline acts as a powerful stabilizer for medical image synthesis. Its unified nature makes it highly practical for real-world hospital settings where the available scans vary per patient. Future research will focus on extending the framework to 3D convolutions for better volumetric consistency and testing its generalization on other organs, such as the liver or heart.

Summary updated

BibTeX
@article{xiao2024fgc,
  author = {Xiaojiao Xiao and Qinmin Vivian Hu and Guanghui Wang},
  title = {{FgC2F-UDiff: Frequency-guided and Coarse-to-fine Unified Diffusion Model for Multi-modality Missing MRI Synthesis}},
  journal = {IEEE Trans. Computational Imaging},
  year = {2024},
  doi = {10.1109/tci.2024.3516574},
}
Figure