TL;DR
This paper introduces a novel MRI reconstruction method that leverages the strengths of both diffusion models and model-driven networks. By using preliminary, albeit noisy, samples from a diffusion model as guidance for a model-driven network, they achieve high-quality MRI reconstructions with reduced artifacts and improved robustness, even with limited training data.
What Problem Does This Paper Address?
Magnetic Resonance Imaging (MRI) is a crucial diagnostic tool, but its long scanning times can lead to patient discomfort and motion artifacts, degrading image quality. Undersampling k-space data can speed up scans but introduces aliasing artifacts. While deep learning methods have shown promise in MRI reconstruction, they often require large datasets and can lack interpretability. Diffusion models, like those based on score matching with Langevin dynamics (SMLD), have emerged as powerful generative tools for MRI reconstruction. However, SMLD methods are sensitive to hyperparameter tuning and can produce severe hallucination artifacts, especially with out-of-distribution data. This paper aims to bridge this gap by developing a robust MRI reconstruction framework that effectively utilizes the generative capabilities of diffusion models while mitigating their artifact-prone nature through a guided model-driven approach.
Methods
The proposed method, termed SGM-Net, consists of three main steps:
- Sampling: Instead of meticulously tuning SMLD hyperparameters or retraining score networks, the authors use a pretrained score network (PSN) to generate preliminary guidance images (PGIs) from undersampled k-space data. This bypasses the need for in-distribution training data and extensive tuning.
- Denoising: Recognizing that PGIs can be noisy and contain artifacts, a denoising module (DM) is introduced. This module employs a score-based information extractor (SIE) and a cross-domain information extractor (CIE) to remove artifacts and noise from the PGIs, producing denoised guidance images (DGIs). The CIE specifically leverages the complementarity between the frequency and image domains.
- Guided Reconstruction: A densely connected unrolled network is designed to reconstruct the MR images. This network takes both the undersampled measurements and the DGIs as input. The DGIs are periodically updated to provide increasingly accurate guidance throughout the reconstruction process, enabling a coarse-to-fine recovery of image details.
The entire pipeline is trained end-to-end.
Datasets
The authors used the following datasets for training and evaluation:
- NYU knee dataset: A publicly available 15-coil dataset used for training and comparison.
- FastMRI dataset: A publicly available dataset for accelerated MRI, used for evaluation.
No new benchmarks were introduced.
Key Insights and Ideas
- Leveraging "noisy" priors: The core idea is that even imperfect samples from diffusion models (PGIs) contain valuable information that can guide a model-driven network, rather than being discarded due to artifacts.
- Complementarity of domains: The CIE module effectively exploits the relationship between the k-space (frequency) and image domains to extract noise patterns, enhancing the denoising process.
- Coarse-to-fine reconstruction: The multi-stage approach, starting with denoising PGIs and then refining with guided reconstruction, allows for effective recovery of both low-frequency patterns and fine details.
- Robustness to sampling steps: The method demonstrates robustness to variations in sampling steps, suggesting that the guidance mechanism can compensate for lower-quality initial samples.
Key Results
- The proposed SGM-Net achieved state-of-the-art results on both the NYU knee and FastMRI datasets, outperforming traditional and several DL-based methods in terms of PSNR and SSIM metrics.
- Quantitative results in Table 1 show SGM-Net consistently achieving higher PSNR and SSIM values across different acceleration factors (4x and 6x) and sequences (Coronal PD, PDFS, FastMRI). For instance, on the FastMRI dataset at 6x acceleration, SGM-Net achieved a PSNR of 36.87±3.03 and SSIM of 0.9094±0.0707.
- Ablation studies (Table 2) demonstrated that combining the SMLD guidance (SG) with the model-driven network (MN) and denoising module (DM) was crucial for achieving optimal performance. Simply combining SG and MN was worse than MN alone.
- The method showed remarkable robustness to limited training data, maintaining stable performance even with significantly reduced training slices, unlike other methods that showed considerable deterioration.
Limitations and Future Work
- Training data requirement: While the workflow is robust to noisy SMLD inputs, the unrolled network still requires some training data to achieve superior performance.
- Future work: The authors suggest exploring self-supervising strategies to further optimize the workflow and potentially reduce the need for labeled training data. They also mention that the workflow is compatible with optimized SMLD steps, suggesting potential for future improvements by integrating higher-quality references.
Summary updated