Conditional Denoising Diffusion Model-Based Robust MR Image Reconstruction from Highly Undersampled Data

TL;DR

This paper tackles the challenge of robust MRI reconstruction from highly undersampled data by introducing a conditional denoising diffusion model with enforced data consistency. The approach conditions the diffusion process on undersampled inputs and injects a fidelity term into every reverse-diffusion step, thereby tightly coupling the learned prior with the MRI measurement physics. Trained on paired undersampled and ground-truth data, the method achieves state-of-the-art performance on fastMRI in terms of SSIM, PSNR, and LPIPS, while exhibiting strong generalization across acceleration factors. The work demonstrates a practical pathway to accelerated, reliable MRI reconstruction with potential for integration into clinical workflows and real-time imaging scenarios.

Abstract

Magnetic Resonance Imaging (MRI) is a critical tool in modern medical diagnostics, yet its prolonged acquisition time remains a critical limitation, especially in time-sensitive clinical scenarios. While undersampling strategies can accelerate image acquisition, they often result in image artifacts and degraded quality. Recent diffusion models have shown promise for reconstructing high-fidelity images from undersampled data by learning powerful image priors; however, most existing approaches either (i) rely on unsupervised score functions without paired supervision or (ii) apply data consistency only as a post-processing step. In this work, we introduce a conditional denoising diffusion framework with iterative data-consistency correction, which differs from prior methods by embedding the measurement model directly into every reverse diffusion step and training the model on paired undersampled-ground truth data. This hybrid design bridges generative flexibility with explicit enforcement of MRI physics. Experiments on the fastMRI dataset demonstrate that our framework consistently outperforms recent state-of-the-art deep learning and diffusion-based methods in SSIM, PSNR, and LPIPS, with LPIPS capturing perceptual improvements more faithfully. These results demonstrate that integrating conditional supervision with iterative consistency updates yields substantial improvements in both pixel-level fidelity and perceptual realism, establishing a principled and practical advance toward robust, accelerated MRI reconstruction.

Figures (6)

• Figure 1: The forward process (left to right) and the reverse inference process (right to left) in conditional denoising diffusion models.
• Figure 2: Flowchart of training process and sampling process of conditional denoising diffusion model with enforced data consistency.
• Figure 3: MRI reconstructions with an acceleration factor of x = 8 on the same subject using models: UNet, self-attention UNet (SA-UNet), conditional denoising diffusion (Diffusion), the proposed model (DiffDC), and sampling patterns (Sampling-Ptn): Gaussian 1D (G1D), Gaussian 2D (G2D), Uniform 1D, and Poisson.
• Figure 4: MRI reconstructions with an acceleration factor of x = 8 on four different subjects using models: UNet, self-attention UNet (SA-UNet), conditional denoising diffusion (Diffusion), the proposed model (DiffDC), and sampling patterns (Sampling-Ptn): Gaussian 1D (G1D), Gaussian 2D (G2D), Uniform 1D, and Poisson.
• Figure 5: MRI reconstructions with an acceleration factor of x = 4 on the same subjects by applying the proposed model (DiffDC) trained at an acceleration factor of x=8, and different sampling patterns (Sampling-Ptn): Gaussian 1D (G1D), Gaussian 2D (G2D), Uniform 1D, and Poisson.