Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution

TL;DR

This work tackles the inefficiency and distortion issues of diffusion-model-based MRI super-resolution by introducing DiffMSR, which applies a diffusion process in a highly compact latent space to generate a latent prior $Z \in \mathbb{R}^{4 \hat{C}}$ guiding a novel PLWformer decoder. The training is split into Prior Extraction (PE) to obtain a reliable latent prior and diffusion-based refinement to produce $\hat{Z}$, with inference using a conditional latent from the LR input and a reference HR image to reconstruct high-quality SR images. DiffMSR achieves state-of-the-art PSNR/SSIM on four datasets (public FastMRI Knee and clinical Brain, Tumor, Pelvic) with lower FLOPs and faster inference, thanks to the latent-space diffusion and a large-window transformer that leverage prior knowledge without excessive computation. The approach demonstrates substantial practical impact for fast, accurate multi-contrast MRI SR, enabling high-detail reconstruction with reduced measurement times and improved diagnostic fidelity.

Abstract

Recently, diffusion models (DM) have been applied in magnetic resonance imaging (MRI) super-resolution (SR) reconstruction, exhibiting impressive performance, especially with regard to detailed reconstruction. However, the current DM-based SR reconstruction methods still face the following issues: (1) They require a large number of iterations to reconstruct the final image, which is inefficient and consumes a significant amount of computational resources. (2) The results reconstructed by these methods are often misaligned with the real high-resolution images, leading to remarkable distortion in the reconstructed MR images. To address the aforementioned issues, we propose an efficient diffusion model for multi-contrast MRI SR, named as DiffMSR. Specifically, we apply DM in a highly compact low-dimensional latent space to generate prior knowledge with high-frequency detail information. The highly compact latent space ensures that DM requires only a few simple iterations to produce accurate prior knowledge. In addition, we design the Prior-Guide Large Window Transformer (PLWformer) as the decoder for DM, which can extend the receptive field while fully utilizing the prior knowledge generated by DM to ensure that the reconstructed MR image remains undistorted. Extensive experiments on public and clinical datasets demonstrate that our DiffMSR outperforms state-of-the-art methods.

Figures (12)

• Figure 1: Comparison with DM-based MRI reconstruction methods on FastMRI dataset. Note that the experimental settings used by these methods are the same as those in Sec. \ref{['exper']}. As can be seen, our method has the best reconstruction metric and it only requires 4 iteration steps. Note that DiffuseRecon peng2022towards, MC-DDPM xie2022measurement, Score-MRI chung2022score, and AdaDiff gungor2023adaptive are employed for single-contrast SR (SCSR) reconstruction. DisC-Diff mao2023disc is specifically designed for multi-contrast SR (MCSR) reconstruction.
• Figure 2: The overall architecture of our proposed DiffMSR, which mainly divided into two parts: (1) Diffusion Model; (2) Prior-guide Large Window Transformer (PLWformer) including $N$ PLWformer layers and a image reconstruction module. CATL: Cross-Attention Transformer Layer.
• Figure 3: The process of training stage and inference stage.
• Figure 4: The architecture of prior-guide large window self-attention and prior-guide feed-forward network.
• Figure 5: Qualitative comparison of SOTA MCSR methods on four datasets with an upsampling scale of 4$\times$. The top, second, third, and bottom rows are the SR results under the FastMRI, clinical brain, clinical tumor and clinical pelvic datasets. Please zoom-in for details.