Cross-Contrast Diffusion: A Synergistic Approach for Simultaneous Multi-Contrast MRI Super-Resolution

Abstract

Purpose: Diffusion-based deep-learning frameworks have been recently used in MRI resolution enhancement, or super-resolution. Multi-contrast MRI share common anatomical structures while holding complementary soft-tissue information. This study aims to incorporate both shared and contrast-specific features of multi-contrast MRI into a diffusion-based deep-learning framework for simultaneous multi-contrast MRI super-resolution.Methods: Public IXI brain dataset consisting of 576 healthy participants was used. High resolution (HR) multi-contrast MRI (T1-w and T2-w, 0.9375mm x 0.9375mm in 2D) were defined as groundtruth. Low resolution (LR) multi-contrast MRI (T1-w and T2-w, 3.75mm x 3.75mm in 2D) were simulated by low-pass filtering in K-space and used as input. LR T1-w and T2-w images were combined into a dual-channel representation in the forward diffusion process. In the backward process, a cross-contrast encoder with a spatial self-attention module was added to capture shared structural features and contrast-specific information. The separation was guided by a disentanglement term. All extracted features were processed through a Squeeze-and-Excitation module for channel-wise weighting. Charbonnier loss and a progressive training strategy were employed for training stability. The method was examined on the IXI dataset with training, validation, and testing split of 500:6:70 and compared with several state-of-the-art super-resolution methods, including Bicubic Interpolation, EDSR, SwinIR, Guided Diffusion, MINet, MASA-SR under the metrics of PSNR and SSIM.Results: The method achieved simultaneous high-quality T1-w and T2-w super-resolution, enhancing resolution by 2 and 4 times. On 2x super resolution, the PSNR/SSIM were 37.42 dB/0.9869 for T1-w and 37.70 dB/0.9879 for T2-w MRI. On 4x super resolution, the PSNR/SSIM were 31.61 dB/0.9570 for T1-w and 30.85 dB/0.9443 for T2-w MRI. Notably different from other multi-contrast methods, our method does not need HR MRI as input, while achieving the highest PSNR and SSIM on both magnification factors among the state-of-the-art (SOTA) methods. Conclusions: To our knowledge, this is the first study of simultaneous multi-contrast MRI super-resolution using diffusion-based framework. Our network effectively achieves high quality super-resolution of two MRI contrasts without need of high-resolution guidance. Future work includes extension studies in additional contrasts and 3D image scenarios.