Advancing 5T Cine MRI Reconstruction with Cross-Domain Diffusion Model and Temporal Consistency

Background: Cine CMR is the reference for assessing ventricular function, motion, and myocardial mechanics. Higher field strengths such as 5T offer improved SNR, spatial resolution, and tissue characterization^1,2, but cine imaging at 5T is limited by long acquisition, particularly for retrospective scans requiring repeated breath-holds. Deep learning unrolled networks can accelerate cine MRI at 1.5T and 3T³, but they require large fully sampled cine datasets, which are not yet available for 5T due to scanner's novelty. In contrast, diffusion models provide flexible image priors learned from diverse MRI datasets and can generalize across tasks without retraining^4,5. Here, we propose using a diffusion prior not trained on cardiac data for undersampled 5T cine reconstruction, combined with a novel temporal regularization framework, to reduce scan times while preserving temporal fidelity.

Methods: Retrospectively gated cardiac cine images were acquired from 3 subjects on a 5T scanner (uMR Jupiter, United Imaging) using a 2D spoiled GRE sequence (FOV 360×320 mm², matrix 224×179, slice thickness 8 mm, TE 2.27 ms, TR 3.91 ms, FA 9°, 20 cardiac phases). For each subject, 2-chamber, 4-chamber, and three short-axis slices were acquired. Tenfold retrospective undersampling was applied to simulate accelerated acquisition.

The diffusion model, implemented with a score-based generative framework⁶, was trained unconditionally on over 10,000 MR slices from brain, knee, breast, and abdomen (public and private datasets). As shown in Figure 1, the model serves as a prior during reconstruction, combined with an ADMM-based⁷ temporal TV regularization to enforce data fidelity and temporal coherence. Conventional CG-SENSE and temporal-TV CS reconstructions were used for comparison.

Results: Figure 2 shows reconstruction results from undersampled cine data alongside the fully sampled reference. Figures 2(A–C) display short-axis, 4-chamber, and 2-chamber views. The diffusion reconstruction reduced artifacts and improved boundary definition, particularly at the atrioventricular valves, compared with TV reconstruction. Figure 2(D) presents a y–t plot from a short-axis line, showing that diffusion reconstruction better preserves temporal motion, whereas TV reconstruction oversmooths dynamics.

Table 1 summarizes PSNR and SSIM across 304 frames. The proposed method achieved slightly higher PSNR but slightly lower SSIM than TV reconstruction.

Conclusion: We show that diffusion model priors, even without cardiac training, can effectively reconstruct undersampled 5T cine MRI. With ADMM-based temporal regularization, the method preserves temporal fidelity and improves image quality, particularly at anatomical boundaries. Although its SSIM was slightly lower than TV reconstruction—likely reflecting TV's oversmoothing rather than clinical quality—radiologist evaluation will be needed in future studies. This approach has the potential to accelerate 5T cine acquisitions while maintaining high-resolution functional assessment.

Table 1. Average PSNR and SSIM of different reconstruction methods compared to the fully sampled ground truth. The numbers in the brackets show the standard deviation. The results were calculated over all the 304 frames.

	CG-SENSE Reconstruction	Temporal TV Reconstruction	Diffusion Model Reconstruction
PSNR	25.84 (1.39)	35.58 (2.15)	35.95 (1.92)
SSIM	0.6496 (0.0408)	0.9247 (0.0205)	0.8876 (0.0231)

Figure 1. The proposed undersampled cine MRI reconstruction pipeline using diffusion model and data consistency with temporal TV regularization.
Figure 2. Reconstruction images from different methods at (A) short-axis view, (B) 4-chamber view, (C) 2-chamber view, and (D) y-t view along a line through short-axis image. The red arrows point to regions that the proposed diffusion reconstruction shows sharper boundaries.