Patch-based diffusion model for accelerated cardiac cine reconstruction

Background: Real-time (RT) cine imaging offers the advantages of free-breathing and robustness to arrhythmias over breath-held segmented acquisitions. However, RT cine commonly uses k-space undersampling with sparsity-based compressed sensing (CS) reconstruction [1] to achieve spatial and temporal resolution adequate for diagnostic imaging. At the high acceleration rates needed to meet these requirements, CS reconstructions often become over-smoothed, degrading fine cardiac structural details. Diffusion models [2] yield strong learned priors that can benefit this ill-posed inverse problem. We propose and evaluate a patch-based diffusion model for accelerated 2D cardiac cine imaging. Pretrained weights and code are available at https://github.com/OSU-MR/CineDiff.

Methods: Patch-based video diffusion was implemented in PyTorch [3] and trained using the Elucidated Diffusion Models (EDM) schedule [4]. Training data comprised of fully sampled 2D cine images (3T, 1.5T; short/long axis views, 4712 slices) from the OCMR [5] and CMRxRecon 2023–2025 databases [6]. Cine series were randomly partitioned into complex-valued 64×64×8 (H×W×T) patches. For inference, we stitched the overlapping patches together and used Diffusion Posterior Sampling (DPS) [7]. Each test case was reconstructed eight times with different noise seeds, and the outputs were averaged to boost peak signal-to-noise ratio (PSNR). For retrospective assessment, we undersampled multi-coil k-space from held-out OCMR (10 slices) and CMRxRecon 2024 (20 slices) images, each at acceleration rates (R) of 5, 10, and 15. We compared the proposed method against a CS reference [8] using PSNR, SSIM [9], LPIPS [10], and DISTS [11] metrics. An MRI expert blindly scored the results on the scale of 1 (worst) to 5 (best) based on overall image sharpness and visualization of structural details of the heart. This analysis was conducted on OCMR data (3T and 1.5T, 10 slices at R = 9) and from custom collected low-field data (0.55T, 5 slices each from human and porcine data; R = 10).

Results: Across held-out OCMR and CMRxRecon retrospectively undersampled images, the proposed method outperformed CS at R = 5, 10, and 15, increasing PSNR/SSIM and reducing LPIPS/DISTS, as summarized in Table 1a. A representative example is shown in Fig. 1. In the prospective study, the proposed method received higher expert scores than CS (Table 1b and Fig. 2). In all cases, the proposed method generated visibly sharper images.

Conclusion: Patch-based diffusion with DPS enables high-quality reconstruction of undersampled cardiac cine MRI, consistently outperforms CS on image-quality metrics, and generalizes to unseen low-field (0.55T) human and porcine cine. These findings support clinical translation of the proposed method for accelerated cine MRI. Future work will reduce runtime for clinical use, which is currently 10 to 20 minutes per cine series for 1000 diffusion steps on a Nvidia H100 GPU.

Table 1: Quantitative Comparison of the Proposed Method vs CS.

(a) Image quality metrics for retrospectively undersampled data.
	PSNR ↑	SSIM ↑	LPIPS ↓	DISTS ↓
R = 15
CS	29.08	0.8766	0.12007	0.14068
Proposed	30.89	0.9192	0.07334	0.10406
R = 10
CS	31.04	0.9155	0.08931	0.11628
Proposed	32.67	0.9398	0.05936	0.09317
R = 5
CS	34.19	0.9537	0.05189	0.08264
Proposed	35.43	0.9611	0.04337	0.07964
(b) Expert scores for prospectively undersampled data (higher is better).
	3T/1.5T OCMR(R = 9)	0.55T human(R = 10)	0.55T porcine(R = 10)
CS	4.0	3.2	2.8
Proposed	4.4	3.6	4.2

Note, italic numbers indicate better performance (higher PSNR/SSIM; lower LPIPS/DISTS; higher expert score).Figure 1: A representative example from the retrospectively undersampled data (R = 15) from CMRxRecon 2024, with red arrows highlighting details that are better preserved in the proposed method.
Figure 2: Representative examples from the prospectively undersampled data with the proposed method shown in the top row and CS in the bottom row. In general, CS images show blocky artifacts and loss of fine details.