CineGen: Conditional Flow Matching-based Super-Resolution for High-Quality Free-Breathing Real-Time Cartesian Cine MRI

Background: Free-breathing real-time cine MRI is clinically valuable as it overcomes the limitations of traditional segmented Cartesian cine imaging by eliminating the need for multiple breath holds, reducing motion artifacts, and preserving image quality in patients with irregular heart rates or limited breath-hold capacity (1). While compressed sensing of Cartesian cine MRI achieves higher acceleration and is clinically adopted, pushing beyond 7–8× often results in degraded image quality. Combining low-resolution acquisitions with AI-based super-resolution (SR) enable higher acceleration (2), which is essential for achieving the high temporal resolution required in real-time cine MRI. Previously, we showed that diffusion models can super-resolve cardiac MR images (3, 4), but flow matching (FM) (5) simplifies the process and achieves faster sampling. Here, we propose a FM–based SR approach that reconstructs high-quality cine MRI images with faster inference time, achieving ~21× acceleration and outperforming the GAN-based SR method (2), thereby establishing a foundation for high resolution, high-quality inline Cartesian real-time cine MRI.

Methods: Dataset: (a) Retrospective data: ~2,000 standard breath-hold bSSFP cine MRI short-axis slices from 200 patients (1.5T Aera, Siemens) were used for training, and 120 slices from 12 patients for testing. (b) Prospective 21x accelerated data: Three patients (3T Vida, Siemens) were scanned with ~38% central k-space (R = 2.6) and compressed sensing (R = 8), resulting ~21x acceleration. Each slice was acquired in one heartbeat with 1.4×1.4 mm² spatial and 30–40 ms temporal resolution.

Model: We used the conditional rectified FM model as shown in Figure 1A. For model hyperparameters, we used five levels in U-Net, with the number of channels in each level being [64,128,256,512,512], and the model was trained for 200,000 iterations with a learning rate of 3e-5 and a batch size of 64. During training, timesteps was set to 1000; during inference, 20 timesteps were used.

Results: For synthetic testing data with multiple slices, CineGen outperformed REGAIN (2) with improved contrast, sharper myocardial borders, and higher image quality relative to HR references (Figure 1B). Another synthetic example with multiple temporal frames (Figure 2A) and corresponding temporal profiles showed that CineGen was superior in temporal fidelity and sharpness (Figure 2B). Quantitatively, CineGen achieved lower nRMSE and higher PSNR and SSIM than REGAIN (Figure 2C). For prospectively acquired 21x accelerated data, CineGen provided better myocardium–blood pool contrast compared to REGAIN, closely matching reference breath-hold cine images (Figure 3).

Conclusion: We developed CineGen, a conditional FM–based SR model that enables highly accelerated real-time cine MRI with ~21x acceleration. By incorporating LR acquisitions with compressed sensing, CineGen bridges the gap between image acquisition speed, image generation speed and reconstruction quality.

Figure 1. Conditional flow matching–based generative modeling for cardiac cine MRI super resolution, and comparisons of REGAIN and CineGen on a retrospectively synthetic testing example. (A) Schematic of the conditional flow matching (CFM) framework: starting from Gaussian noise 𝑥_0 using a paired low resolution cine MRI image 𝑧 as inputs, cardiac MR images are iteratively denoised by a U-Net model. At each timestep 𝑡, the model predicts the velocity field 𝑣_𝜃 (𝑥_𝑡, 𝑡, 𝑧), which is trained to match the rectified flow target u_𝑡 (𝑥_𝑡│𝑧). The model was updated by the CFM loss​, enabling efficient reconstruction of high-resolution cine MRI from LR inputs. Low-resolution and high-resolution (LR/HR) pairs were synthesized by retaining 30–40% of central k-space lines, zero-padding, and center-cropping both LR and HR images to 128×128. (B) Synthetic example images of multiple short-axis slices for a patient. The low-resolution cine MR images in the first row were simulated from the high-resolution reference images in the last row, using 38% central phase encoding lines (R=2.6) and zero-padding reconstruction. Low-resolution images were super-resolved by REGAIN (second row) and CineGen (third row), respectively. CineGen better preserves anatomical details and contrast as compared to low-resolution images, REGAIN super resolution, and reference standard breath-hold cine MRI.
Figure 2. Evaluation of CineGen on a retrospective patient data across multiple cardiac phases, and temporal profile plots. (A) Example images for a patient with synthetic low resolution across multiple cardiac phases. The low-resolution input was zero-padded using central 38% phase-encoding lines. CineGen demonstrates sharper myocardial borders and reduced artifacts compared to REGAIN and shows better consistency with the reference HR images. (B) Corresponding spatiotemporal (x–t and y–t) profiles extracted from the boxed regions. CineGen generates clearer myocardial wall motion patterns with preserved temporal fidelity, reducing aliasing and temporal artifacts observed in low-resolution and REGAIN super-resolved images. (C) Quantitative comparisons were performed on all cine MRI images, including a total of 120 short-axis slices from 12 independent subjects using HR cine MRI (spatial resolution 1.4-1.5 × 1.4-1.5mm2) as the reference in terms of Structural Similarity Index (SSIM), normalized Root-Mean-Square-Error (nRMSE) and Peak Signal-to-Noise Ratio (PSNR). CineGen demonstrated better performance on all three metrics compared to REGAIN, with lower nRMSE, higher PSNR, and higher SSIM (P < 0.05 for all metrics using paired t-test).
Figure 3. Comparisons of REGAIN and CineGen for super-resolution of prospectively acquired free-breathing real-time cardiac cine MRI from one patient with ~21-fold acceleration (38% phase-encoding lines, compressed sensing rate 8). Representative short-axis frames across the cardiac cycle reconstructed with different methods were shown, including low-resolution image reconstructed from 38% phase-encoding lines and 8-fold compressed sensing, REGAIN super-resolution, and CineGen super-resolution. The real-time acquisition was performed at 1.4 × 1.4 mm² spatial resolution, 30.7 ms temporal resolution. CineGen improves myocardial border sharpness, image contrast and overall image quality compared to low-resolution input and REGAIN and showed superior consistency with the reference breath-hold images separately acquired at the same location.