Abstract
Accelerating Magnetic Resonance Imaging (MRI) acquisition is critical for cardiac imaging, where patient motion and physiological constraints limit scan duration. While deep learning methods like Generative Adversarial Networks (GANs) have shown promise in reconstructing images from under-sampled k-space data, they often suffer from training instability and hallucinated artifacts. This paper presents DiffCMR, a framework utilizing conditional denoising diffusion probabilistic models (DDPMs) for high-fidelity cardiac MRI reconstruction. By framing reconstruction as a conditional denoising task and implementing a multi-round ensembling strategy, DiffCMR achieves state-of-the-art performance on the CMRxRecon benchmark, demonstrating superior stability and structural fidelity compared to traditional deep learning baselines.
1. Introduction
Magnetic Resonance Imaging (MRI) is a cornerstone of non-invasive cardiac diagnostics, yet its clinical utility is often hampered by slow acquisition speeds. Scanners must collect data points in the frequency domain, known as k-space; in Cardiac MRI (CMR), the constant motion of the heart necessitates rapid acquisition to minimize motion artifacts and patient discomfort during breath-holds.
To address this, clinicians employ under-sampling, collecting only a fraction of the required k-space data. However, reconstructing images from such sparse data typically results in significant aliasing and blurring. While standard deep learning architectures like U-Nets and GANs have been applied to this problem, GANs in particular are prone to mode collapse and instability. DiffCMR proposes a diffusion-based alternative that leverages the generative power of DDPMs while maintaining strict adherence to the underlying cardiac anatomy through conditional guidance.
2. Methodology
2.1 Conditional Diffusion Framework
DiffCMR is built upon the conditional DDPM framework. Unlike unconditional models that generate images from pure Gaussian noise, DiffCMR is conditioned on the under-sampled MRI slice. The model learns to reverse a diffusion process, iteratively removing noise to recover the high-quality image while being guided by the features of the original under-sampled input.
2.2 Feature Fusion Architecture
The architecture utilizes a modified U-Net designed for feature fusion. It incorporates two distinct encoders:
- Encoder F: Processes the noisy signal at the current diffusion timestep.
- Encoder G: Extracts features from the under-sampled conditioning image.
The features from both encoders are integrated and passed into a decoder (E) to estimate the noise component. This dual-encoder approach ensures that the model retains the structural context of the under-sampled scan throughout the denoising process.
2.3 K-Space Preprocessing
To prevent spatial bias and artificial edge artifacts, the authors implement zero-padding in the k-space rather than the image domain. By padding the frequency data to a standard $512 \times 512$ matrix before performing the Fourier transform, the model maintains the integrity of the noise source and ensures a consistent input distribution for the diffusion process.
2.4 Multi-Round Ensembling
Because diffusion is a stochastic process, individual reconstruction passes may contain minor random variations. To mitigate this, DiffCMR employs a multi-round ensembling strategy. During inference, the model generates multiple reconstructions ($R$) for the same input and averages the results. This "wisdom of the crowd" approach significantly stabilizes the output and improves the Peak Signal-to-Noise Ratio (PSNR).
3. Experiments and Results
3.1 Dataset and Tasks
The framework was validated using the MICCAI 2023 Cardiac MRI Reconstruction Challenge (CMRxRecon) dataset, which includes single-coil and multi-coil data from 120 training and 60 validation cases. The model was evaluated on two primary tasks: Cine reconstruction (dynamic heart motion) and T1/T2 mapping (tissue characterization), across acceleration factors of 4x, 8x, and 10x.
3.2 Performance Metrics
DiffCMR demonstrated significant improvements over standard U-Net and conditional GAN (cGAN) baselines:
- Cine Reconstruction (4x): DiffCMR achieved a PSNR of 36.10, outperforming cGAN (33.85) and U-Net (33.11).
- T1/T2 Mapping (4x): DiffCMR reached a PSNR of 34.60, compared to 32.67 for cGAN.
- Robustness: At an extreme 10x acceleration factor, DiffCMR maintained a PSNR above 34 for cine tasks, whereas baseline performance degraded significantly.
Furthermore, the Normalized Mean Square Error (NMSE) was consistently lower for DiffCMR, indicating higher structural fidelity and fewer reconstruction artifacts.
4. Discussion and Limitations
4.1 Stability and Ensembling
A key finding of this work is the inherent stability of diffusion models compared to GANs. While GANs require meticulous hyperparameter tuning, DiffCMR proved robust across various tasks. The study also quantified the impact of ensembling, showing that increasing the number of rounds from 1 to 8 consistently improved image quality, albeit at a higher computational cost.
4.2 Computational Trade-offs
Despite the "Fast" designation in the title, the iterative nature of diffusion models presents a challenge for real-time deployment. The current implementation uses 1000 steps for inference and multiple ensembling rounds, making it computationally more intensive than single-pass architectures like U-Net. Future research may focus on "distilled" diffusion models to reduce the number of required timesteps without sacrificing image quality.
5. Conclusion
DiffCMR represents a significant step forward in applying generative diffusion models to medical imaging. By combining conditional denoising with k-space preprocessing and ensembling, the framework provides a robust and high-fidelity solution for cardiac MRI reconstruction, even at high acceleration factors. This approach offers a promising alternative to GAN-based methods, prioritizing both image quality and training stability.
Summary updated