Abstract
Cardiac MR (CMR) CINE imaging is essential for assessing heart function but is limited by the trade-off between scan time and image quality. Traditional motion-compensated reconstruction (MCMR) faces a "chicken-and-egg" dilemma: accurate motion estimation requires high-quality images, while high-quality reconstruction requires accurate motion data. This paper introduces an unrolled, joint optimization framework that iteratively refines both image reconstruction and motion estimation. Utilizing a lightweight GRAFT (Group-wise Registration and ATtention) network, the proposed method achieves high-fidelity reconstructions at acceleration rates up to 16x, significantly outperforming conventional non-iterative approaches.
Introduction: The "Chicken-and-Egg" of Cardiac MRI
CMR CINE imaging is the gold standard for evaluating cardiac structure and function. To capture the heart's rapid motion, technicians often undersample data to speed up scans, which typically introduces blurring and artifacts. Engineers address this using Motion-Compensated Reconstruction (MCMR), a two-step process involving motion estimation and image reconstruction.
However, traditional methods usually estimate motion only once at the start. If the initial undersampled image is of poor quality, the motion estimation is flawed, leading to a degraded final reconstruction. This paper proposes a framework to break this cycle by updating motion and image data simultaneously through an iterative process.
Methodology: Unrolled Joint Optimization
The researchers developed an unrolled, joint optimization framework that converts a traditional iterative mathematical algorithm into a series of neural network layers. This architecture alternates between two primary components:
- CG-SENSE Block: Reconstructs the image based on the current motion data and undersampled k-space data.
- GRAFT Block (Group-wise Registration and ATtention): A lightweight network (5 million parameters) that estimates motion by looking at the entire sequence simultaneously rather than frame-by-frame. This efficiency allows for registration in seconds rather than hours.
The framework is trained using self-supervised learning, ensuring that warped frames match the ground truth of target frames, which forces the network to learn genuine physical motion rather than simply masking artifacts.
Experimental Setup and Datasets
The study utilized a dataset of 40 subjects (healthy volunteers and patients) scanned on a 1.5T Siemens scanner. The training set included 304 image sequences from 35 subjects, while testing was performed on 15 sequences from 5 subjects. To simulate high-speed clinical scans, the data was retrospectively undersampled at acceleration rates of R=8, 12, and 16.
Results and Discussion
The proposed MCMR method demonstrated a "virtuous cycle" where improvements in image quality led to more precise motion estimation, which in turn enabled sharper reconstruction.
Quantitative and Qualitative Performance
At the highest acceleration rate (R=16), the proposed method achieved a PSNR of 34.78 dB, a massive improvement over the 21.74 dB achieved by standard non-motion-compensated (N-CG-SENSE) methods. Qualitatively, the model preserved sharp myocardial boundaries and recovered fine details in the spatio-temporal plane that were lost to streaking artifacts in other methods.
Efficiency and Convergence
By using the lightweight GRAFT network, the full reconstruction sequence takes approximately 18 seconds, making it suitable for clinical workflows. The researchers found that just three unrolled iterations were sufficient to reach a high performance plateau.
Conclusion and Future Work
While the current model significantly improves CMR reconstruction, it does not yet strictly guarantee "diffeomorphic" motion (preventing physical impossibilities like tissue folding). Future work will involve integrating scaling and squaring layers to ensure physical plausibility and testing the model on prospective, real-time clinical data. Combining this explicit motion correction with implicit deep learning denoisers may further push the boundaries of cardiac imaging acceleration.
Summary updated