Abstract
Cardiac Magnetic Resonance (CMR) imaging is the clinical gold standard for assessing heart function, yet it remains highly sensitive to patient motion. Motion during acquisition corrupts raw k-space data, leading to ghosting and blurring that can compromise clinical diagnosis. This paper presents an end-to-end deep learning framework that automatically detects motion-corrupted k-space lines and treats them as missing data. By reframing artifact correction as an undersampled reconstruction task, the proposed dual-network architecture—comprising a 3D CNN for detection and a Recurrent Convolutional Neural Network (RCNN) for reconstruction—significantly improves image fidelity. Evaluated on the UK Biobank dataset, the method achieves a PSNR of 37.1, outperforming standard denoising and GAN-based approaches while maintaining high computational efficiency.
Introduction
Cardiac Magnetic Resonance (CMR) imaging is essential for evaluating cardiac morphology and function. However, the acquisition process is notoriously sensitive to motion, such as respiratory movement, patient discomfort, or arrhythmias. These factors corrupt the raw data stored in k-space, resulting in artifacts that obscure myocardial boundaries and can cause automated analysis tools to fail.
Currently, the standard clinical response to significant motion artifacts is to re-scan the patient, which increases costs and reduces throughput. This research aims to "fix it in post" by identifying exactly which parts of the raw data are corrupted and reconstructing the image without them, effectively treating artifacts as a data-filling problem rather than a noise-filtering problem.
Methods
The authors propose a dual-network architecture that integrates detection and correction into a single end-to-end pipeline.
- The Detection Network: A spatio-temporal 3D Convolutional Neural Network (CNN) analyzes the 2D+time k-space data to classify which lines are corrupted by motion.
- The "Data Consistency" Bridge: This is the core innovation. Instead of attempting to "clean" corrupted data, the detection network creates a mask. The corrupted lines are discarded, transforming the problem into an undersampled reconstruction task.
- The Reconstruction Network: A Recurrent Convolutional Neural Network (RCNN) fills in the missing k-space information by exploiting temporal dependencies across different frames of the cardiac cycle.
The entire system is trained using a joint loss function that balances detection accuracy (cross-entropy) and image reconstruction quality (mean squared error). This synergy allows the reconstructor to "inform" the detector about which types of errors are most detrimental to final image quality.
Experimental Results
The study utilized 300 cine SSFP CMR datasets from the UK Biobank. To provide ground truth for training, high-quality scans were retrospectively corrupted by swapping random k-space lines to simulate realistic motion effects.
Quantitative Performance
The proposed end-to-end method was compared against several state-of-the-art techniques, including standard denoising (DNCNN) and GAN-based reconstruction (Automap-GAN). The results demonstrated significant improvements:
- PSNR (Peak Signal-to-Noise Ratio): The proposed method achieved 37.1, significantly higher than the baseline (26.3) and standard denoising (32.2).
- SSIM (Structural Similarity Index): Reached 0.890, indicating high structural fidelity to the original uncorrupted heart images.
- Speed: Once trained, the network can correct a single image sequence in less than one second, making it viable for real-time clinical use.
Qualitative Analysis
Visual results showed much sharper myocardial boundaries and better recovery of the septum compared to other methods. By treating artifacts as missing data, the network avoids the over-smoothing often seen with traditional denoising filters.
Conclusion
This paper demonstrates that reframing motion artifact correction as an undersampled reconstruction task is a powerful strategy for CMR. The end-to-end framework effectively leverages temporal redundancies in the heart cycle to recover missing information without introducing blurring. Future work will focus on validating the method against "naturally" corrupted clinical data and scaling the approach to cover the full heart volume across the entire UK Biobank cohort.
Summary updated