{"id":5,"url":"https://pm.philipcastiglione.com/papers/5.json","title":"Deep learning-based rigid motion correction for magnetic resonance imaging: A survey","read":false,"authors":"Yuchou Chang, Zhiqiang Li, Gulfam Saju, Hui Mao, Tianming Liu","year":2023,"auto_summary":"The paper titled \"Deep learning-based rigid motion correction for magnetic resonance imaging: A survey\" provides a comprehensive review of the current state of deep learning methods applied to correct rigid motion artifacts in MRI scans. These artifacts, primarily caused by patient movement during scans, can lead to image distortions that affect diagnostic accuracy. The survey discusses various deep learning strategies that have been developed to address these issues, comparing them with traditional motion correction techniques and hybrid methods that combine elements of both.\n\nKey points from the paper include:\n\n1. **MRI and Motion Artifacts**: MRI is a preferred imaging modality due to its superior soft tissue contrast and resolution. However, its sensitivity to patient movement during long scan times results in motion artifacts, which are a significant source of image degradation.\n\n2. **Types of Motion Correction**: The paper categorizes motion correction techniques into prospective motion correction (PMC), retrospective motion correction (RMC), and hybrid methods. PMC involves real-time adjustments during the scan, while RMC corrects images after acquisition. Hybrid methods combine both approaches.\n\n3. **Deep Learning Motion Correction (DLMC)**: DLMC methods treat the scan as an input-output system, learning the relationship between motion-corrupted and motion-free images. These methods can be categorized based on whether they consider variations in pulse sequences during data acquisition.\n\n4. **DLMC with Pulse Sequence Consideration**: Techniques that incorporate pulse sequence information, such as those using multishot acquisitions or parallel imaging, have shown promise in improving motion correction. These methods often involve complex models like generative adversarial networks (GANs) to enhance image quality.\n\n5. **DLMC without Pulse Sequence Consideration**: These methods use large datasets to train models to directly map motion-corrupted images to motion-free ones, without specific sequence information. Convolutional neural networks (CNNs) and other architectures like residual networks and variational autoencoders are commonly used.\n\n6. **Challenges and Future Directions**: The paper highlights challenges such as the need for large datasets, the generalization of models to new motion patterns, and the integration of DLMC with existing PMC and RMC methods. The authors suggest that future work should focus on improving model transparency and developing hybrid methods that combine deep learning with traditional approaches.\n\n7. **Evaluation and Validation**: The paper discusses the importance of both quantitative and qualitative evaluations of DLMC methods. While quantitative metrics like MSE, SSIM, and PSNR are commonly used, qualitative assessments by radiologists are crucial for clinical validation.\n\nOverall, the paper emphasizes the potential of deep learning to significantly improve motion correction in MRI, while also acknowledging the need for further research to address existing limitations and enhance clinical applicability.","notes":{"id":7,"name":"notes","body":null,"record_type":"Paper","record_id":5,"created_at":"2024-12-10T03:51:57.459Z","updated_at":"2024-12-10T03:51:57.459Z"},"created_at":"2024-12-10T03:51:08.303Z","updated_at":"2024-12-10T03:52:11.776Z"}