Related Experiment Video
Updated: Dec 21, 2025
Dynamic Lung Tumor Tracking for Stereotactic Ablative Body Radiation Therapy
Published on: June 7, 2015
Deep learning-based image reconstruction and motion estimation from undersampled radial k-space for real-time
Maarten L Terpstra1, Matteo Maspero, Federico d'Agata
1Department of Radiotherapy, University Medical Center Utrecht, Utrecht, The Netherlands. Computational Imaging Group for MR Diagnostics & Therapy, Center for Image Sciences, University Medical Center Utrecht, Utrecht, The Netherlands.
Deep learning motion estimation with conventional MRI reconstruction enables accurate 2D deformation vector fields for real-time adaptive radiotherapy. This combination achieves high temporal resolution and low latency, crucial for precise radiation delivery.
Area of Science:
- Medical Physics
- Radiotherapy Technology
- Artificial Intelligence in Medicine
Background:
- Real-time adaptation in MRI-guided radiotherapy requires rapid motion estimation.
- Deep learning (DL) offers potential to reduce latency by enabling higher undersampling in MRI acquisition.
- Investigating DL's impact on image reconstruction and motion estimation is key for improving real-time adaptive radiotherapy.
Purpose of the Study:
- To evaluate the benefit of DL for image reconstruction and motion estimation in obtaining accurate deformation vector fields (DVFs).
- To compare conventional and DL-based methods for real-time motion estimation in MRI-guided radiotherapy.
- To assess the trade-offs between undersampling, DVF accuracy, image quality, and latency.
Main Methods:
- Retrospective analysis of 2D cine MRI from 135 abdominal cancer patients at 1.5 T.
- Simulation of retrospectively undersampled radial golden angle acquisitions.
- Computation of DVFs using combinations of conventional and DL-based image reconstruction and motion estimation methods.
Main Results:
- Conventional methods yielded lowest DVF error and highest SSIM up to a specific undersampling factor.
- For higher undersampling factors, conventional reconstruction with DL-based motion estimation achieved the lowest DVF error and highest SSIM.
- This combination produced accurate DVFs (RMSE < 1 mm, SSIM > 0.8) within 60 ms, even with high undersampling.
Conclusions:
- Conventional image reconstruction combined with DL-based motion estimation enables high-quality 2D DVFs from highly undersampled k-space.
- This approach provides high temporal resolution and minimal latency, suitable for real-time adaptive MRI-guided radiotherapy.
- The findings support the use of DL for motion estimation to enhance adaptive radiotherapy precision.
