MR image reconstruction from undersampled data for image-guided radiation therapy using a patient-specific deep
Jace Grandinetti1, Yin Gao1, Yesenia Gonzalez1
1Innovative Technology of Radiotherapy Computations and Hardware (iTORCH) Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, TX, United States.
Frontiers in Oncology
|December 8, 2022
Summary
This study introduces a novel deep learning method to improve Magnetic Resonance (MR) image quality for radiotherapy (RT) image guidance. The patient-specific approach significantly enhances MR image reconstruction from undersampled data, benefiting tumor targeting.
Area of Science:
- Medical Imaging
- Radiotherapy
- Machine Learning
Background:
- Advancements in radiotherapy integrate Magnetic Resonance (MR) imaging for precise tumor targeting.
- MR image acquisition acceleration via undersampling compromises image quality.
- High-quality patient MR images from treatment planning offer a unique prior for reconstruction.
Purpose of the Study:
- To develop a patient-specific deep learning method for high-quality MR image reconstruction from undersampled data.
- To leverage patient-specific MR image priors for improved image guidance in radiotherapy.
Main Methods:
- A deep auto-encoder was trained on patient-specific image patches to form a manifold.
- This manifold was used as regularization to restore undersampled MR images.
- The method was validated through simulation, real patient studies (liver cancer), and phantom experiments.
Main Results:
- The patient-specific method significantly improved peak-signal-to-noise ratio (4.46dB) and structural similarity index measure (28%) in simulations compared to patient-generic methods.
- Experimental studies demonstrated visually superior image quality reconstructions.
Conclusions:
- Exploiting patient-specific MR image priors with deep learning enables high-quality MR image reconstruction.
- This approach enhances image guidance capabilities in MR-guided radiotherapy.
- The method shows promise for improving accuracy and safety in radiation oncology.
