Generative evidential synthesis with integrated segmentation framework for MR-only radiation therapy treatment
Lina Mekki1, Matthew Ladra2, Sahaja Acharya2
1Department of Biomedical Engineering, Johns Hopkins University, Baltimore, Maryland, USA.
Medical Physics
|April 12, 2025
Summary
This study introduces a novel deep learning method for MR-only radiation therapy (RT) planning, enabling fast and reliable CT synthesis from MRI and organ segmentation. The approach streamlines RT planning by providing accurate synthetic CTs and estimating their reliability.
Area of Science:
- Medical Imaging
- Radiation Oncology
- Artificial Intelligence
Background:
- Radiation therapy (RT) planning traditionally relies on CT scans for electron density information, but MRI offers superior soft-tissue contrast for contouring.
- Image registration between MRI and CT introduces uncertainties, leading to contouring errors and dose inaccuracies in RT planning.
- MR-only RT planning aims to eliminate CT scans and registration by synthesizing CT from MRI, but lacks methods to assess synthetic CT reliability.
Purpose of the Study:
- To develop a rapid and reliable method for simultaneous CT synthesis from MRI, uncertainty estimation, and organ-at-risk (OAR) segmentation.
- To enable MR-only RT planning by providing a trustworthy synthetic CT without the need for ground truth comparison.
Main Methods:
- Applied deep evidential regression within a multi-task vision transformer framework for MR-only brain RT planning.
- Utilized a nested encoder with separate convolutional decoders for CT synthesis and OAR segmentation, incorporating an evidential layer for uncertainty estimation.
- Trained and validated the framework on 119 paired T1-weighted MRI and CT scans with OAR contours.
Main Results:
- Achieved high performance in CT synthesis (SSIM: 0.820 ± 0.039, MAE: 47.4 ± 8.49 HU, PSNR: 23.4 ± 1.13) and OAR segmentation (e.g., Dice for brainstem: 0.953 ± 0.024).
- Completed the entire process, including synthesis and segmentation, in a rapid 6.71 ± 0.25 seconds.
- Evidential uncertainty estimation effectively identified regions of uncertainty, correlating with Monte Carlo-based methods, confirming the approach's reliability.
Conclusions:
- Developed a novel framework for joint CT synthesis, uncertainty prediction, and OAR segmentation in a single inference, leveraging deep evidential regression.
- The proposed method significantly streamlines MR-only RT planning by offering a reliable synthetic CT without ground truth.
- This approach holds potential to enhance efficiency and accuracy in radiation therapy planning.
