Related Experiment Video
Updated: Oct 4, 2025
08:34
Proton Therapy Delivery and Its Clinical Application in Select Solid Tumor Malignancies
Published on: February 6, 2019
20.5K
Facilitating MR-Guided Adaptive Proton Therapy in Children Using Deep Learning-Based Synthetic CT
Chuang Wang1, Jinsoo Uh1, Thomas E Merchant1
1Department of Radiation Oncology, St Jude Children's Research Hospital, Memphis, TN, USA.
International Journal of Particle Therapy
|February 7, 2022
Summary
A novel deep-learning model using self-attention cycle-generative adversarial networks (cycle-GANs) accurately generates synthetic computed tomography (sCT) for pediatric brain tumor patients. This AI-driven sCT facilitates adaptive proton therapy by enabling precise treatment adjustments.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Radiation Oncology
- Medical Physics and Imaging
Background:
- Adaptive proton therapy requires accurate patient imaging for treatment adjustments.
- Conventional methods for generating synthetic computed tomography (sCT) may lack precision, especially in pediatric patients with brain tumors.
- Deep learning, specifically generative adversarial networks (GANs), offers potential for improving image synthesis.
Purpose of the Study:
- To evaluate the efficacy of self-attention cycle-generative adversarial networks (cycle-GANs) in generating accurate sCT for pediatric brain tumor patients.
- To determine if the generated sCT can effectively support adaptive proton therapy planning.
Main Methods:
- A novel cycle-GAN model incorporating a self-attention mechanism was developed to enhance tissue interfaces and reduce noise in sCT generation.
- The model was trained using CT and T1-weighted MRI data from 125 pediatric brain tumor patients.
- The performance of the self-attention cycle-GAN was validated on a test set of 7 patients, comparing MRI-based sCT with replanning CT (ground truth).
Main Results:
- The self-attention cycle-GAN significantly reduced Hounsfield unit-mean absolute error compared to conventional cycle-GAN (65.3 ± 13.9 vs. 88.9 ± 19.3, P < .01).
- High 3D gamma passing rates (2%/2 mm) were achieved for both original and adapted plans using self-attention cycle-GAN generated sCT (97.6% ± 1.2% and 98.9% ± 0.9%, respectively).
- Plan adaptation was appropriately triggered in all test patients, with minimal dose differences observed in the clinical target volume (CTV) and distal falloff regions.
Conclusions:
- The self-attention cycle-GAN model demonstrates superior performance over conventional cycle-GAN for sCT generation in pediatric brain tumor patients.
- The generated sCT shows promising dosimetric accuracy, supporting its use in adaptive replanning.
- This AI-driven approach can help identify pediatric patients who would benefit from adaptive proton therapy.
