Related Experiment Video
Updated: Aug 10, 2026
08:34
Proton Therapy Delivery and Its Clinical Application in Select Solid Tumor Malignancies
Published on: February 6, 2019
Patient-Specific Deep Reinforcement Learning for Proton Beam Delivery Under Inter-Phase Variations
Mélanie Ghislain1, Estelle Loÿen1, Antoine Aspeel2
1UCLouvain (ICTEAM), Place du Levant 3, Louvain-la-Neuve, 1348, Belgium.
International Journal of Particle Therapy
|August 9, 2026
Summary
This study introduces a deep Reinforcement Learning (RL) framework for real-time proton therapy adaptation, improving lung tumor motion management. The RL approach enhances target coverage and reduces healthy tissue irradiation compared to conventional methods.
Area of Science:
- Medical Physics
- Radiation Oncology
- Artificial Intelligence in Medicine
Background:
- Proton therapy for lung tumors faces challenges from respiratory motion.
- Conventional methods like safety margins or robust optimization can irradiate healthy tissues.
- Real-time plan adaptation during treatment delivery is a promising strategy to mitigate motion effects.
Purpose of the Study:
- To develop and evaluate a patient-specific deep Reinforcement Learning (RL) control framework for proton pencil beam scanning.
- To address real-time plan adaptation for respiratory motion in proton therapy.
- To mitigate intrafractional motion effects during lung tumor treatment.
Main Methods:
- A deep RL-based control framework was developed for proton pencil beam scanning.
- RL agents were trained on patient CT data to control beam position and spot delivery.
- The learned policy was applied to 4DCT data without online retraining, using 2D observations of target geometry, beam position, and delivery history.
- The approach was compared against static Gross Tumor Volume (GTV)-based and Internal Target Volume (ITV)-based planning strategies.
Main Results:
- The RL-based approach improved target coverage under respiratory motion compared to GTV-based plans, with an average gain of 4.54 Gy in D95GTV for Patient 1.
- Compared to ITV-based plans, the RL approach reduced organ-at-risk dose exposure.
- Specific reductions included an average of 0.42, 4.12, and 3.14 Gy in DmeanLung-GTV for the three patients, and 1.41 Gy in DmeanHeart for Patient 3.
Conclusions:
- The study demonstrates the potential of RL-based control for proton therapy delivery under intrafractional motion.
- This framework serves as a proof-of-concept for future dynamic and adaptive proton therapy treatments.
- Further development can lead to more effective motion management in radiation oncology.
