Related Experiment Video
Updated: Jun 26, 2026
Dynamic Lung Tumor Tracking for Stereotactic Ablative Body Radiation Therapy
Published on: June 7, 2015
Automatic beam angle optimization in brain tumor radiotherapy using deep reinforcement learning
Han Guo1, Zhiqing Xiao1, Huandi Zhou2
1Department of radiation oncology, the Second Hospital of Hebei Medical University, Heping west road, Shijiazhuang, 050000, Asia, China.
Purpose:
This study aims to develop an automatic beam arrangement algorithm based on deep reinforcement learning (DRL) to explore DRL's clinical potential in radiotherapy.
Methods And Materials:
The Soft Actor-Critic (SAC) algorithm was used with planning data from brain tumor patients at our institution. A three-dimensional dose distribution incorporating target areas and isocenters was used as input. Data interaction with the ECLIPSE planning system was facilitated through ESAPI scripts. To enhance sampling efficiency, multi-agent parallel sampling was employed to generate beam distribution plans.
Results:
A total of 236 brain tumor patients were included. In the validation set, the model achieved a score of 73.48±23.17, compared to an initial plan score of 66.16±26.76. A random selection of 48 cases from the training set revealed that the best beam arrangement plans generated during training scored 83.33±25.18, while the initial plans scored 64.43±24.37. Furthermore, 14 cases not conforming to clinical beam arrangement practices were manually arranged, with the DRL group scoring 87.25±16.67 and the manual group scoring 79.94±20.02. All P-values were less than 0.05, indicating statistical significance.
Conclusions:
In this investigated cohort of brain tumor cases and under the evaluated technical configuration, the proposed DRL framework generated beam arrangement plans with improved plan scores, supporting its feasibility and potential value for beam angle optimization.
