Click here to

Session: Artificial Intelligence in Treatment Planning and Delivery [Return to Session]

Towards Interpretable Intelligent Automatic Treatment Planning in Radiotherapy: Understanding the Decision-Making Behaviors of a Hierarchical Deep Reinforcement Learning Based Virtual Treatment Planner Network

C Shen*, L Chen, X Jia, (1) University of Texas Southwestern Medical Center, Dallas, TX


SU-A-TRACK 6-6 (Sunday, 7/25/2021) 10:30 AM - 11:30 AM [Eastern Time (GMT-4)]

Purpose: Intelligent automatic treatment planning (IATP) in radiotherapy builds a virtual treatment planner network (VTPN) that automatically operates a treatment planning system (TPS) to generate high-quality plans by adjusting dose constrains in TPS like a human. It is important to interpret VTPN’s decision-making behaviors to gain confidence for clinical applications. This study proposed a VTPN of a hierarchical architecture (HieVTPN) motivated by the decision-making process of human planners, and performed numerical studies to analyze and understand its behaviors in planning process.

Methods: HieVTPN consists of three networks, i.e. Structure-Net, Parameter-Net, and Action-Net. In each step controlling the TPS to improve a plan, HieVTPN sequentially determines the anatomical structure to adjust dose constraints, the specific constraint of the selected structure, and the detailed adjustment action, respectively. A hierarchical deep reinforcement learning approach was developed to train HieVTPN to operate an in-house developed TPS to generate plans for prostate cancer intensity modulated radiotherapy. We analyzed the decision-making behaviors of the established HieVTPN on anatomical structure, parameter, and adjustment action levels, respectively. Scenarios triggering distinct decisions were characterized to understand the rationale of the decision-making behaviors, which were then compared with human intuitions.

Results: HieVTPN were successfully trained to operate the in-house developed TPS to generate high-quality plans for all 59 testing patient cases, achieving an average plan quality score of 8.62 (±0.83), with 9 being the maximal score. We found that the decision-making behaviors of HieVTPN were understandable. By observing a plan, HieVTPN tended to pick the anatomical structure that needs the most attention, and adjusted the appropriate parameter corresponding to the selected structure towards the direction to improve the plan. Such decision-making behaviors generally agreed with the human experience.

Conclusion: Our study demonstrated the agreement between the decision-making behaviors of HieVTPN in planning process with that of human based on intuitions.



    Not Applicable / None Entered.


    TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation

    Contact Email