E Wang, C Shen*, Y Gonzalez, X Jia, The University of Texas Southwestern Medical Center, Dallas, TX

• 
  C Shen

SU-H330-IePD-F7-5 (Sunday, 7/10/2022) 3:30 PM - 4:00 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 7

Purpose: In image-guided adaptive high-dose-rate brachytherapy of cervical cancer, treatment planning optimization relies on a human-planner to adjust parameters in the optimization problem, such as weights of clinical tumor volume (CTV) and organs at risk (OARs), potentially yielding suboptimal plans due to planner inexperience and time pressure. This study develops an automated planning process using deep reinforcement learning (DRL) to build a virtual planner trained to autonomously make human-like decisions in adjusting planning parameters to generate high-quality plans that optimize treatment outcomes.

Methods: Plan quality score was defined as tumor control probability (TCP) divided by the average normal tissue complication probability (NTCP) of the bladder, rectum, and sigmoid colon. TCP and NTCP were calculated from combined brachytherapy and external beam radiotherapy dose based on prior studies investigating dose-response relationships. The virtual planner was represented by a deep neural network that observes CTV D90 and OARs’ D2cc and decides adjustment actions to planning parameters in the optimization engine. We trained this neural network via an end-to-end DRL process to maximize a reward function of the plan quality score. Experience replay and an ε-greedy algorithm were implemented. Over 24,000 state-action pairs generated from real patient cases were used for training. We tested the virtual planner to autonomously plan additional patient cases.

Results: Plans generated by the virtual planner attained qualities surpassing corresponding human-generated plan quality 5 of 6 times, attaining 6.50% higher scores on average. This improvement was attributable to improvements in average TCP by 1.41% and reduced average NTCP by 5.46%.

Conclusion: A DRL-based virtual planner was trained to autonomously operate the treatment planning optimization engine, generating plans of higher clinical quality than human planners. Our study demonstrates the immense potential of DRL-guided approaches to maximize clinical outcomes in treatment planning.

Keywords

Brachytherapy, Treatment Planning

Taxonomy

TH- Brachytherapy: Treatment planning using machine learning/automation

Contact Email