199 - A Reinforcement Learning Approach to Automate Breast Radiation Therapy Treatment Planning Using Electronic Compensation (ECOMP)

D. Yang¹, X. Li¹, S. Yoo¹, R. C. Blitzblau², S. McDuff¹, S. J. Stephens³, P. Segars⁴, Q. J. J. Wu⁵, F. F. Yin⁶, and Y. Sheng¹; ¹Duke University Medical Center, Durham, NC, ²Duke University Medical Center, Department of Radiation Oncology, Durham, NC, ³Department of Radiation Oncology, Duke University Medical Center, Durham, NC, ⁴Department of Radiology, Duke University Medical Center, Durham, NC, ⁵Duke University, Durham, NC, ⁶Duke Kunshan University, Kunshan, China

Purpose/Objective(s): The electronic compensation (ECOMP) technique for breast radiation therapy provides excellent dose conformity and homogeneity. However, the manual fluence painting process presents a challenge for efficient clinical operation. This study utilized deep reinforcement learning (RL) to enable ECOMP automatic planning via fluence map editing under clinical objectives guidance. Materials/

Methods: With institutional review board (IRB) approval, seventy patients treated with a 6MV tangential ECOMP plan were included in this study (20/50 for training/testing). Pixel-wise RL agents were trained to perform fluence editing actions. The backbone of the RL framework was deep Q learning. Taking a beam’s-eye-view (BEV) projected dose profile as state input, the RL agents selected actions based on the predicted Q values to edit the fluence intensity, aiming to maximize expected plan quality improvement by balancing dose coverage and hot spot control. After fluence map editing, the dose was updated and the RL reward was calculated from the change in dose coverage and hot spot control. The reward was then utilized as feedback to update the RL network parameters. RL generated testing plans were evaluated with dose distribution and dosimetric endpoints (Breast V90%, Breast V95%, Breast V105%, Lung V20Gy, Heart V5Gy, Heart Dmean and Dmax) and compared with clinical plans.
Results: The RL framework averaged 90 seconds to generate a treatment plan with 20 iterations of editing. The RL plans exhibited comparable plan quality to clinical plans in terms of isodose distribution and dosimetric endpoints. Dosimetric endpoints statistics are presented in Table 1 and significance was evaluated with Wilcoxon Signed-Rank test. While maintaining comparable dose coverage (Breast V95%), the RL plans showed improved hotspot control with an average 25.9 cc breast V105% reduction. Furthermore, the RL framework demonstrated reduced variation and improved consistency compared to clinical plans, as indicated by a lower standard deviation of Breast V105%. OAR sparing was comparable between the groups as demonstrated by close Lung V20Gy(%), Heart V5Gy(%), and Heart Dmean(Gy) metrics.
Conclusion: We developed a self-driven automatic treatment planning tool using deep Q learning for breast ECOMP radiation therapy. Results demonstrate reliability and consistency of the tool, and importantly, significantly enhanced planning efficiency.

Abstract 199 – Table 1

Dosimetric Endpoints	RL	Clinical	P Value
Breast V90% (%)	82.60±8.11	82.80±8.12	<.05
Breast V95% (%)	77.76±8.90	78.57±9.09	<.001
Breast V105% (cc)	8.52±11.47	34.30±36.30	<.001
Lung V20Gy (%)	10.35±6.21	10.28±6.03	<.001
Heart V5Gy (%)	0.15±0.39	0.15±0.40	0.593
Heart Dmean (Gy)	0.51±0.24	0.52±0.24	<.05
Dmax (%)	108.19±1.29	107.96±1.15	0.263
Note: P Value <.05 was indicative of significant difference.