A reinforcement learning approach for finding optimal policy of adaptive radiation therapy considering uncertain tumor biological response

Abstract

Recent studies have shown that a tumor's biological response to radiation varies over time and has a dynamic nature. Dynamic biological features of tumor cells underscore the importance of using fractionation and adapting the treatment plan to tumor volume changes in radiation therapy treatment. Adaptive radiation therapy (ART) is an iterative process to adjust the dose of radiation in response to potential changes during the treatment. One of the key challenges in ART is how to determine the optimal timing of adaptations corresponding to tumor response to radiation. This paper aims to develop an automated treatment planning framework incorporating the biological uncertainties to find the optimal adaptation points to achieve a more effective treatment plan. First, a dynamic tumor-response model is proposed to predict weekly tumor volume regression during the period of radiation therapy treatment based on biological factors. Second, a Reinforcement Learning (RL) framework is developed to find the optimal adaptation points for ART considering the uncertainty in biological factors with the goal of achieving maximum final tumor control while minimizing or maintaining the toxicity level of the organs at risk (OARs) per the decision-maker's preference. Third, a beamlet intensity optimization model is solved using the predicted tumor volume at each adaptation point. The performance of the proposed RT treatment planning framework is tested using a clinical non-small cell lung cancer (NSCLC) case. The results are compared with the conventional fractionation schedule (i.e., equal dose fractionation) as a reference plan. The results show that the proposed approach performed well in achieving a robust optimal ART treatment plan under high uncertainty in the biological parameters. The ART plan outperformed the reference plan by increasing the mean biological effective dose (BED) value of the tumor by 2.01%, while maintaining the OAR BED within +0.5% and reducing the variability, in terms of the interquartile range (IQR) of tumor BED, by 25%.

Keywords: Adaptive radiation therapy.; Biological tumor response; Radiotherapy; Reinforcement learning.