Improving Efficiency of Training a Virtual Treatment Planner Network via Knowledge-guided Deep Reinforcement Learning for Intelligent Automatic Treatment Planning of Radiotherapy

TL;DR

This paper introduces a knowledge-guided deep reinforcement learning method that significantly accelerates training of a virtual treatment planner network for radiotherapy, enabling efficient development for complex clinical scenarios.

Contribution

The study proposes a novel KgDRL approach that incorporates human expert rules to improve training efficiency of VTPN in radiotherapy planning.

Findings

KgDRL reduces training time from over a week to 13 hours.

VTPN trained with KgDRL achieves comparable plan quality to traditional DRL.

Incorporating human knowledge enhances training efficiency and scalability.

Abstract

We previously proposed an intelligent automatic treatment planning framework for radiotherapy, in which a virtual treatment planner network (VTPN) was built using deep reinforcement learning (DRL) to operate a treatment planning system (TPS). Despite the success, the training of VTPN via DRL was time consuming. Also the training time is expected to grow with the complexity of the treatment planning problem, preventing the development of VTPN for more complicated but clinically relevant scenarios. In this study we proposed a knowledge-guided DRL (KgDRL) that incorporated knowledge from human planners to guide the training process to improve the training efficiency. Using prostate cancer intensity modulated radiation therapy as a testbed, we first summarized a number of rules of operating our in-house TPS. In training, in addition to randomly navigating the state-action space, as in the DRL using the epsilon-greedy algorithm, we also sampled actions defined by the rules. The priority of sampling actions from rules decreased over the training process to encourage VTPN to explore new policy that was not covered by the rules. We trained a VTPN using KgDRL and compared its performance with another VTPN trained using DRL. Both VTPNs trained via KgDRL and DRL spontaneously learned to operate the TPS to generate high-quality plans, achieving plan quality scores of 8.82 and 8.43, respectively. Both VTPNs outperformed treatment planning purely based on the rules, which had a plan score of 7.81. VTPN trained with 8 episodes using KgDRL was able to perform similarly to that trained using DRL with 100 episodes. The training time was reduced from more than a week to 13 hours. The proposed KgDRL framework accelerated the training process by incorporating human knowledge, which will facilitate the development of VTPN for more complicated treatment planning scenarios.