Purpose: Virtual treatment planner network (VTPN) has been previously developed to operate treatment-planning systems (TPSs) and adjust planning parameters to generate a plan in lieu of human planners. Previous applications were limited to relatively simple planning tasks on unrealistic in-house TPSs. This study reports the development of VTPN that is able to operate a realistic TPS for a real clinical task of prostate cancer SBRT treatment planning. We evaluate it objectively against human performance in the case in the planning challenge of ProKnow 2016 AAMD/RSS Plan Study.
Methods: Due to technical challenge of directly interacting with a commercial TPS, we built a realistic TPS equivalent to a commercial TPS in terms of inverse planning. It uses the actual dose engine in the real TPS, and solves the identical optimization problem defined by the user-adjustable dose-volume constraints. A hierarchical VTPN was constructed to perform multi-level decision-making in planning process to adjust constraints to improve plan quality. We trained VTPN via end-to-end deep reinforcement learning with human experience as guidance. The established VTPN was applied to the case in the planning challenge of ProKnow 2016 AAMD/RSS Plan Study to generate a 7-field intensity-modulated radiotherapy plan. The resulting plan quality was compared against top plans in the challenge generated by expert human planners.
Results: VTPN was successfully trained to operate the TPS, achieving an average plan score of 139.8 on validation cases. When applying the trained VTPN to plan for the case used in the planning challenge, it achieved a plan score of 140.5, ranked as the third highest score among all the 31 plans submitted for competition.
Conclusion: VTPN can spontaneously learn planning strategies and handle complex prostate cancer SBRT treatment planning. This is the first time that VTPN achieves performance similar to expert human planners.
Not Applicable / None Entered.