Exhibit Hall | Forum 4
Purpose: Tradeoffs between a target’s prescription and the max dose in proximal organs at risk require iterative manipulation of planning parameters to reach an optimal plan. This work develops the ability for an auto-planning agent trained by reinforcement learning to shape the dose in regions of competing constraints.
Methods: Regions with competing constraints were formulated via max dose and dose falloff for the control of an auto-planning agent. The agent was given control over the dosimetric constraints and morphology of the regions. The agent was then trained on a set of 50 Pancreas SBRT plans with a prescription of 25 Gy and a simultaneous boost to 33 Gy. A maximum dose constraint was set to 29 Gy for the GI structures. The agent was tested by planning on a testing set of 15 cases that were not included in the training set. The convergence of the training was measured using a derived metric that measures the stability of the agent’s learning function.
Results: The agent’s performance on the testing set was compared to the clinically delivered plans as well as to a previously reported baseline agent that has non-region-specific control by comparing the final dose metrics for the PTV33 and GI structures. Compared to the baseline agent, the introduction of the dose falloff control resulted in better plans 83% of the time. When the GI structures were considered of equal or lesser importance the agent out-performed the clinical plans 88% of the time and 83% with the GI considered more important.
Conclusion: The dose falloff control allowed an auto-planning agent better control of the dose gradient in regions of overlap or proximity of competing optimization constraints resulting in clinical quality plans.
Not Applicable / None Entered.