Room: Exhibit Hall | Forum 5
Purpose: Online adaptive radiotherapy (OART), enabled by the recently-introduced MR-Linac, can generate and deliver high-quality plans. However, plan quality ranking in the current OLAR systems is subjective and can be time consuming. This work aims to develop a method to automatically rank plans for OART.
Methods: Thirty-nine dose-volume parameters (DVPs) were calculated and reduced to representative parameters using a recursive feature elimination (RFE) algorithm with a random forest (RF) estimator. An ensemble of RFs were trained with representative parameters using either the acceptable/not acceptable or the superior/inferior plan quality labels assigned by a planner. In the former case, the quality scores (1-100) assigned by the trained RF were used to determine the quality of a plan / rank, while in the latter case the assigned superior scores were used to determine the superior plan in a pair of plans. The method was trained/evaluated using 100 paired (1 reposition and 1 adaptive) plans created based on daily CTs from 20 prostate cancer patients. The machine predictions were analyzed and compared to the plannerâ€™s evaluation.
Results: The ensemble of RF trained with acceptable/not acceptable labels assigned higher quality scores to the plans with the higher target coverage and/or normal tissue sparring. When the RFs were trained with the superior and inferior labels, the difference in superior score between paired plans implied a greater difference between plans. When the superior score difference was greater than 20 points, the trained RF correctly predicted 90% of the time the superior plan as determined by the planner. The trained RF running on a quad-core CPU with a processor base frequency of 3.4 GHz evaluated and compared plans within 10 seconds.
Conclusion: We have developed a method that can automatically evaluate and compare RT plans, thus, can be used to accelerate OART.
Funding Support, Disclosures, and Conflict of Interest: MCW Fotsch Foundation
Not Applicable / None Entered.