Author: Liqin HU, Tao He, Jing JIA, Pengcheng LONG, Wei Meng, Yang Yuan 👨🔬
Affiliation: SuperAccuracy Science & Technology Co. Ltd. 🌍
Purpose: A multi-criteria optimization method based on reinforcement learning and adaptive boosting(RLAB MCO) has been developed to enhance radiotherapy plan quality by offering reasonable and effective solutions to handle the complex optimization problems in the clinical practice.
Methods: A reinforcement learning and adaptive boosting strategy has been added to the multi-criteria optimization process. An optimization mode classifier is created after the trade-off is completed, enabling it to determine the subsequent optimization strategy based on the user-selected trade-off results. The strategies include dose aligning, weighted interpolation, and pareto plan projection, each targeting different optimization goals. The classifier uses adaptive boosting and is trained with reinforcement learning to augment available planning samples. During the training process, the user’s trade-off results serve as the input and the ground truth serves as the output. To enhance the classifier performance, the planning target volume (PTV) dose distribution should align with the actual dose distribution while minimizing the dose of organs-at-risk (OARs).
Results: The proposed method was compared with baseline plans for ten patient cases in VMAT. Dose-volume histogram (DVH) results indicated that the new method reduced OAR dose and improved the conformity index (CI) and homogeneity index (HI). The difference in PTV dose distribution between the new and conventional methods was within 40 cGy.
Conclusion: The proposed RLAB MCO method overcomes the limitations of conventional optimization techniques, which often fail to deliver high-quality MCO plans. When the user’s trade-off is suboptimal, the proposed method can generate a plan that delivers a desired dose distribution comparable to conventional methods, without requiring additional adjustments.