2357 - A Robust Auto-Planning Algorithm Based Artificial Intelligence for Pencil Beam Scanning Proton Therapy

Y. Liu^1,2, X. Shang^1,2, Z. R. Li³, N. Li⁴, Z. Wang⁴, C. Fang⁴, Y. Zou⁴, W. Zhao⁵, X. Le², Q. Zhou³, G. Zhang², and S. Xu¹; ¹National Cancer Center/ National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China, ²School of physics, Beihang University, Beijing, China, ³Department of Research Algorithms, Manteia Technologies Co., Ltd, Xiamen, Fujian, China, ⁴Department of Oncology, Yizhou Pronton Therapy Center, Zhuozhou, Hebei, China, ⁵Beihang Hangzhou Innovation Institute, Hangzhou, Zhejiang, China

Purpose/Objective(s): While proton therapy has the theoretical potential for high precision, in practice, it still requires addressing uncertainties introduced by factors. However, conventional planning systems struggle to fastly optimize plans while accounting for these numerous sources of uncertainty. We developed a robust auto-planning algorithm based on artificial intelligence to overcome the challenges. Materials/

Methods: The robust automatic planning algorithm has two modules. The first is a GPU-accelerated neural network dose influence matrix (NN-D_ij) engine, which resulted that the D_ij matrix calculation being ultra-speedly in specific gantry and uncertainty settings. The second is an interface for reference DVHs as the optimization objectives for robust auto-planning. It allows input of reference plans or results of dose predictions. When applying a weight matrix to the D_ij matrix, each ROI has numbers (equal to uncertainty number) of DVHs, then calculate the punishment function (target/OAR DVH should be consistent/lower to reference) and find the worst DVH of each ROI for optimization. The robust auto-planning module ensures that all DVHs (in different uncertainties) in a plan are consistent/lower to reference DVHs. We set nine uncertainties and verified the algorithm on the lungs and head. By inputting the plan from the treatment planning system as a reference plan and the plan optimization was automatically completed.
Results: Evaluation used the difference between DVHs. For the target and OARs, the difference was 0.32±0.26Gy and 0.06±0.315Gy for the lung site(a detail in Table 1) and the mean value for the head site was -0.16±0.19Gy and -2.59±3.06Gy. The target D₉₅ difference for lung head cases was 0.31±0.25Gy and 0.13±0.47Gy. The time cost of NN-D_ij and robust auto-plan was within 2 minutes on RTX3090-server.
Conclusion: The robust automatic planning algorithm has a significant advantage in PBS planning, for automation, efficiency, quality and robustness. This not only enhances the quality and safety of clinical treatments but also promotes the modernization and efficiency of the entire proton radiotherapy process. Table 1. The value of auto-plan DVHs minus reference DVHs for a lung case

Uncertainty DVH	0	X +5mm	X -5mm	Y +5mm	Y -5mm	Z +5mm	Z -5mm	SP +5%	SP -5%
CTV D95 (Gy)	0.65	0.09	0.07	0.79	0.32	0.17	0.28	0.04	0.37
CTV Dmax (Gy)	-0.8	-0.3	-0.7	-0.2	-0.1	-0.3	0	0	-0.2
Spinal Cord Dmax (Gy)	-0.4	-0.5	-0.5	-0.5	-0.2	-0.2	-0.3	-0.4	-0.2
Esophagus Dmax (Gy)	-6	-4	-5.7	-6.9	-3	-6	-4.5	-7.1	-4.4
Esophagus V5 (%)	-2.3	-0.7	-2.1	-2.3	2.89	-2.2	-1.7	-2.3	0.65
Heart Dmax (Gy)	-0.1	1.67	0.61	2.1	-0.6	0.76	1.73	-3.8	1.71
Heart Dmean (Gy)	-0.1	-0.1	-0.8	0	-0.1	0.31	-0.1	-0.9	0.63
Lung-R Dmax (Gy)	-0.9	-0.6	-0.7	-0.4	-0.2	-0.5	-0.1	-0.2	-0.3
Lung-R Dmean (Gy)	0.47	0.66	0.11	0.62	0.33	0.62	0.22	0.5	0.31
Lung-R V5 (%)	1.78	2.78	-0.6	2.18	1.89	3.15	1.09	0.04	2.97
Lung-R V20 (%)	1.89	2.06	1.56	2.15	1.5	1.94	1.49	2.06	1.4