3666 - Unsupervised Deep Learning Model for Fast Energy Layer Pre-Selection in Delivery-Efficient Proton Arc Therapy
Presenter(s)
G. Liu1, B. Yang2, X. Ding3, P. Y. Chen4, Y. Luo5, G. Peng1, S. Zhang1, K. Yang1, and J. Huang1; 1Cancer Center, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, 2Electronic Information School, Wuhan University, Wuhan, Hubei, China, 3Corewell Health William Beaumont University Hospital, Royal Oak, MI, 4Department of Radiation Oncology, Corewell Health William Beaumont University Hospital, Royal Oak, MI, 5School of Computer Science, National Engineering Research Center for Multimedia Software and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, Wuhan, Hubei, China
Purpose/Objective(s):
Proton arc therapy is an emerging and attractive technique in the field of radiotherapy due to the comprehensive advantages compared the traditional intensity modulated proton therapy. However, it is time consuming to determine the optimal energy sequence from multitude of control points and extensive energy layers(ELs). This study is to develop an unsupervised deep learning framework that facilitates fast pre-select the optimal EL, minimizing energy layer switch while maintaining the superior plan quality and delivery efficiency.Materials/Methods: The number of spots that penetrates target and OAR are sorted by gantry angles and energy layers and aggregated into a matrix. A U-Net style architecture SPArc_dl takes in the matrices as proxy of radiation coverage to minimize a tri-objective function, maximizing the number of spots on target, minimizing the number of spots on OAR and minimizing energy switch time. We evaluate SPArc_dl on 63 patients with nasopharyngeal radiotherapy using plan quality and delivery efficiency benchmarked against proton arc plans generated by SPArc_partice swarm (SPArc_ps) algorithm.
Results: SPArc_dl demonstrates a remarkable capability to generate the optimal energy sequence with exceptional speed and maintains excellent plan quality. More specifically, the optimal energy layers is pre-selected in SPArc_dl faster with time cost 4832.21±1450.32s(SPArc_ps) vs. 0.03± 0.01s(SPArc_dl),p<0.01. Additionally, SPArc_dl produces a minimum number of energy layers, resulting in an 18% reduction in energy switch time compared to SPArc_ps (p < 0.01). Importantly, the plan quality remains comparable between SPArc_dl and SPArc_ps in terms of target coverage and normal tissue sparing.
Conclusion: The proposed algorithm is both swift and efficient, generating high-quality proton arc therapy (PAT) plans by strategically pre-selecting EL. This innovation effectively reduces delivery time while maintaining excellent plan quality, thereby enhancing the application of proton arc therapy in clinical settings.