2379 - The Feasibility of Newly Developed AI-Based Automatic Radiotherapy Planning for Lung Cancer in SBRT-VMAT: Comparison with Manual Treatment Planning

H. Nemoto^1,2, M. Saito¹, T. Komiyama¹, Y. Kimura³, N. Kadoya², R. Tozuka¹, and H. Onishi¹; ¹University of Yamanashi, Chuo, Japan, ²Department of Radiation Oncology, Tohoku University Graduate School of Medicine, Sendai, Japan, ³AiRato. Inc, Sendai, Japan

Purpose/Objective(s): SBRT-VMAT planning is created manually, resulting in inconsistent plan quality and a time-consuming process. To improve efficiency and quality of planning, we evaluated the feasibility of AI-based automatic SBRT-VMAT planning for lung cancer. Materials/

Methods: We developed the AI-based dose distribution prediction model (installed in RatoGuide prototype, AiRato. Inc) which was trained using manually optimized plans for 120 cases of stage I lung cancer (peripheral cases (55 Gy / 4 Fr) and central cases (50 Gy / 8 Fr), 54 cases each, both with PTV D95% prescription). Test data consisted of 6 treatment plans each of central and peripheral lung cancer. These plans were manually optimized and used for treatment, previously. The automatic plan was created following procedures: (i) predicting dose distribution based on target and OAR contours and CT images using RatoGuide, (ii) converting the predicted dose distribution into contours and importing them into the treatment planning system, and (iii) generating mimicked dose distributions using predefined optimization parameters and constraints. We measured the time to create plans and compared the DVH metrics between automatic and manual plans. Two expert radiation oncologists evaluated automatic and manual plans based on both visual inspection and DVH metrics in following aspects: (a) clinically acceptable or not and (b) preference for automatic or manual plan in clinical practice.
Results: DVH metrics and creation time for the 2 plans are shown in Table 1. No significant differences were observed in all DVH metrics between automatic and manual plans for peripheral cases, whereas significant differences were observed in some metrics of lungs for central cases. Radiation oncologists reviewed all automatic plans as clinically acceptable. The percentages of cases which 2 radiation oncologists preferred automatic plans to manual plans were 66.7% and 33.3%, respectively, for peripheral cases, and 91.7% and 50.0%, respectively, for central cases.
Conclusion: Our newly developed method of automatic SBRT-VMAT treatment planning of lung cancer was fast and comparable to manually optimized plans in dose distribution. It can be feasible for clinical use, but we will focus on improving the accuracy of the prediction model and its workflow. Table 1 - Abstract 2379

	Peripheral			Central
	Manual	Automatic	p-value	Manual	Automatic	p-value
PTV D98% (Gy)	53.25±0.4	53.66±0.15	0.16	48.27±0.58	48.32±0.16	1.00
PTV D2% (Gy)	70.25±1.05	67.92±1.57	0.63	60.01±0.9	60.89±0.48	0.16
Confomity index	0.87±0.04	0.87±0.02	0.69	0.9±0.02	0.89±0.01	0.56
Lungs V20Gy (%)	3.41±1.39	3.68±1.54	0.22	5.56±1.58	5.57±1.51	1.00
Lungs V5Gy (%)	12.84±4.37	17.12±7.17	0.06	16.47±3.62	26.07±8.2	0.03*
Creatoin time (min)	-	5.25±0.98	-	-	4.95±0.47	-