1129 - A Multi-Center Study on the Clinical Deliverability of Fully Automated Planning for Stereotactic Body Radiation Therapy in Lung Cancer Using a Deep Learning Transfer Method

J. Pan^1,2, F. Zhuo³, C. Ruan¹, Z. Liang⁴, D. Li⁵, H. Zhao⁶, Y. Liu⁷, M. Han⁸, J. Xiao², Q. Song⁹, H. Jiang², and X. Li¹; ¹Department of Radiation Oncology, Renmin Hospital, Wuhan University, Wuhan, Hubei, China, ²School of Electronic Information, Wuhan University, Wuhan, Hubei, China, ³Department of Radiation Oncology, Jiangling County Peoples Hospital, Jingzhou, Hubei, China, ⁴Cancer Center, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, ⁵Department of Radiation Oncology, Hubei Cancer Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, ⁶Department of Radiation and Medical Oncology, Zhongnan Hospital of Wuhan University, Wuhan, China, ⁷Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, ⁸Shanghai United Imaging Healthcare Co., Ltd., Shanghai, Shanghai, China, ⁹Cancer Center, , Renmin Hospital of Wuhan University, Wuhan, Hubei, China

Purpose/Objective(s): This study presents an advanced fully automated planning solution, verifying the feasibility of deep learning-based transfer learning method in the automated planning of multi-center stereotactic body radiation therapy (SBRT) for lung cancer radiation treatment plans with diverse technology types. Materials/

Methods: We randomly collected radiation treatment plans with manual planning from 181 patients across four centers. Center ? and ? contributed 131 Intensity-Modulated Radiation Therapy (IMRT) plans (P_IMRT), on which a basic U-Net based dose prediction model (M_IMRT) was trained and tested. The predicted dose distribution informed the selection of DVH indices as the optimization constraints to guide the generation of automated IMRT plans (AP_IMAT). Center ? and ? provided 50 cases of Volumetric Modulated Arc Therapy (VMAT) plans (P_VMAT), which were used to refine and test a model (M_Trans) through transfer learning method from M_IMRT, aiming to create fully automated VMAT plans (AP_Trans). The plans were assessed via a quantitative comparison of dosimetric parameters and blind review by a physician with 10 years of experience.
Results: The M_IMRTs dose distribution predictions for 13 P<sub>IMRT test cases closely matched those of manual plans. Performance comparisons between AP_IMRT and AP_Trans, each leveraging M_IMRT and M_Trans respectively, were evaluated on 10 P<sub>VMAT cases. The dosimetric indices presented in Table 1 are the averages plus or minus one standard deviation (avg±std) for the corresponding plans. Statistically, AP_Trans aligned more closely with P<sub>VMAT than AP_IMRT, illustrating that transfer learning effectively adapts the dose distribution style to P<sub>VMAT. Moreover, both AP_IMRT and AP_Trans demonstrated lower standard deviations than P<sub>VMAT, indicating the enhanced robustness of automated planning. All 10 auto-generated plans were clinically approved upon reevaluation by a physician with extensive experience, and 7 were preferred by the radiologist.
Conclusion: This study shows no statistically significant difference in dosimetric parameters between the AP_Trans and P<sub>VMAT regarding overall plan quality. It confirmed the applicability of transfer learning in multi-center SBRT for lung cancer across various technologies, supporting its generalization and potential for clinical implementation. Table 1 Results of manual and automated plans

ROI	Index	PVMAT	APIMRT	APTrans P-value
PTV	Dmax [cGy]	7453±339	7266±134	7295±173 0.21
PTV	CI	1.01±0.03	1.10±0.06	1.10±0.06 0.00
D2cm	Dmax [cGy]	3310±673	3143±136	3220±120 0.67
Ipsilateral Lung	V20 [%]	9.30±6.10	10.8±6.50	9.70±6.00 0.90
Ipsilateral Lung	V10 [%]	17.9±9.50	19.5±10.3	17.39±9.05 0.92
Ipsilateral Lung	V5 [%]	24.9±11.6	26.9±12.8	25.4±11.8 0.93