Author: Zixu Guan, Takahiro Iwai, Takashi Mizowaki, Mitsuhiro Nakamura, Michio Yoshimura 👨🔬
Affiliation: Kyoto University, Department of Radiation Oncology and Image-applied Therapy, Graduate School of Medicine, Kyoto University 🌍
Purpose:
The goal of this study is to develop a fully automated treatment planning approach for VMAT in pancreatic cancer that can convert patient anatomy into LINAC machine parameters. In this work, we focus on predicting the MLC aperture sequence, a key step in generating a deliverable DICOM-RT plan.
Methods:
The dataset consisted of 140 full-single-arc VMAT cases in pancreatic cancer with various prescriptions (42 Gy, 45 Gy, 48 Gy in 15 fractions), divided into 123 for training, 10 for validation, and 7 for testing. 3D contours and 3D dose distribution were projected along the BEV at all 178 control points as input to the network. A customized 3D-UNet was trained with 3 input groups: contours only (Model-R), dose only (Model-D), and the combination of contours and dose (Model-C), predicting the MLC apertures as the output. The predicted MLC sequence and clinical MUs were used to create DICOM-RT plans, which are imported into Eclipse® for dose calculation and evaluation by the clinical requirement. All predicted plans underwent dose normalization consistent with the clinical plan.
Results:
All predicted plans met the clinical requirement for D95% in PTV-PRV (volume receiving the prescribed dose). For D98% in PTV, 100% of Model-C plans pass, compared to 85.7% for Model-D and Model-R. The average Dmax values in body are 114.6% (Model-C), 115.7% (Model-D), and 142.4% (Model-R). For V42Gy in duodenum, the pass rates are 85.7% (Model-C and Model-D), and 28.6% (Model-R). For V39Gy in stomach, the pass rates are 71.4% (Model-C) and 57.1% (Model-D and Model-R). For other OARs (liver, kidneys, spinal cord), all predicted plans meet dose constraints.
Conclusion:
Model-C, using both dose and contours as inputs, achieved the highest clinical pass rates among the three models, meeting all target dose requirements and demonstrating its potential for fully automated treatment planning.