S1869
Physics - Dose prediction/calculation, optimisation and applications for photon and electron planning
ESTRO 2026
1 Department of Advanced Medical Physics, Graduate School of Medicine, Kyoto University, Kyoto, Japan. 2 Department of Radiation Oncology and Image- Applied Therapy, Graduate School of Medicine, Kyoto University, Kyoto, Japan Purpose/Objective: This study aims to develop and validate a deep learning framework for fully automated VMAT treatment planning in pancreatic cancer, capable of generating deliverable DICOM RT plans from anatomical contours and dose distribution. Material/Methods: We retrospectively collected 200 single full-arc VMAT cases for pancreatic cancer (Prescription: 42 Gy / 15 fractions), divided into training (n=170), validation (n=10), and testing (n=20) sets. The baseline comprised a standard 3D-UNet1 architecture (an encoder and a MLC decoder) and a separate MU decoder. We proposed a MATE-UNet (Multi-modal, Attention & Transformer-Enhanced U-Net), which enhanced the baseline by incorporating: (i) a multi-stage feature fusion module, (ii) attention gates2 in the skip connections, and (iii) a Transformer3 module at the bottleneck. For network input, 3D contours, including PTV-PRV (volume receiving the prescribed dose), PTV, stomach, and duodenum, and the 3D dose distribution were projected onto the BEV at all 180 control points. The network was trained to predict the corresponding MLC sequences and MUs, which were subsequently used to generate deliverable plan files. These plans were then imported into the Eclipse™ TPS for dose calculation and evaluation against clinical requirements. All predicted plans were normalized, consistent with clinical protocol, ensuring 100% of the prescribed dose covered 95% of the PTV-PRV. Results: MATE-UNet successfully met all clinical goals in 100% (20/20) of the testing plans, substantially outperforming the baseline model, which achieved all requirements in 70% (14/20) plans. Dosimetric analysis (Table 1) confirmed that MATE-UNet yielded statistically significant improvements (p < 0.05) over the baseline in crucial OAR sparing, including Dmax for body, V39Gy for stomach and duodenum, and V36Gy for stomach. Regarding pass rates (plans that achieved the goal), the baseline model failed to meet the criteria for PTV D98% in 2 cases (10%) and Stomach V39Gy in 4 cases (20%). A visual comparison (Figure 1) for a representative case demonstrated the superior dose conformity of the MATE-UNet plan. As highlighted in the magnified inserts, the proposed model achieved both excellent sparing of the duodenum (contoured in pink) and a visibly tighter 36 Gy isodose line (green) compared to the baseline.
Figure 2. Box plots for dose and plan quality parameters derived from the automated plans, clinical plans, and the predicted doses. The dashed line represents the DHV-parameter and gamma pass-rate goals used in our clinic. Conclusion: We developed and validated a deep learning based automated treatment planning workflow for left-sided breast cancer RT. The automated treatment planning workflow produced clinically acceptable treatment plans comparable to the original clinical treatment plans. Automated plan generation may significantly reduce the treatment planning workload and aid in standardising clinical planning practices. Keywords: automation, treatment planning, dose prediction,
Proffered Paper 2036
Direct Generation of VMAT Plan from Dose and Contours Using Deep Learning-based Auto- planning in Pancreatic Cancer Zixu Guan 1 , Yukine Shimizu 1 , Takahiro Iwai 2 , Michio Yoshimura 2 , Takashi Mizowaki 2 , Mitsuhiro Nakamura 1
Made with FlippingBook - Share PDF online