ESTRO 2026 Congress Report I Physics track I Proffered Paper
Authors:
Zixu Guan1, Yukine Shimizu1, Takahiro Iwai2, Michio Yoshimura2, Takashi Mizowaki2, Mitsuhiro Nakamura2
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:
Volumetric Modulated Arc Therapy (VMAT) is an essential modality in modern pancreatic cancer radiotherapy, offering highly conformal dose distributions. The traditional treatment planning workflow is a manual, iterative and labour-intensive process that requires a significant time investment. To improve planning efficiency, various AI-based frameworks have been developed. However, most existing methods only predict dose distributions or fluence maps; consequently, further optimisation within a Treatment Planning System (TPS) remains necessary to generate the final plan, which is still time-consuming. To achieve a rapid, optimisation-free workflow, we developed a multi-modal, attention & transformer-enhanced UNet (MATE-UNet) architecture. This system is uniquely designed to directly map anatomical contours and dose distributions to linear accelerator (LINAC) machine parameters, specifically multi-leaf collimator (MLC) sequences and monitor units (MUs).
Methods:
We retrospectively collected clinical single full-arc VMAT cases from pancreatic cancer patients to serve as the dataset for model training, validation, and independent evaluation. All cases utilised a standard clinical prescription of 42 Gy delivered in 15 fractions. To overcome the limitations of a standard 3D-UNet baseline architecture, our proposed MATE-UNet framework integrates a multi-stage feature fusion module, attention gates within skip connections, and a transformer module at the network bottleneck. Rather than relying on conventional three-dimensional volumetric data, the network operates on Beam’s Eye View (BEV) projections. Specifically, patient anatomical contours, including targets and crucial organ-at-risk (OARs), along with the 3D dose distributions, were projected onto the BEV across all 180 control points. The model directly predicts the corresponding MLC sequences and MUs. To ensure clinical validity, these predicted machine parameters were utilised to generate the DICOM RT plan files, which were then imported back into the Eclipse™ TPS for independent dose calculation and evaluated against clinical protocol criteria.
Significant Findings:
The core findings presented onsite demonstrated that the MATE-UNet framework substantially outperformed the baseline 3D-UNet model. Key highlights include:
- 100% Clinical Acceptance: MATE-UNet plans successfully satisfied all clinical goals in 100% (20/20) of the testing cases. In contrast, the baseline model plans achieved only a 70% (14/20) acceptance rate, specifically failing to meet PTV D98% coverage in two cases and stomach V39Gy constraints in four cases.
- Superior OAR Sparing: Dosimetric analyses confirmed statistically significant improvements (p < 0.05) in OAR sparing. The proposed model notably reduced the maximum dose to the body, as well as the high-dose volumes for the stomach (V39Gy, V36Gy) and duodenum (V39Gy).
Research Implications:
The successful validation of the MATE-UNet framework demonstrates the feasibility of an end-to-end radiotherapy planning system. Specifically, if coupled with a high-performance dose prediction model, this framework can directly translate patient data (CT images and anatomical contours) into LINAC machine parameters (MLC and MU). This integration has the potential to establish a rapid, direct auto-planning pipeline that completely circumvents the need for conventional TPS optimisation. Consequently, this optimisation-free approach can improve overall treatment planning efficiency, providing a highly practical and essential foundation for optimising real-time adaptive radiotherapy workflows.
Zixu Guan
PhD student
Department of Advanced Medical Physics
Kyoto University
Kyoto, Japan
zixu.guan.32i@st.kyoto-u.ac.jp
https://medicalphysics.hs.med.kyoto-u.ac.jp/