Zixu Guan, Yukine Shimizu, Takahiro Iwai, Michio Yoshimura, Takashi Mizowaki, Mitsuhiro Nakamura
This study demonstrates the feasibility of MATE-UNet for the direct prediction of VMAT machine parameters without iterative optimization. By leveraging multi-modal BEV inputs and a Transformer-enhanced architecture, the proposed framework represents a valuable step toward bridging the gap between dose distributions and clinically usable treatment plans. Although an additional normalization step is required to determine the absolute MUs, MATE-UNet has the potential to serve as a downstream component of a future fully automated anatomy-to-plan pipeline.
BACKGROUND: Pancreatic cancer Volumetric Modulated Arc Therapy (VMAT) planning presents a significant dosimetric challenge due to the high-dose gradients required to spare adjacent, radiosensitive organs at risk (OARs) like the stomach and duodenum. This anatomical complexity has limited the scope of automated planning for this site. Broadly, while deep learning (DL) has been introduced to streamline treatment planning, most existing models only predict intermediate outputs, such as dose distributions or fluence maps, which still necessitate a subsequent, computationally expensive inverse optimization step on a treatment planning system.
PURPOSE: To address these challenges, we aim to develop an optimization-free fully automated VMAT planning framework for pancreatic cancer. As a key component of this system, this study introduces a DL model designed to directly generate machine parameters from dose distributions and anatomical contours.
METHODS: A total of 200 vmat plans for pancreatic cancer (prescription: 42 Gy in 15 fractions) were retrospectively collected. The dataset was randomly split into training (n = 170), validation (n = 10), and testing (n = 20) sets. The proposed Multi-modal, Attention & Transformer-Enhanced U-Net (MATE-UNet) utilizes beam's-eye-view (BEV) projections of the reference dose distribution and anatomical contours to directly predict machine-executable multi-leaf collimator (MLC) apertures and Monitor Units (MUs). The proposed model was benchmarked against baseline U-Nets (using contour-only, dose-only, and combined inputs) on the testing set. Model accuracy was assessed using the Dice Similarity Coefficient (DSC) for MLC and Mean Absolute Error (MAE) for MU, while plan quality was evaluated using clinical dose-volume histogram (DVH) metrics and Conformity Index (CI), Homogeneity Index (HI), and Gradient Index (GI).
RESULTS: MATE-UNet achieved a clinical acceptance rate of 100% (20/20), compared to 70% for the best-performing baseline model. In terms of prediction accuracy, the proposed model achieved a DSC of 0.9553 ± 0.0042 for MLC apertures and an MAE of 1.960 ± 0.396 for MU. Dosimetric evaluation demonstrated that, relative to the reference plans, MATE-UNet maintained comparable target coverage, CI (0.773), and HI (0.100), while achieving a significantly improved GI (3.732, p < 0.05). Furthermore, MATE-UNet significantly reduced the V39Gy of the stomach and duodenum, as well as the global maximum dose, compared with the baseline models (p < 0.05).
CONCLUSIONS: This study demonstrates the feasibility of MATE-UNet for the direct prediction of VMAT machine parameters without iterative optimization. By leveraging multi-modal BEV inputs and a Transformer-enhanced architecture, the proposed framework represents a valuable step toward bridging the gap between dose distributions and clinically usable treatment plans. Although an additional normalization step is required to determine the absolute MUs, MATE-UNet has the potential to serve as a downstream component of a future fully automated anatomy-to-plan pipeline.