Unigeodose: A Universal Real-Time Dose Verification Model for All Beam Energies and Treatment Units Via Beam Characteristics Embedded Geometry Encoding
Abstract
Purpose
Timely dose verification is essential for quality assurance (QA) in modern radiotherapy (RT), particularly in online adaptive RT, where measurement-based QA is often impractical. Current approaches are limited by machine/energy-specific designs, hindering scalability in real-world clinical applications. To address this, we propose UniGeoDose, a universal framework that supports real-time dose verification for all treatment machines and beam energies using a single model, providing a deploy-once, use-everywhere solution for all clinical settings.
Methods
To generalize across diverse clinical settings, UniGeoDose incorporates machine/beam characteristics as model input via a designated encoding block. Specifically, line profiles and percentage depth dose curves are integrated to represent both spatial and depth-dose characteristics of each beam energy on each LINAC. An autoencoder was trained to extract key features and concatenated with dosimetric leaf gap and leaf transmission factor for a comprehensive representation of beam characteristics. This representation conditions a geometry-encoded U-Net, a novel architecture that resolves the exact treatment delivery geometry by directly embedding control points from treatment plans into CT volumes, to specify the machine/energy under evaluation. To build the UniGeoDose model, a cohort of 995 prostate cancer cases were collected, and randomly split into 850 cases for training, 74 cases for validation, and 71 cases for independent testing with stratified sampling by machine and energy.
Results
Compared to dose calculated with the treatment planning system, UniGeoDose achieved an average γ-passing rate (3%/2mm, 10% threshold) of 99.89% ± 0.19% for all machines and energies available in our clinic with mean computation time <40ms per case, demonstrating superior performance comparing against the same model established without beam information encoded.
Conclusion
By introducing a novel beam-aware design, this study established a highly scalable approach to achieve real-time 3D dose verification, providing a universal solution for any beam energies, LINACs, and clinical settings.