From Predicted Dose to Computationally Verified Plan: An End-to-End Deep Learning and Monte Carlo Pipeline for Automated VMAT Glioblastoma Radiotherapy.

Authors

Abstract

Deep learning dose prediction can accurately reproduce clinical dose distributions but typically stops short of generating machine-parameterized treatment plans. The present study introduces an end-to-end pipeline that converts a predicted dose into a machine-parameterized volumetric modulated arc therapy (VMAT) approximation, evaluating dosimetric consistency via an independent Monte Carlo (MC) dose recomputation for cranial glioblastoma (GBM). A three-dimensional U-Net predicts voxel-level dose from the planning CT and structure masks. The framework was evaluated on 30 held-out test patients. Each prediction is converted into a discretized VMAT sequence via direct aperture optimization (DAO) and voxel-wise dose mimicking. This optimization is driven by a size-normalized squared-deviation objective that equalizes per-voxel optimization pressure across targets and organs at risk (OARs) of drastically different sizes. The resulting plans undergo an independent computational dose recomputation step using the ompMC photon Monte Carlo engine, reporting gamma analysis at 3%/2 mm evaluated over the high-dose (≥ 50%) and whole-field (≥ 10%) regions in accordance with TG-218 recommendations. The predicted dose matched the clinical dose to within 0.9-1.9% on PTV dose-volume metrics. The dose-mimicking pipeline successfully inverse-engineered these predictions into discretized VMAT plans that reproduced clinical PTV coverage while adhering to strict OAR tolerances. Independent Monte Carlo recomputation confirmed the dosimetric consistency of the generated plans. Utilizing a highly converged MC simulation (σ ≈ 0.8%), the high-dose-region gamma pass rate (3%/2 mm) reached a median of 98.2% [IQR 97.1-99.0%], confirming excellent agreement between the pencil-beam optimization and MC dose transport. The proposed pipeline bridges the critical gap from a deep learning-predicted dose to a machine-parameterized, computational VMAT plan. The size-normalised mimicking objective is identified as the key enabler for stable inverse-engineering. While true clinical deliverability requires further validation on a physical linear accelerator, the independent MC dose-recomputation step provides a rigorous computational QA framework that standard dose prediction studies currently omit.

Tags

Topics