Dl-Dose Framework: Introducing a Fast Dose Calculation Engine for Radiotherapy-Based Deep Learning Publisher



Mahdavi N ; Zafarghandi MS ; Setayeshi S ; Nedaie HA
Source: Iranian Journal of Medical Physics Published:2026

Abstract

Introduction: This study aimed to develop a dose prediction framework based on deep learning that utilizes water dose distribution and the characteristics of heterogeneous phantoms as inputs. Material and Methods: A dataset of two hundred heterogeneous phantoms with identical geometry and variable bone and lung layer thicknesses was generated using the DOSXYZnrc Monte Carlo (MC) code. The thickness of each inhomogeneity was randomly assigned between 1 and 5 cm at different positions along the z-axis of the phantom. A deep learning–based dose prediction framework was then developed to estimate three-dimensional dose distributions. The model used five input channels, including water dose distribution, mass density, CT number, voxel distance from the radiation field center, and a binary radiation field mask. The network's output was the predicted dose distribution for each heterogeneous phantom. Results: The accuracy of the predicted results by the DL-Dose Framework was evaluated against those obtained through the Monte Carlo method, using the delta index in heterogeneous phantoms. In the water medium, before encountering heterogeneities, 100% of the dose distribution for voxels with deviations of less than 1% from the maximum dose, consistent with the results measured by the MC method. Furthermore, 94.2% of the dose distribution for voxels in lung heterogeneity areas and 98.1% for voxels in bone heterogeneity regions were comparable to the MC method results, with deviations of less than 1% from the maximum dose. Conclusion: The developed DL-Dose Framework can accurately predict dose distribution in heterogeneous phantoms. © 2026 Mashhad University of Medical Sciences. All rights reserved.