Applying the Deep Learning Method for Dose Estimation in Radiotherapy Publisher



Mahdavi N ; Zafarghandi M ; Setayeshi S ; Nedaie H A
Source: International Journal of Radiation Research Published:2026

Abstract

Background: The present study's main objective is to develop the model based on the U-net architecture to estimate dose distribution in heterogeneous phantoms by applying water phantom dose distribution and the characteristics of a heterogeneous phantom. Materials and Methods: The proposed model was developed based on the U-Net architecture, which includes five input channels and one output channel with dimensions 32×32×64 in the associated x, y, and z directions. Two hundred heterogeneous phantoms with water, lung, and bone materials at various depths were simulated using the Dosxyznrc code. Subsequently, the external photon source was applied to irradiate all the heterogeneous phantoms. The dose distribution of the water phantom served as the initial input channel, along with other input channels containing physical components of the heterogeneous phantoms. The dose distributions calculated by the Dosxyznrc code for each heterogeneous phantom were used as output. The developed model was trained using all the provided input and output data associated with the training data on the Google Colab execution platform. Results: The findings show that over 98% of total voxels met the 3%/3 mm gamma test requirements in the water area before heterogeneity. Furthermore, more than 99% of the voxels in the heterogeneous lung and bone medium satisfied the gamma test requirement of 3%/3 mm. Conclusion: The dose distribution in heterogeneous phantoms can be predicted quickly and accurately using the developed model based on a U-net architecture, which leverages the distribution of dose in a water phantom and the physical properties of the heterogeneous phantom. © 2026, Novin Medical Radiation Institute. All rights reserved.