Feasibility study on deep learning-based human organ dose estimation in monoenergetic point-source exposures

Rapid and accurate estimation of organ-absorbed doses is critical for effective radiological protection and emergency response. While Monte Carlo (MC) simulations provide high fidelity, their computational cost hinders real-time application. This study proposes a deep neural network (DNN)-based surrogate model to predict organ-absorbed doses and effective doses for an adult male using the ICRP Publication 145 mesh-type reference computational phantom. A dataset comprising approximately 54,000 cases was generated using Geant4 simulations for point sources emitting 1 MeV photons isotropically. Various data transformation strategies, including log and Box-Cox transformations, were evaluated to optimize model performance. The proposed DNN model reproduced MC-calculated doses with a mean absolute relative error (ARE) of less than 2.5%, which is comparable to the inherent statistical uncertainty of the training data. Notably, the model demonstrated a processing time of 0.76 s for 5000 cases, significantly outperforming conventional gamma-constant-based methods in both accuracy and real-time applicability. An error analysis suggested that prediction residuals were partially associated with the statistical noise of the simulation data and related to geometric complexities, such as anatomical shielding and steep dose gradients near the body surface. These findings validate the feasibility of the proposed AI-driven approach as a robust tool for real-time computational dosimetry.