Deep Neural Network for risk assessment via organ dose estimation in inhomogeneous radiation fields

When radiation exposure occurs, evaluating the radiation dose is necessary to assess the risk and implement appropriate protective measures. Typically, radiation workers use personal dosimeters, which conservatively calculate the effective dose based on measured values. However, in scenarios involving potential high radiation exposure, emergency response tasks, or accidental exposure, precise dose evaluation is crucial. Conventional methods estimate human dose from dosimeter readings by applying dosimeter-to-human dose conversion factors under the assumption of a parallel radiation field, but this can introduce significant errors when the radiation field is inhomogeneous.
In this study, Deep Neural Networks (DNN) was applied to rapidly estimate absorbed doses to humans and dose conversion factors in inhomogeneous radiation fields. It was assumed that the radiation field can be described by the location and energy distribution of point sources, and the basic exposure scenario was set to external exposure by a standing adult male from a point source. Through GEANT4-based Monte Carlo simulations, absorbed doses to radiation-sensitive organs and whole-body were calculated, and conversion factors between chest-worn dosimeters and organ doses were determined.
Due to significant skewness in dose data, statistical techniques that transform the data to approximate a normal distribution, such as log transformation and Box-Cox transformation, to facilitate more effective training. The transformed dataset was divided into training, validation, and test sets. Optimization of model hyperparameters was performed using training and validation data and optimized model was trained. The model's predictive performance was verified by evaluating its relative error rate in predicting test data compared to ground truth values. The prediction results demonstrated an acceptable relative error rate, factoring in the uncertainty inherent in simulation-derived data. The results of this study are expected to provide a foundation for easily and quickly assessing the predicted risk to the human body when radiation exposure occurs in an unexpected inhomogeneous source distribution situation. This will help to quickly determine follow-up measures.