Evaluation of artificial neural network-based dose estimation models for emergency dose assessment in external exposure scenarios

Accurate and rapid assessment of individual dose in radiological emergencies is crucial. Conventional Monte Carlo (MC) simulation-based dose reconstruction offers high accuracy but is time-consuming and labor-intensive, while pre-calculated dose conversion coefficients are fast but lack flexibility for various exposure conditions. To overcome these limitations, this study developed an artificial intelligence (AI) based dose estimation and conversion model for external exposure scenarios. We generated 700,000 exposure datasets using the Geant4 code for a 1 MeV point source randomly positioned within a 3 m radius around the International Commission on Radiological Protection adult male mesh-type reference computational phantom. Based on this data, an artificial neural network (ANN) model was trained to estimate whole-body dose, organ doses, and absorbed doses in fortuitous dosimeters within personal belongings. The ANN model's performance was systematically optimized against various factors, including batch size, data volume, model architecture, etc. The developed ANN model showed a high correlation with the MC simulation results. Notably, the whole-body dose estimation model demonstrated high accuracy, with a 1st–99th percentile relative difference range of ±0.4%. Model performance was highly dependent on the statistical noise of MC results, and the accuracy of organ dose estimation tended to decrease for smaller organ masses. Although some high-error outliers were observed in the dose estimations of fortuitous dosimeters, outlier reduction techniques consisting of source position uncertainty and En-value evaluation effectively suppressed the outlier ratio to below 0.004% in the test set. In conclusion, the AI-based dose estimation model proposed in this study presents a novel methodology that satisfies the speed and flexibility required in radiological emergencies.