Feasibility study on dose conversion using a deep learning algorithm for retrospective dosimetry

Abstract

The application of deep learning-based artificial intelligence (AI) models for dose estimation has garnered significant attention in dosimetric research, aiming to supplement or replace Monte Carlo (MC) particle transport simulations. The present study explores the feasibility of AI-based dose conversion techniques for retrospective dosimetry in radiological emergencies, particularly focusing on scenarios where a rapid estimation of body dose is required using measured doses from fortuitous dosimeters placed on or near the body during exposure. With the modeling of an International Commission on Radiation Units and Measurements (ICRU) slab phantom (presuming a human body) and glass plates (presuming fortuitous dosimeters), a large amount of dose data was generated through MC simulations with respect to randomly generated point sources (¹⁹²Ir, ¹³⁷Cs, and ⁶⁰Co) within a radius of 3 m from the phantom center. A deep learning (DL) model was trained to estimate doses and dose conversion coefficients (DCCs) between the phantom and the glass plates using the input of exposure structures, i.e. the position and energy of the source. Data scaling, such as logarithmic or power transformations, was essential for the dose data due to its highly biased distribution. The results showed that 98% of the estimated doses had relative differences (RDs) within ±3% when compared to MC simulations. To assess the impact of data volume on performance, datasets of varying sizes (55 k, 108 k, 216 k, and 432 k) were used for training, revealing a strong dependence of model performance on data volume. Outlier reduction methods, such as dose averaging and data reduction near the center, were applied, reducing the max-min RD range by a factor of 3–10. From the results, the potential and necessity of an AI dose estimation model for more complicated geometries, such as those involving anthropomorphic phantoms, were discussed.