Comparative analysis of machine learning-based dose assessment algorithms for TL dosimetry

Abstract

This paper explores the implementation of machine learning-based algorithms for TL dose assessment. It focuses on the radiation field classification, performance quotient evaluation, and shallow and deep dose equivalent assessment of ANN and LGBM, in comparison to the traditional method of DT. We evaluate these algorithms based on the element response data measured by TLD. A data set was built for training, and the base element responses of test categories were amplified, and normalized to 1 mSv Cs-137 within the range of ±3 %. Both algorithms consist of five subset models for classifying radiation fields and identifying ratios of mixed fields. The LGBM showed the best accuracy in classifying considered radiation fields and the lowest performance quotients. By comparing the tolerance levels of deep dose and shallow dose equivalents among the three algorithms, the LGBM yields the smallest difference between the predicted and true dose equivalents. This smaller difference implies the LGBM offers the least bias and standard deviation in the expected value, giving higher accuracy and precision in dose assessment over the traditional DT method. The findings from this study further contribute to the adoption of ML-based algorithms for TL dose assessment, underscoring its importance in the field.