Machine learning for advancing gamma source localization in a human phantom using NaI (Tl) detector spectra in radiobioassay

This study introduces a machine learning (ML) based framework for the anatomical localization of gamma radiation sources in a human phantom (i.e., RMCII) using energy spectra from a NaI (Tl) scintillation detector, aiming to enhance internal dose assessment accuracy in radiobioassay. Spectra from Co-60, Cs-137, and Ba-133 placed in GI, lung, thyroid, and whole-body (WB) organs were recorded using a whole-body counter. After background subtraction and Min–Max normalization, ten statistical features (mean, skewness, kurtosis, entropy, interquartile range, etc.) have been extracted from each spectrum. A modular classification strategy trained four ML models (i.e., logistic regression (LR), decision tree (DT), support vector machine (SVM), multilayer perceptron (MLP)) for each isotope-detector configuration. LR achieved 100% localization accuracy when the detector was positioned in front of the GI organ with thyroid spectra excluded, while DT achieved up to 97.4% under the same conditions. However, performance dropped for Ba-133 due to spectral overlaps and attenuation, and for the WB detector position due to geometric symmetry. SVM and MLP performed poorly due to limited data and feature sensitivity. This modular, feature-based approach demonstrates the feasibility of ML for source localization in constrained radiobioassay settings, with LR and DT emerging as the most reliable classifiers, while highlighting challenges for low-energy isotopes and symmetric geometries that require future investigation.