A team has used machine learning to localize gamma radiation sources inside a human phantom from spectra recorded by a sodium iodide scintillator detector activated with thallium, NaI(Tl). The work falls within radiobioassay, the discipline that quantifies radioactivity incorporated by the human body and that is essential for monitoring exposed workers and for responding to radiological emergencies.

The technical challenge is that a single NaI(Tl) spectrum is not by itself sufficient to determine the source position: the deposited energy depends on attenuation in tissues, Compton scattering and the geometry of the body itself. Conventional reconstruction methods usually require multiple measurements or assumptions about the activity distribution, which limits their use in real situations. The new approach trains machine learning models on simulated or measured spectra so that the network learns to associate the shape of the spectrum with the location of the source.

The strategy combines radiation transport physics with the ability of deep learning models to extract subtle patterns from spectra. By training on a broad set of configurations, the system can generalize and provide an estimate of the source position without requiring full body scans or costly imaging equipment. The use of NaI(Tl), an inexpensive scintillation material widely available in dosimetry laboratories, makes the proposal easily transferable to existing radiation protection infrastructure.

The authors highlight that the method improves accuracy over approaches based solely on analysis of the photopeak and that it can be extended to other detector geometries and different radioisotopes. The next planned step is to validate performance with experimental measurements on anthropomorphic phantoms and to broaden the training to spectra with higher background noise, with the goal of bringing the technique into operational radiological surveillance protocols.