Capability of deep learning to predict recoil protons for neutron dosimetry with Fluorescent Nuclear Track Detectors

Fluorescent Nuclear Track Detectors (FNTDs) provide high spatial resolution, wide linear energy transfer coverage, and reusability, making them well-suited for high-energy neutron dosimetry. When neutrons traverse a polyethylene converter, recoil protons are generated, and their tracks are stored inside the FNTDs and visualised through optical readout. Traditional analysis of FNTD images relies on deterministic algorithms or machine learning methods with explicit feature definition, limiting their general extension. In contrast,
deep learning networks can extract image features enabling generalisation across different neutron energy spectra and dose values. In this study, a deep learning network was trained on images of FNTDs irradiated at six mono-energetic neutron energies and tested on images of FNTDs exposed to a broad-spectrum $^{241}$Am-Be neutron source. Using raw images of irradiated FNTDs as input, the network predicted the proton tracks which were later counted. For the $^{241}$Am-Be test dataset, a dose-response curve of identified tracks over ambient dose equivalent was fitted, and the sensitivity in terms of 𝐻∗(10) was extracted from the slope. ... mehrWhen the fit was applied on the whole 𝐻∗(10) range, from 0 mSv up to 100 mSv, the predicted sensitivity for $^{241}$Am-Be was 𝑆$_{𝑝𝑟𝑒𝑑}$ = (2280 ± 20) tracks mSv$^{−1}$ cm$^{−2}$. The relative deviation of this predicted sensitivity from the reference sensitivity was 5.8 %. When the fit considered only the 𝐻∗(10) range of the training dataset, namely from 5 mSv to 15 mSv, the predicted sensitivity for $^{241}$Am-Be was 𝑆$_{𝑝𝑟𝑒𝑑}$= (2500 ± 60) tracks mSv$^{−1}$ cm$^{−2}$. This led to a relative deviation from the reference sensitivity of only 1.2 %. Despite being trained solely on mono-energetic data, the model successfully generalised to the $^{241}$Am-Be energy spectrum.