Extending information relevant for personal dose monitoring obtained from glow curves of thermoluminescence dosimeters using artificial neural networks

No more glowing in the dark: How deep learning improves exposure date estimation in thermoluminescence dosimetry

“…The TL-DOS readout automate however does not allow for a temperature measurement as a result of its design for high throughput routine dose monitoring. As small uncertainties deviations in the transformation step can lead to large uncertainties on the subsequent deconvolution, this step was found to be a driving factor of uncertainties in the previous studies (Theinert 2018, Kröninger et al 2019, Mentzel et al 2020). In addition, the underlying physical model of the glow curve deconvolution itself is still a field of open research (Eliyahu et al 2017(Eliyahu et al , 2014, making the development of robust fit routines for glow curves resulting from various irradiation scenarios an even more challenging task.…”

“…In a prior publication it was demonstrated that using a fully-connected deep neural network (DNN) with multiple glow curve parameters as input obtained from the glow curve deconvolution (GCD) after the temperature reconstruction can significantly improve the estimation precision of the irradiation date compared to a one-dimensional fit approach (Mentzel et al 2020). The univariate approach comprised a one-dimensional fit function of a single variable computed from multiple glow curve parameters in dependence of the irradiation date.…”

“…Figure 12: Comparison of prediction uncertainties for t pre of the new deep neural network trained with features derived from the time-resolved glow curve (green triangles pointing right) and the convolutional network model red squares) with the previously presented models: univariate analysis of features from glow curve analysis in temperature space (blue triangles pointing left, reproduced from (Mentzel et al 2020)), deep neural network trained with features from glow curve analysis in temperature space(orange circles, reproduced from (Mentzel et al 2020)). The plot below shows the respective improvement compared to the univariate baseline model in percent…”

“…The measured data presented in the following was used for the first time in (Mentzel et al 2020) for a glow curve analysis.…”

“…In prior work on the TL-DOS system it was reported, that multivariate analyses of the detector signal, the so-called glow curves, can supply additional insights into the circumstances of an irradiation. In recent studies, the date of a single exposure within a monitoring interval of 42 days was reconstructed with a prediction uncertainty of 2-4 days using neural networks (Theinert 2018, Kröninger et al 2019, Mentzel et al 2020.…”