Low-activity hotspot investigation method via scanning using deep learning

Small areas of elevated activity are a concern during a final status scan survey of residual radioactivity of decommissioned and contaminated sites. Due to the characteristics of scanning, the lower limit of detection is relatively high because the number of counts is low due to the short measurement time. To overcome this, an algorithm capable of finding hotspots with little information through deep learning was developed. The developed model using an artificial neural network was trained with the scan survey data acquired from a Monte Carlo-based computational simulation. A random mixing method was used to obtain sufficient training data. In order to respond properly to the experimental data, training and verification were conducted in various situations, in this case, in the presence or absence of random background counts and collimators and various source concentrations. Experimental data were obtained using a conventional detector, in this case, the 3″ × 3″ NaI(Tl). The advantages and limitations to the proposed method are as follows. Results were well predicted even in cases at less than 1 Bq/g, which is lower than the scanned minimum detectable concentration (MDC) of the detection system. It is a great advantage that it can detect contaminated areas that are lower than the existing scan’s minimum detectable concentration. However, the limitation is that it cannot be predicted, and the accuracy is low in multi-sourced scans. The source position and size are also important in residual radioactive evaluations, and scanning data images were evaluated in artificial neural network modes with suitable prediction results. The proposed methodology proved the high accuracy of hotspot prediction for low-activity sites and showed that this technology can be used as an efficient and economical hotspot scanning technology and can be extended to an automated system.