Classifier Comparison For Radionuclide Identification With Gamma-ray Spectra

John-Ryan R Romo - Lawrence Livermore National Laboratory
Kai Tyrus Nelson - Lawerence Livermore National Laboratory
Mateusz Monterial - Lawerence Livermore National Laboratory
Karl E Nelson - Lawerence Livermore National Laboratory
Adam Hecht - University of New Mexico
Simon E Labov - Lawerence Livermore National Laboratory
File Attachment
a457.pdf1.41 MB
Classifier Comparison for Radionuclide Identification with Gamma-ray SpectraJohn-Ryan Romo, Kai Nelson, Mateusz Monterial, Karl Nelson, Adam Hecht, Simon Labov Machine learning is increasingly applied for gamma-ray spectra analysis, particularly gamma-ray identification. Radionuclide identification is a multiclass multilabel classification problem which can be tackled with methods including artificial neural networks, extreme gradient boosted tress (XGBoost) and Random Forest. In this study, we compare the performance of these classifiers using the Benchmark Algorithm for Radionuclide Identification (BARNI) as a framework for processing raw spectra into pre-selected features extracted using a peak search algorithm. Sampled spectra from a 3”x3” NaI detector with a library of 33 radionuclides under a spanning set of shielding configurations was used for training and testing. The overall performance of each classifier was assessed using the F1-score. We also break down the per nuclide performance, and demonstrate which classifiers are better for identification of particular radionuclides. The experimental results show that XGBoost was the optimal choice when performance and training time was valued. The Multi-Layer Perceptron (MLP) neural network showed potential as a runner up although lacked the proper amount of optimization and tuning in order to surpass the XGBoost classifier. The Random Forest classifier was the weakest perform in both performance and training time.