A Semi-Supervised Model for Multi-Label Radioisotope Classification and Out-of-Distribution Detection

Abstract

In the machine learning problem of multi-label classification, the objective is to determine for each test instance which classes the instance belongs to. In this work, we consider multi-label classification in the context of multi-label radioisotope classification for gamma spectra data. By viewing spectra as discrete distributions we tackle a more challenging variant of multi-label classification where the goal is to ascribe a proportion to each class label, not just a binary variable. Motivated by this application to radioisotope identification, we aim to simultaneously predict label proportions while also performing out-of-distribution (OOD) detection. To achieve this goal, we introduce a novel semi-supervised loss function that combines a traditional supervised loss with an unsupervised reconstruction error penalty. This work demonstrates that the proposed model can successfully perform radioisotope identification in a realistic test scenario. We also show how to extend this approach to perform OOD detection which can determine when the model prediction should not be trusted due to the presence of an anomalous source. The semi-supervised model, trained on gamma spectra based on a measurement of a real fission source containing a mixture of 30 distinct radioisotopes (labeled by a spectroscopist), learned to estimate in-distribution (ID) samples with about 39% error while simultaneously being able to differentiate (with 95% confidence) between ID and OOD samples with an anomaly contribution as small as 10%.