Automatic detection of anomalous thermoluminescent dosimeter glow curves using machine learning

Abstract

Computerized Glow Curve Analysis (CGCA) has been, and still is, an intensively-investigated subject for the past two decades. CGCA has applied different methods for various applications, from glow curve deconvolution into isolated peaks, through semi-automatic software tools for detecting outliers, to software that discovers exceptional curves by using predefined rules. The method presented herein addresses the subject using a new approach in which a completely automatic algorithm is used for the accurate detection of anomalies in thermoluminescent dosimeter (TLD) glow curves. A Support Vector Machines (SVM) technique, which is a machine learning classification algorithm, is used for the first time for radiation dosimetry applications. The algorithm classifies glow curves into two categories: an acceptable i.e. 'regular' curve, or a curve that exhibits any kind of anomaly i.e. an 'anomalous' curve. The classification method treats the glow curves raw data as a large ensemble of statistical data, and tries to identify exceptional glow curve shapes by statistical means. This classification method is performed in three steps. First, a library of glow curves is manually classified by a human user of the system into the above two classes. Then an iterative training algorithm is applied to these glow curves. The final stage applies a method of comparison between an unidentified glow curve and these two pre-classified sets, and assesses a classification probability to each of the two classes. The results show between 96.2% and 97.7% accuracy of the correct classification to either one of the classes, depending on the admissible false negatives rate.