Computerized categorization of TLD glow curve anomalies using multi-class classification support vector machines

Abstract A novel method for automatic categorization of thermoluminescent dosimeter (TLD) glow curve (GC) anomalies is presented. This automatic categorization will improve the metrological process of dose estimation by enhancing both its repeatability and its accuracy. Moreover, it will help external dosimetry laboratories to forecast some of the malfunctions of their TLD readers. A degenerated automatic approach was previously used in order to differentiate between a regular GC and an anomalous one, without being able to distinguish between different types of anomalies. That approach is now substantially extended to implicitly enable the categorization of GCs into five different kinds of anomalies. The machine learning algorithm applied for this purpose is support vector machines (SVM). The SVM algorithm categorizes TLD GCs into either a ‘good’ GC or into five types of TLD GC anomalies. When applied on an uncategorized GC, SVM associates it with a classification probability for each of the six categories. Results show an accuracy rate between 87.5% and 89% for the correct categorization of GCs to either of the six classes, depending on the presence of 'spikes' class in the data.