LVQ and SVM classification of FDG-PET brain data

Abstract

We apply Generalized Matrix Learning Vector Quantization (GMLVQ) and Support Vector Machine (SVM) classifiers to fluorodeoxyglucose positron emission tomography (FDG-PET) brain data in the hope to achieve better classification accuracies for parkinsonian syndromes as compared to the decision tree method which was used in previous studies. The classifiers are validated using the leave- one-out method. The obtained results show that GMLVQ performs better than the previously studied decision tree (DT) method in the binary classification of group comparisons. Additionally, GMLVQ achieves a superior performance over the DT method regarding multi-class classification. The performance of the considered SVM classifier is comparable with that of GMLVQ. However, in the binary classification, GMLVQ performs better in the separation of Parkinson’s disease subjects from healthy controls. On the other hand, SVM achieves higher accuracy than the GMLVQ method in the binary classification of the other parkinsonian syndromes.