Supervised Classification of Mean Diffusivity in Substantia Nigra for Parkinson's Disease Diagnosis

The substantia nigra (SN) is one of the brain regions commonly affected by neurodegenerative diseases, including Parkinson's disease (PD). The neurodegenerative process could be characterized through diffusion tensor imaging (DTI). Recent DTI studies corroborated this assumption that significantly lower fractional anisotropy (FA) and higher mean diffusivity (MD) values are observed in the SN in patients with PD compared to healthy control (HC) subjects. Since the accuracy of diagnosis based on clinical criteria is low, an automated learning-based method such as a support vector machine (SVM) is required to diagnose the disease based on FA and MD signal alternation. The aim of this study was to optimize and compare five commonly used classifications methods, including K-nearest neighbor (KNN), support vector machine (SVM), tree classification, Naïve Bayes, and ensemble classifiers to find the best technique to discriminate early PD from HC based on MD values in the SN. K-nearest neighbor (KNN) and Ensemble classifier outperformed other classification methods in the diagnosis of PD with 81%, 78% accuracy, and 78%, 82% sensitivity, respectively. Our results showed that although SVM in most cases exhibited good performance in classifying PD patients based on DTI data, for classification of the PD and HC based on the MD values of left and right SN, KNN performed better. In this regard, the Ensemble method with an AUC of 75% exhibited superior performance to SVM.