Comparing the Performances of PDF and PCA on Parkinson's Disease Classification Using Structural MRI Images

Detection and diagnosis of neurodegenerative diseases, including Parkinson's disease, using computer-aided technology have been widely studied, recently. In this paper, automatic classification of Parkinson's disease by using structural magnetic resonance imaging have been investigated. Voxel-based morphometry technique is used to compare the local and global differences of each grey matter and white matter with Parkinson's disease versus healthy controls. The differences are considered as volumes of interests. VBM is used with DARTEL in order to increase inter-group registration, provide more precise, and accurate localization of structural differences of sMRI data. In order to generate 3D masks both for gray as well as white matter modalities, the general linear models are configured and two sample t-test based statistical method is used. Principle component analysis and probability distribution function based feature selection methods are used to select the volumes of interest. Support vector machine with Gaussian kernel is taken into account for classification of 40 Parkinson's disease patients and 40 healthy controls. The experiments are investigated to compare the performances of two methods with original and raw data for white and gray matter modalities. Since the number of voxels in obtained clusters are a few, some of which may be counted as noise and hence amount of extracted features are not extremely high, using principle component analysis for 3D mask data gives better results than probability distribution function based feature selection methods. It removes unimportant voxels, whereas histogram technique uses their impacts.