Whole brain volume and cortical thickness based automatic classification of wilson's disease

Wilson's disease (WD) is a progressive autosomal-recessive genetic disorder of copper metabolism that can induce cognitive, physical and psychiatric symptoms. Despite the wide use of machine learning methods in neuroimaging analysis, WD-related research has been very rare. In this work, we proposed and validated an efficient pipeline for an automated WD classification based on whole brain segmentation volumes and cortical thicknesses obtained from T1-weighted magnetic resonance images (MRIs). Three well-known supervised machine learning algorithms, including support vector machine (SVM), linear discriminant analysis (LDA), and logistic regression (LR), were evaluated and compared in the setting of WD classification. A total of 51 images, including 27 acquired from WD patients and 24 from age-matched healthy controls, were used in the validation analysis. Univariate feature selection was conducted to eliminate non-relevant features and retain features contributing to the classification performance. Two nested leave-one-out cross validations were adopted, with the inner folder used for optimal parameter estimation and the outer folder for classification performance evaluation. Experimental results showed that when employing volume features, SVM significantly outperformed both LDA and LR, yielding an overall accuracy of 96.1%, a sensitivity of 92.6% and a specificity of 100%. LR could also reach such best classification performance when using a combination of volumes and thicknesses as input features. This study provides a new non-invasive tool (MRI-based) for an automated detection of WDs.