fMRI based computer aided diagnosis of schizophrenia using fuzzy kernel feature extraction and hybrid feature selection

Functional magnetic resonance imaging (fMRI) is a useful technique for capturing deformities in brain activity patterns of several disorders. Schizophrenia is one such serious psychiatric disorder that, in absence of any standard diagnostic tests, is detected from behavioural symptoms observed externally. Thus, fMRI can be used for building an effective decision model for computer aided diagnosis of schizophrenia. However, fMRI data has huge dimension compared with the number of subjects; therefore it is essential to reduce the data dimension to avoid poor generalisation performance of the decision model. In the present work, we propose a three-phase dimension reduction that comprises of segmentation of voxels of 3-D spatial maps (independent component score-maps or β-maps) into anatomical brain regions; feature extraction from each region using a novel fuzzy kernel principal component analysis; and a novel hybrid (filter-cum-wrapper) feature selection for determining a reduced subset of discriminative features. These features are used as input to support vector machine classifier for learning a decision model. The method is carried out within leave-one-out cross-validation. Classification accuracy, sensitivity, and specificity are utilised to estimate the performance on two different balanced datasets D1 and D2 (respectively acquired on 1.5 T and 3 T scanners). Both the datasets contain fMRI data of age-matched healthy subjects and schizophrenia patients for auditory oddball task, obtained from FBIRN multisite dataset. The proposed method attains best classification accuracy of 95.6% and 96.0% for D1 and D2 respectively. The proposed method shows enhanced performance over the state-of-the-art methods. Further, the discriminative brain regions identified are in accordance with the findings in related literature and may be used as potential biomarkers.