A novel fuzzy rough selection of non-linearly extracted features for schizophrenia diagnosis using fMRI

Schizophrenia is a severe brain disorder primarily diagnosed through externally observed behavioural symptoms due to the dearth of established clinical tests. Functional magnetic resonance imaging (fMRI) can capture the distortions caused by schizophrenia in the brain activation. Hence, it can be useful for developing a decision model that performs computer-aided diagnosis of schizophrenia. But, fMRI data is huge in dimension. Therefore dimension reduction is indispensable. It is additionally required to identify the discriminative brain regions. Hence, we aim to build an effective decision model that incorporates suitable dimension reduction and also identifies discriminative brain regions.

We propose a three-phase dimension reduction. First phase involves spatially-constrained fuzzy clustering of 3-dimensional spatial maps (obtained from general linear model and independent component analysis). In the second phase, non-linear features are extracted from each cluster using a generalized discriminant analysis. In the third phase, a novel fuzzy rough feature selection is proposed. The features obtained after the third phase are used for learning a decision model by the help of support vector machine classifier. This complete method is implemented within leave-one-out cross-validation on two balanced datasets (respectively acquired on 1.5Tesla and 3Tesla scanners). Both these datasets are created using Function Biomedical Informatics Research Network multisite data and contain fMRI data acquired during auditory oddball task performed by age-matched schizophrenia patients and healthy subjects. A permutation test is also carried out to ensure that no bias is involved in the learning.

The results indicate that the proposed method achieves maximum classification accuracy of 97.1% and 98.0% for the two datasets respectively. The proposed method outperforms the state-of-the-art methods. The results of the permutation test show that p-values are lesser than the significance level i.e. 0.05. Therefore, the classifier has found a significant class structure and does not involve any bias. Further, discriminative brain regions are identified and are in agreement with the findings in related literature.

The proposed method is able to derive suitable non-linear features and the related brain regions for effective computer-aided diagnosis. The fuzzy and rough set based approaches help in handling uncertainty and ambiguity in real data.