Voxel selection in MRI through bagging and conformal analysis: application to detection of obsessive compulsive disorder

In this work we apply a multivariate feature selection method based on bagging linear SVMs to construct a classifier able to differentiate among control subjects and patients with obsessive compulsive disorder (OCD). Our method selects sets of voxels that are relevant for the detection of the disease. The voxel selection is completed with a conformal analysis based refinement that controls over fitting and dramatically reduces the test error rate of the final classifier. Furthermore, the resulting discrimination pattern is organized in regions that show great agreement with those found by traditional methods used in OCD problems, achieving cleaner and more accurate region maps.