Diffusion tensor imaging (DTI) provides rich information about brain tissue structure especially in the white matter, which is known to be affected in several diseases like schizophrenia. Identifying patterns of brain changes induced by pathology is therefore crucial to clinical studies. However, the high dimensionality and complex structure of DTI make it difficult to apply conventional linear statistical and pattern classification methods to identify such patterns. In this paper, we present a novel framework that uses a combination of DTI-based anisotropy and geometry features to effectively identify brain regions with pathology-induced abnormality, and to classify brains into the diseased and healthy groups. Our method first directly estimates the underlying overlap between the patient and control groups, based on a semi-parametric Bayes error estimation method. By ranking voxels based on these estimation results, the method identifies abnormal brain regions from which features are extracted through Kernel Principal Component Analysis (KPCA) for subsequent classification. Application of the method to a dataset of controls and patients with schizophrenia, demonstrates promising accuracy of this framework in identifying brain patterns to separate two groups, and hence aiding in prognosis and treatment.