Ordinal Pattern Based Brain Network Classification With Applications

Recently, with the development of medical imaging techniques, it has attracted much attention to study brain disease based on neuroimaging data. Brain network, which is constructed from neu-roimaging, is a new powerful way to describe the function and structure of human brain. The brain network based classification is very important for automatic diagnoses and analyses of brain diseases (Alzheimer's disease, mild cognitive impairment, attention deficit hyperactivity disorder, etc.) and also is the main content of this paper. Specifically, the main innovation and work of this paper are summa-rized as follows:First, a new kind of feature named ordinal pattern for brain networks is proposed. An ordinal pattern is constructed by some ordinal rules of weighted edges. Different from existing brain network features, ordinal patterns have two advantages. Firstly, ordinal patterns are defined on weighted brain networks directly and no need for thresholding, which can highly save the weights information. Secondly, or-dinal patterns contain local topological structures of brain networks, which can reflect local structure disorders. Furthermore, based on the Apriori property of ordinal patterns, we propose an algorithm to effectively mine frequent ordinal patterns from brain network set. In two real brain disease datasets, we test the efficiencies of our proposed algorithm. The frequent ordinal patterns represent common information in a brain network set, which have important significances for brain network classifications and analyses.Second, a discriminative ordinal pattern based brain network classification method is proposed. Specifically, ratio score function is applied to measure the discriminative powers of frequent ordinal patterns. And the frequent ordinal patterns with higher ratio scores are selected as discriminative ordi-nal patterns. After that, based on discriminative ordinal patterns, we construct the feature matrix of all samples. Then, support vector machine is applied to build classifiers for brain diseases. Three real brain disease datasets are adopted to evaluate our proposed discriminative ordinal pattern based classifica-tion method. The classification results show that, comparing with existing features based classification method, our proposed method can achieve better performance. Furthermore, the detailed analyses of discriminative ordinal patterns show that, the abnormal brain regions we find are all in line with existing studies.