| Autism spectrum disorders(ASD)and Parkinson’s disease(PD)and other neurological disorders of the brain cause severe emotional and financial stress to patients and their families.Currently,clinical diagnosis mainly relies on physicians’ experience and scale scores,and this manual diagnosis lacks a quantitative diagnosis,which is prone to miss diagnosis and misdiagnosis.In recent years,magnetic resonance imaging(MRI),a non-invasive neuroimaging technique,has become the diagnostic aid of choice for neurological brain diseases such as ASD and PD,as it helps researchers to study the brain quickly and effectively from the perspectives of both abnormal functional brain activity as well as structural tissue alterations.Brain network feature based on MRI can reflect functional or structural connectivity changes in different regions of interest(ROI)of brain,so it is important to carry out research on classification methods based on brain network feature.In the classification method based on brain network feature,feature optimization and classifier construction are important aspects.Since brain network features are high-dimensional and nonlinear,how to enhance the feature representation of brain network feature from the perspective of feature mapping and feature selection deserves further study.In addition,there are many classification methods based on brain network feature,from which cascaded framework stand out in two main forms: cascading the same classifier and cascading different classifiers.In particular,the method of cascading the same classifier mainly improves the classification accuracy by adjusting the parameters of the cascaded framework,and the method of cascading different classifiers mainly improves the classification accuracy by jointly utilizing several different classifiers.Nevertheless,the optimization of the classifier structure of these two forms of cascade framework to improve the classification accuracy of the model is a problem that needs further research.In this paper,the research is carried out in two aspects of brain network feature optimization and classifier improvement in this context,and the main work is as follows:(1)A cascaded classification method based on empirical kernel mapping(EKM)to enhance the expression of brain network feature is proposed to optimize the feature expression by introducing an EKM module to enhance the differentiability of brain network feature from the feature mapping perspective.At the same time,the cascaded framework method of cascading the same classifier is also improved by using self paced learning(SPL)to realize the improvement of the classification method in terms of classification accuracy.Compared to existing methods,the method proposed in this paper is validated on public datasets of two neurological brain disorders,ASD and PD,and improved in classification accuracy by 6.55% and 7.05%,respectively.(2)A cascaded classification method based on a priori ROI sparse brain network feature to optimize feature representation by designing a priori brain region sparse(PRI)module to realize the sparse brain network feature from the feature selection perspective.The edge weight coding(EWC)module is also designed to improve the cascaded framework method of cascading different classifiers to realize the classification method in terms of classification accuracy.Compared to existing methods,the proposed method is validated on a publicly available dataset of ASD and improved 13.41% in classification accuracy. |
