Brain Network Classification Based On Topological Information

Mild cognitive impairment(MCI)is the early stage of Alzheimer's disease(AD),and MCI patients will be in this stage for a long time before the onset of dementia.Therefore,the timely diagnosis of mild cognitive impairment provides an important means to prevent the deterioration of disease.Medical imaging technology can obtain structural or functional brain image more conveniently and quickly,and then construct the brain connection network,which can be used for analysis and diagnosis.Machine learning can analyze the difference of local or global topological structure between MCI patients and normal people brain networks,and then use the differences to distinguish the unknown brain networks,which has become an important method to diagnose mild cognitive impairment.This paper has conducted an in-depth study of the brain network classification method based on topological structure information,and mainly carried out the following aspects: First,a classification method based on the fusion of multiple topological attributes from the perspective of local node attributes is proposed.The classification effects of betweenness centrality,eccentricity and degree centrality of weighted brain networks and unweighted brain networks under different thresholds are analyzed,and the dominant topological attributes of each type of brain network under the optimal threshold are selected by feature selection for integration to fully reflect the difference between the local topological structure of the brain network of MCI patients and normal people;Secondly,a feature selection algorithm based on correlation is proposed to remove redundant brain region features in the same topological attribute;Thirdly,the disease can cause changes in the connection relationship between brain regions.Therefore,the number of occurrences of a certain shortest path that can reflect the connection relationship in the two types of sample groups are significantly different.Use the difference to select the discriminative brain regions and construct the key subnet,and then use Weisfeiler-Lehman subtree kernel for similarity discrimination,which can effectively diagnose the disease.In order to verify the effectiveness of the above methods,the acquired brain images of MCI patients and normal people were preprocessed and the brain networks were constructed,and the following comparative experiments were carried out.The fusion method of multiple topological attributes under the weighted and unweighted brain network is compared with the method of classification using only the topological attributes calculated by the unweighted brain network,and the classification accuracy is significantly improved.Compared with the method that does not remove redundant features,the proposed feature selection algorithm based on correlation improves the classification accuracy by 4.5%.The weighted brain network and the unweighted brain network that implement the key subnet selection algorithm based on the shortest path,and the whole brain network that does not implement the algorithm made a comparison from the two aspects of classification effect and the number of iterations of Weisfeiler-Lehman subtree kernel.The results show that the algorithm can effectively capture the abnormal structure of brain network caused by the disease,which is beneficial to the diagnosis of disease.The comparisons of the experimental results of the weighted sub-network and the unweighted sub-network obtained after the implementation of the algorithm show that the weight information representing the strength of interaction between brain regions can affect the calculation of the shortest path,thereby affecting the capture of abnormal structures.