结合表型信息的阿尔兹海默症图卷积神经网络分类方法研究

阿尔兹海默症(AD)的早期检测与发现具有重要的临床和社会意义.由于AD患者的功能性脑网络拓扑性质存在异常变化,并且不同表型类型人群中阿尔兹海默症的患病率也存在着较大差异,因此将脑网络特征和表型信息结合构建训练特征,用于阿尔兹海默症不同阶段的分类.同时,图卷积神经网络(GCN)分类方法被证明是目前对图数据学习任务的最佳选...

Study on Classification Method of Alzheimer's Disease Convolutional Neural Network Combined with Phenotypic Information

The early detection and diagnosis of Alzheimer's disease (AD) has important clinical and social significance. Because of the abnormal changes in the topological properties of the functional brain network in AD patients and the large differences in the prevalence of Alzheimer's in different phenotype types, this study combined brain network features and phenotypic information to construct training features for classification at different stages of Alzheimer's disease. In recent years, the graph convolutional neural network (GCN) classification method has proved to be the best choice for graph data learning tasks. Therefore, this paper applied GCN to the classification study of AD and completed for healthy control (CN), early mild cognitive impairment (EMCI), late mild cognitive impairment (LMCI) and AD four types of classification. Herein we used the basic framework of population map convolutional neural network to classify 300 subjects in the ADNI database, and improved methods were proposed in terms of the similarity between the subjects in the population graph and the characteristics of the brain network of the subjects. In terms of the similarity between graph subjects, the construction of other phenotypic graph structures was performed using methods such as addition, improved initial value, only feature similarity, only phenotypic similarity, and four other combination methods; in terms of a brain network feature, combined with multi-modal thinking, the phenotypic information was converted into binary features, and the brain network features were spliced into total features. In addition, this study tried to use different types of phenotype information to the experiments. Finally, 10-fold cross-validation was used to verify the results. The results showed that the improvement in both aspects increased the accuracy to a certain extent. The classification accuracy was the best while using brain network similarity as the edge weight of graph construction, and the phenotypic information (age or gender) without dimension reduction as the characteristics of subjects (nodes). Compared with the original method, the accuracy was improved from 80% to 82%. It was shown that the brain network features and phenotype information were important features in the brain disease classification task, which can help to improve the accuracy of classification task, therefore is of research implications.