面向脑网络的新型图核及其在 MCI 分类上的应用

摘要

Graph kernel,as a similarity measure of graphs,has been proposed for computing the similarity of a pair of brain networks and applied for classification of brain diseases,such as Alzheimer’s disease (AD)as well as its early stage,i.e.,mild cognitive impairment (MCI). However,existing graph kernels are mainly constructed on general graphs and thus ignore the intrinsic property of brain networks,such as the uniqueness of each node,i.e.,each node corresponds to a unique brain regions,which may affect the performance of brain network analysis (classification).To address this problem,in this paper,a novel graph kernel is proposed for measuring the similarity of brain networks.Specifically,a group of sub-networks are first constructed on each node to reflect the local and multi-level topological properties of brain network.Then, according the uniqueness of each node,a function is defined to measure the similarity of a pair of sub-network groups across different subjects.Finally,the graph kernel on brain network can be defined through computing the similarity of all pairs of sub-network groups.Different from existing graph kernels,our proposed graph kernel not only considers the specific property of brain networks,but also preserves the local connectivity properties of brain networks.The experimental results on both real MCI datasets show that our proposed graph kernel can significantly improve the classification performance in comparison with state-of-the-art graph kernels.%作为一种图的相似性度量，图核已经被提出用于计算脑网络的相似性，并用于分类一些脑疾病，如阿尔茨海默病（Alzheimer’s Disease，AD）以及它的早期阶段，即轻度认知功能障碍（Mild Cognitive Impairment，MCI）。然而，已有图核主要面向一般图而构建，从而忽略了脑网络自身特有的特性，如节点的唯一性（即每个节点对应着唯一的脑区），这可能影响到脑网络分析（分类）性能。为了解决这个问题，构建一种面向脑网络的图核，用于测量一对脑网络的相似性。具体而言就是：首先，以网络中每一个节点为中心，构建一组子网络来反映网络的局部多层次拓扑特性。而后，利用节点的唯一性，构建测量每对子网组之间相似性函数，从而获得用于测量一对脑网络的相似性的图核。不同于已有的图核，提出的图核充分考虑到脑网络自身特有的特性，以及保留了脑网络局部连接特性。在两个真实的 MCI 数据集上，实验结果表明，相对于现阶段的图核，文中提出的图核能够显著提高分类的性能。