基于最小生成树的多特征融合的脑网络分类研究

f MRI classification method of multiple features fusion based on minimum spanning tree

针对传统图论方法设置阈值等问题,采用最小生成树方法进行脑网络的构建。传统的分类方法是从脑网络中提取一些可量化指标用于分类,忽视了多个脑区之间的拓扑信息。针对此问题,提出了一种脑区特征和连接模式相结合的多特征融合的分类方法,利用不同类型的特征来量化不同的网络性能。结果表明,抑郁症患者的最小生成树更趋向于随机网络,且局部属性出现显著异常的脑区集中在边缘系统—皮层—纹状体—苍白球—丘脑神经环路(limbic-cortical-striatal-pallidal-thalamic,LCSPT)。此外,与单一类型特征的分类方法相比,多特征融合的方法能够有效地提高分类精度。进一步分析表明,最小生成树方法可以用于抑郁症的辅助诊断,不同形式的特征表示方法具有信息描述方面的互补性。

To avoid complicated threshold setting problem by applying traditional graph theory method,this paper utilized the minimum spanning tree(MST)method to construct brain network.The traditional classification method extracted quantifiable attributes from the brain network for classification,ignoring the topology information between multiple brain regions.To overcome this problem,this paper proposed a novel method combining brain region and subgraph features for classification.This method utilized two different types of features to quantify two different properties of network.The result indicates that MST of depression are more similar to random networks,and exhibits significantly difference in some regions concentrating on the LCSPT circuit.In addition,the classification method of multiple features fusion can effectively improve the classification accuracy compared with the method using only single type of features.Further analysis demonstrates that the MST method can be used in the auxiliary diagnosis of depression.Meanwhile,different forms of feature representation have the complementary in information description.