基于频率排序的判别子图筛选及在精神分裂症分类中的应用

Discriminant Subgraph Screening Based on Frequency Sorting and Its Application to Schizophrenia Classification

【目的】从脑网络中提取准确的判别性特征作为分类特征,可以提高SCZ的分类准确率,避免依靠量表的主观诊断造成缺陷。传统的脑网络特征如介数、聚类系数等丢失了拓扑信息,最小生成树丢失了部分脑区连接,子图虽然保留了拓扑信息,但传统的判别子图的筛选会产生部分冗余信息,进而影响分类准确率。【方法】提出一种基于频率排序的判别子图的筛选方法(frequently scoring screen, FSS),使用FSS筛选判别子图,在不损失原有判别信息的情况下,消除冗余信息。使用OpenfMRI的公开数据,对比了使用不同特征和不同分类算法的分类性能。【结果】FSS特征的分类性能优于其他传统脑网络特征,且该特征不受分类算法影响,分类准确率优于已有SCZ分类文献。

【Purposes】Studies of brain networks in schizophrenia(SCZ)have shown that both structural and functional networks are altered in patients.Extracting accurate discriminative fea-tures from brain networks as classification features can improve the classification accuracy of SCZ and avoid the deficiencies caused by subjective diagnosis relying on scales.Traditional brain net-work features such as betweenness centrality and clustering coefficients lose topological informa-tion,and minimum spanning tree loses some brain region connections.Although subgraphs re-tain topological information,the screening of traditional discriminative subgraphs generates some redundant information,which in turn affects the classification accuracy.【Methods】In this paper,a screening method for discriminant subgraphs based on frequency ranking(Frequency Scoring Screen,FSS)is propsed.FSS screens discriminative subgraphs and eliminates redundant information without losing the original discriminative information.The classification performance of using different features and different classification algorithms is compared by using publicly avail-able data from OpenfMRI.【Findings】The results show that the classification performance of the FSS feature is better than that of other traditional brain network features,the feature is not affected by the classification algorithm,and the classification accuracy is better than existing SCZ classification literatures.