Accurate diagnosis of schizophrenia is of great importance to patients and clinicians. Recent studies have found that different frequency bands contain complementary information for diagnosis and prognosis. However, conventional multiple frequency functional connectivity (FC) networks using Pearson’s correlation coefficient (PCC) are usually based on pairwise correlations among different brain regions on single frequency band, while ignoring the interactions between regions in different frequency bands, the relationship among different networks, and the nonlinear properties of blood-oxygen-level-dependent (BOLD) signal. To take into account these relationships, we propose in this study a multiple networks fusion method for schizophrenia diagnosis. Specifically, we first construct FC networks within the same and across frequency from the resting-state functional magnetic resonance imaging (rs-fMRI) time series by using extended maximal information coefficient (eMIC) based on four frequency bands: slow-5 (0.01–0.027 Hz), slow-4 (0.027–0.073 Hz), slow-3 (0.073–0.198 Hz), and slow-2 (0.198–0.25 Hz). Then, these networks are combined nonlinearly through network fusion, which generates a unified network for each subject. Features extracted from the unified network are used for final classification. Experimental results demonstrated that the interaction between distinct brain regions across different frequency bands can significantly improve the classification performance, comparing with conventional FC analysis based on specific or entire low-frequency band. The promising results suggest that our proposed framework would be a useful tool in computer-aided diagnosis of schizophrenia.

The flowchart of proposed classification framework.

精神分裂症的准确诊断对患者和临床医生非常重要。最近的研究发现，不同的频带包含用于诊断和预后的补充信息。但是，使用Pearson相关系数（PCC）的常规多频功能连接（FC）网络通常基于单个频带上不同大脑区域之间的成对相关性，而忽略了不同频带之间区域之间的相互作用，不同网络之间的关系，以及血氧水平依赖性（BOLD）信号的非线性特性。考虑到这些关系，我们在本研究中提出了一种用于精神分裂症诊断的多网络融合方法。特别，我们首先通过使用基于四个频段的扩展最大信息系数（eMIC），在静止状态功能磁共振成像（rs-fMRI）时间序列的相同频率和跨频率范围内构建FC网络：慢5（0.01–0.027 Hz ），慢4（0.027-0.073 Hz），慢3（0.073-0.198 Hz）和慢2（0.198-0.25 Hz）。然后，这些网络通过网络融合进行非线性组合，从而为每个主题生成一个统一的网络。从统一网络中提取的特征用于最终分类。实验结果表明，与基于特定或整个低频频带的常规FC分析相比，跨不同频带的不同大脑区域之间的交互作用可以显着改善分类性能。

拟议分类框架的流程图。