基于多模态磁共振影像的精神分裂症患者多特征分类研究

Discriminative analysis of schizophrenia using multi-level features based on multi-modal magnetic resonance imaging

精神分裂症(SZ)的分类研究已经被广泛报道,但是之前比较多的研究都是基于单个模态的或者单个特征的。本研究提出了一种基于多模态磁共振影像的自动分类方法,采集了44例SZ患者与56例健康正常人的结构与功能磁共振影像,基于自动解剖标签模板提取了90个感兴趣区域的灰质体积、局部一致性、低频振荡振幅和度中心度作为分类模型的输入特征。为了降低特征维度,利用递归特征消除法获取对分类有效的主要特征,然后采用支持向量机对SZ患者和正常人进行分类。结果表明,将4种多模态特征组合起来作为分类特征时,分类准确率达到94.47%,明显优于单独将这些特征作为分类模型的输入特征时取得的分类效果,并且研究发现分类权重最大的脑区主要集中在额叶、颞叶和枕叶。研究结果有助于理解SZ患者的病理机制与进展规律及实现自动诊断。

The classification studies of schizophrenia（SZ） reported so far are mostly based on a single modality or a single characteristic. Herein the authors propose a new approach based on multi- modal magnetic resonance imaging（MRI） for automatically discriminating SZ patients from normal control subjects. The structural and functional MRI data of 44 SZ patients and 56 normal control subjects were acquired. Based on automated anatomical labeling atlas, we used the gray matter volume, regional homogeneity, amplitude of low frequency fluctuation, and degree centrality from 90 regions of interest as the input features in the classification model. To reduce the feature dimensions, a recursive feature elimination strategy was applied to determine the effective features for classification. A support vector machine was used to classify SZ patients and normal control subjects. The results showed that the classifier using the combination of all the features as the input features achieved a classification accuracy of 94.47%, and the performance of the proposed classifier was better than that of a classifier using the single- level features. The most discriminative features for classification are located mainly in the frontal, temporal and occipital lobes. The research results are conducive to understanding the pathogenic mechanism and developing computeraided diagnosis of SZ.