目的 结合多模态磁共振成像和人工智能技术,探索建立精神分裂症影像学辅助诊断工具.方法 利用3个独立数据集,基于每名被试的脑结构磁共振成像(structural magnetic resonance imaging,sMRI)数据提取灰质体积(grey matter volume,GMV)、白质体积(white matter volume,WMV)、皮质厚度(cortical thickness,CT)和基于形变场的形态学测量(deformation-based morphometry,DBM)指标,基于功能磁共振成像(functional magnetic resonance imaging,fMRI)数据提取低频振幅(amplitude of low frequency fluctuation,ALFF)、局部一致性(regional homogeneity,ReHo)以及功能连接(functional connectivity,FC)指标.为构建精神分裂症识别模型(即区分精神分裂症患者及健康对照),首先利用机器学习方法基于单指标构建分类器,再融合多模态指标构建融合分类器.分类器的训练及测试分别在数据集内以及跨数据集间进行交叉验证.结果 数据集内的交叉验证结果显示,单指标分类器的精神分裂症诊断准确率最高为86.18％(FC),而多模态指标融合分类器的精神分裂症诊断准确率最高可提升至90.21％.跨数据集的交叉验证结果显示,单指标分类器的精神分裂症诊断准确率最高为69.02％(ReHo),而融合分类器的精神分裂症诊断准确率最高为71.25％.结论 基于功能性指标来识别精神分裂症的性能普遍优于结构性指标,而融合多种模态指标可进一步提升分类准确率,且基于CT、DBM、WM、FC、ReHo五种指标所构建的融合分类器性能最高,具有作为精神分裂症影像学辅助诊断工具的潜力.
Imaging-assisted diagnostic model for schizophrenia using multimodal magnetic resonance imaging
Objective To develop an imaging-assisted diagnostic tool for schizophrenia based on multimodal magnetic resonance imaging and artificial intelligence techniques.Methods Three independent datasets were utilized.For each subject,four brain structural metrics including grey matter volume(GMV),white matter volume(WMV),cortical thickness(CT)and deformation-based morphometry(DBM)indica-tors were extracted from the structural magnetic resonance imaging(sMRI)data,and three brain functional metrics including amplitude of low frequency fluctuation(ALFF),regional homogeneity(ReHo)and func-tional connectivity(FC)were extracted from the functional magnetic resonance imaging(fMRI)data.To distinguish patients with schizophrenia and healthy controls,single-metric classification models and multi-metrics-fusion classification models were trained and tested using a within-dataset and a between-dataset cross-validation strategy.Results The results of within-dataset cross-validation showed that the highest ac-curacy of the single-metric classifications for schizophrenia diagnosis was 86.18％(FC),while the multi-metric-fusion classifications could reach an accuracy of 90.21％.The results of between-datasets cross-vali-dation showed that the highest accuracy of the single-metric classifications for schizophrenia diagnosis was 69.02％(ReHo),while the multi-metric-fusion classifications could reach an accuracy of 71.25％.Conclu-sion The functional metrics generally outperforms the structural metrics for the classification between pa-tients with schizophrenia and heathy controls.Additionally,fusion of multi-modal brain imaging metrics can improve the classification performance.Specifically,the fusion of CT,DBM,WMV,FC and ReHo demonstrates the highest classification accuracy,which is a potential tool for imaging-assisted diagnosis of schizophrenia.
SchizophreniaMachine learningStructural magnetic resonance imagingFunc-tional magnetic resonance imagingMultimodal magnetic resonance imaging
NETL
NSTL
万方数据
