摘要
目的基于功能磁共振成像(functional magnetic resonance imaging,fMRI)数据构建深度学习分类模型,以辅助诊断双相障碍患者,并分析双相障碍关键影像学特征,提高双相障碍识别率。方法收集符合DSM-Ⅳ诊断标准的双相障碍患者146例(患者组)以及健康对照者234名(对照组),进行fMRI扫描。采用局部一致性(regional homogeneity,ReHo)和低频振幅2种方法分析fMRI数据。基于ReHo和低频振幅指标分别采用深度神经网络(deep neural networks,DNN)和双通道卷积神经网络(dual-channel convolution neural networks,DCNN)构建分类模型,并通过比较分类准确率、受试者工作特征曲线曲线下面积获得最佳分类模型;采用准确率较高的影像指标对基于自动解剖标记图谱(anatomical automatic labeling,AAL)的90个大脑区域使用支持向量机(support vector machine,SVM)算法构建基于单个脑区的分类模型,并通过比较准确率指标,识别双相障碍的关键影像学特征。结果基于ReHo和低频振幅指标构建的DCNN分类模型的准确率分别为75.3%和72.6%,优于同指标下准确率分别为67.1%和65.1%的DNN分类模型,且使用ReHo指标构建的分类模型准确率相对优于低频振幅指标;同时基于SVM分类模型使用ReHo指标显示枕叶(枕中回、枕上回、舌回)、海马、丘脑等为识别双相障碍的关键脑区,且准确率均高于65.0%。结论基于ReHo指标的DCNN分类模型可用于双相障碍的辅助诊断;同时枕叶、海马、丘脑可能是辅助识别双相障碍的关键影像学特征脑区。
Objective Construction of deep learning classification models based on functional magnetic resonance imaging(fMRI)data assists the clinicians to achieve better diagnosis of bipolar disorder(BD),which can improve the recognition rate of BD by identifying the critical imaging features.Methods A total of 146 patients who met the diagnosis criteria of BD according to DSM-Ⅳand 234 healthy control(HC)were recruited for fMRI scans.Regional homogeneity(ReHo)and amplitude of low frequency fluctuation(ALFF)were used to analyze fMRI data.Based on ReHo and ALFF,the classification models were constructed by deeping neural network(DNN)and dual-channel convolution neural network(DCNN)respectively,and the best classification model was developed by comparing the accuracy and area under curve(AUC)of the two models.Based on each brain region divided by anatomical automatic labeling(AAL),the support vector machine(SVM)classification model was constructed using imaging index with a better performance,and the critical imaging features were identified by comparing the accuracy of each brain region.Results The performances of the DCNN classification model(accuracy=75.3%,and 72.6%,respectively,based on ReHo and ALFF)were significantly better than the DNN classification model(accuracy=67.1%,and 65.1%,respectively).Meanwhile,the accuracy of classification model constructed using ReHo was higher than ALFF.Based on the SVM classification model,critical brain regions were identified above the accuracy of 65.0%,including the occipital lobe(middle occipital gyrus,superior occipital gyrus and lingual gyrus),hippocampus,and thalamus.Conclusion The computational model based on DCNN using ReHo can help the clinicians to achieve better diagnosis of BD.Furthermore,occipital lobe,hippocampus and thalamus may be the critical imaging features for the auxiliary recognition of BD.
作者
魏鑫茹
段佳
张然
杨景钰
张陆衡
姚菲
董帅
张锡哲
王菲
朱荣鑫
Wei Xinru;Duan Jia;Zhang Ran;Yang Jingyu;Zhang Luheng;Yao Fei;Dong Shuai;Zhang Xizhe;Wang Fei;Zhu Rongxin(Early Intervention Unit,Affiliated Nanjing Brain Hospital,Nanjing Medical University,Nanjing 210029,China;School of Biomedical Engineering and Informatics,Nanjing Medical University,Nanjing 211166,China)
出处
《中华精神科杂志》
CAS
CSCD
北大核心
2022年第1期30-37,共8页
Chinese Journal of Psychiatry
基金
国家杰出青年科学基金(81725005)。
关键词
双相情感障碍
磁共振成像
诊断
计算机辅助
Bipolar disorder
Magnetic resonance imaging
Diagnosis,computer-assisted
