基于改进VGG–16和通道注意力机制的脑电图抑郁症辅助检测

Depression Auxiliary Detection by EEG Based on Modified VGG–16 and Channel Attention Mechanism

利用脑电图信号,结合深度学习方法进行抑郁症辅助诊断目前仍存在特征提取不足及模型诊断准确率不高的问题。为了提取更具抑郁症表征的特征,提高抑郁症辅助诊断的准确率,本文从特征提取和网络框架两个方面进行改进,提出一种结合改进VGG–16(visual geometry group–16)和基于压缩激励网络的通道注意力机制(modified VGG–16 network based on SE–NET,SEMod–VGG)的抑郁症辅助检测模型。首先,提取脑电图信号中α(Alpha)、θ(Theta)和β(Beta)频段的微分熵特征,与对应通道的功率谱密度特征相融合,构成一种同时具有时频属性和能量属性的4维融合特征;其次,针对该4维特征,改进现有的VGG–16模型,同时采用5×5和7×7两种不同尺度的卷积核,在提取脑电信号的时频信息和功率信息的同时,提高特征的泛化表征能力;再将基于压缩激励网络的通道注意力机制与改进的检测模型相结合,对电极通道的权重进行2次标定;最后采用10折交叉验证使得最小二乘支持向量机取得最佳检测准确率。对所提模型在准确率,召回率以及网络性能这3个方面进行实验评估,在MODMA数据集上的结果表明:当使用4维融合特征作为输入时,SEMod–VGG可达到最佳检测性能,其抑郁症检测准确率在3通道、16通道及128通道分别为92.21%、93.47%和95.76%;检测召回率在3通道、16通道以及128通道分别为91.57%、92.46%和96.80%。相较于现有的抑郁症辅助检测模型,本研究所提出的融合特征对抑郁症的表征性更强,且所提出的模型在检测准确率,召回率以及模型效率上均取得明显提升。

At present, using electroencephalogram signals combined with deep learning methods for depression auxiliary detection still suffers from some problems such as insufficient feature extraction and low accuracy of detection. In order to extract more representative features and improve the accuracy of depression auxiliary detection, this study modifies the model from two aspects: feature extraction and network framework.A depression auxiliary detection model was proposed by combining the modified VGG–16 network and the channel attention mechanism based on the Squeeze-and-excitation Networks(SEMod–VGG). Firstly, the differential entropy of frequency bands named α(Alpha), θ(Theta) and β(Beta) in electroencephalogram were fused with the power spectral density for the corresponding channels to form a four-dimensional fusion feature which takes on time-frequency and energy attributes;Then, for the four-dimensional feature, the existing VGG–16 model was modified and two different scales convolution kernels, 5×5 and 7×7 were simultaneously used to improve the generalized characterization of the feature while extracting the time-frequency and power information of the electroencephalogram signals. Subsequently, the channel attention mechanism based on Squeeze-and-excitation Networks was combined with the modified detection model to calibrate the weights of the electrode channels. Finally,the ten-fold cross validation was utilized to optimize the least squares support vector machine, which improves the limitations of traditional support vector machine. Experimental evaluation in three aspects, namely, auxiliary detection accuracy, recall and network performance were conducted in the paper. The results on the MODMA dataset showed that SEMod–VGG yields the best accuracy for detection when four-dimensional fusion features are used as input: 92.21% for 3 channels, 93.47% for 16 channels and 95.76% for 128 channels. Besides, the detection recall of the model also reaches a higher level: 91.57% for 3 channels, 92.46% for 16 channels and 96.80% for 128 channels. Compared with the existing models of depression auxiliary detection, the proposed fusion features are more characterizing for depression, and the model achieves significant improvements in detection accuracy, recall and model efficiency.