基于三维卷积神经网络的多频段癫痫发作类型分类

Classification of Multi-subband Seizure Based on Three-dimensional Convolutional Neural Network

脑电信号包含丰富的大脑活动信息,是诊断癫痫最主要的手段。在现阶段癫痫脑电信号分类研究中,单一域无法全面刻画信号数据,空间信息未被有效利用。并且,随着连续卷积层数增加以及多域特征融合,判别性重要特征遗失。提出一种端到端的三维卷积多频段癫痫发作类型分类模型,运用时-频-空域三维数据构建方法,将时频域联合分布信息根据通道空间位置重构为时-频-空域三维表达。用一系列并行的三维卷积子网络直接提取各频段的时间、频率、空间判别性联合活动特征,通过多层感知机整合输出分类结果。基于世界最大的公开癫痫发作语料库TUSZ数据集进行癫痫发作类型7分类实验,得到该模型分类F1-score为0.9274,准确率92.69%,与其他方法比较有明显的提升,验证了该模型在癫痫发作类型分类任务中的有效性。

In this paper,an end-to-end 3D convolutional multi-subband seizure type classification model is proposed,which reconstructing joint time-frequency domain distribution information into a time-frequency-space domain 3D representation according to the location of the channel.The discriminative joint activity features of each band are extracted using a series of parallel 3D convolutional sub-networks,and the classification results are integrated by a multi-layer perceptron.Based on the TUSZ dataset,the world’s largest publicly seizure corpus,for 7 classification experiments,and the F1-score of this paper’s model classification is obtained as 0.9274 with an accuracy of 92.69%,which is a significant improvement compared with other methods.