基于双重注意力机制增强的复合型眼震分类框架  

Composite nystagmus classification framework enhanced by dual attention mechanism

作　　者：王卓然 方志军[1] 王海玲 高永彬 李玉霞 WANG Zhuoran;FANG Zhijun;WANG Hailing;GAO Yongbin;LI Yuxia(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201600,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201600

出　　处：《中国医学物理学杂志》2024年第9期1093-1103,共11页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(62001284);上海市科委“科技创新行动计划”社会发展科技攻关项目(21DZ1204900)。

摘　　要：针对现有研究仅能识别水平、垂直或轴向上某一方向是否有眼震发生,且未能考虑临床上具有强度变化、由多方向组成的复合型眼震的问题,提出一种基于双重注意力机制增强的复合型眼震分类框架。首先,提出一种眼震视频时空浓缩算法,结合卷积神经网络与霍夫变换,去除无效帧和无效区域的干扰,提高眼震视频质量。然后,采用密集光流算法提取眼球运动光流场。最后,构建一种基于双重注意力机制增强的复合型眼震分类网络,提出一种改进高效通道注意力模块,有效获取光流图不同通道中眼球震颤的方向、强度信息;在Bi-LSTM网络末端添加时间注意力模块,实现不同时序特征对分类结果的显著性表达。在合作医院提供的眼震数据集上,本文方法对复合型眼震分类准确率达到83.17%,在单独的水平、垂直、轴向上眼震分类准确率达到91.03%、89.74%、86.05%。本文方法实现复合型眼震的智能分类,具有一定的临床应用价值。A composite nystagmus classification framework enhanced by dual attention mechanism is proposed to address the problem that the existing researches only identify whether nystagmus occurs in a horizontal,vertical,or axial direction,but fail to consider the issue of composite nystagmus composed of multiple directions with various intensities in clinical practice.A spatiotemporal concentration algorithm for nystagmus videos is presented,and it combines convolutional neural networks and Hough transform to remove interference from invalid frames and regions and to improve the quality of nystagmus videos.Then,a dense optical flow algorithm is used to extract the optical flow field of eye movement.Finally,a composite nystagmus classification network based on dual attention mechanism enhancement is constructed.An improved efficient channel attention module is used to effectively obtain the direction and intensity of nystagmus in different channels of the optical flow map;and a temporal attention module is added at the end of the bidirectional long short-term memory network to achieve significant expression of classification results based on different temporal features.On the nystagmus dataset provided by the cooperating hospital,the proposed method has an accuracy rate of 83.17%for composite nystagmus classification,and achieved accuracy rates of 91.03%,89.74%,and 86.05%for individual horizontal,vertical,and axial nystagmus classifications.The proposed method realizes the intelligent classification of composite nystagmus and is valuable in clinic.

关 键 词：医学图像处理 视频眼震电图 良性阵发性位置性眩晕 深度学习 注意力机制 

分 类 号：R318[医药卫生—生物医学工程] TP391[医药卫生—基础医学]