基于注意力机制的CNN和GRU烟雾病检测方法研究  

Research on Detection of Moyamoya Disease by CNN and GRU Based on Attention Mechanism

作　　者：胡涛 苏佳斌[2] 杨恒[2] 蒋卓韵 余锦华[1] 顾宇翔[2] Hu Tao;Su Jiabin;Yang Heng;Jiang Zhuoyun;Yu Jinhua;Gu Yuxiang(不详;Department of Electronic Engineering,Fudan University,Shanghai 200433,China)

机构地区：[1]复旦大学电子工程系,上海200433 [2]复旦大学附属华山医院,上海200040

出　　处：《航天医学与医学工程》2021年第1期68-74,共7页Space Medicine & Medical Engineering

基　　金：上海市科技创新行动计划新技术领域项目(18511102800)。

摘　　要：目的为提高数字减影血管造影(DSA)进行烟雾病诊断的精准率,建立时序分类网络的烟雾病检测模型。方法首先对DSA图像进行预处理,将预处理后的图像采用卷积神经网络(CNN)来提取特征;然后采用结构简单、具有时间记忆功能的门控循环单元(GRU)融合DSA的时序信息,获取更全面的烟雾病特征信息;最后,再结合注意力机制能获取图像重点关注区域中烟雾病的细节信息,而抑制其他无用信息。结果将该方法与3D-CNN、CNN+LSTM以及CNN+LSTM+Attention进行对比,得到的检测准确率和均方误差分别为98.57%和1.43%。结论本方法在烟雾病检测上具有优势,是一种稳定可靠的检测方法。Objective To improve the accuracy of diagnosis of moyamoya disease on digital subtraction angiography(DSA),a moyamoya disease detection model based on time-series classification network was established.Methods Firstly,the DSA image was preprocessed,and the spatial features were extracted by convolutional neural network(CNN).Then a simple and time memory gated recurrent unit(GRU) was used to fuse the time sequence information of DSA to obtain more comprehensive characteristic information of moyamoya disease.Finally,combining the attention mechanism,the details of moyamoya disease in the important area of the image were obtained and other useless information in the image was suppressed.Results Compared with 3D-CNN,CNN+LSTM and CNN+LSTM+Attention,the detection accuracy and mean square error were 98.57% and 1.43% respectively.Conclusion This method has advantages in the detection of moyamoya disease and it is a stable and reliable detection method.

关 键 词：烟雾病检测 数字减影血管造影 卷积神经网络 门控循环单元 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]