多输入神经网络的肺气肿识别  

Emphysema Recognition by Multi-input Neural Network

作　　者：郭涛[1] 古依聪 刘启明 李成 石帅 GUO Tao;GU Yicong;LIU Qiming;LI Cheng;SHI Shuai(State Key Laboratory of Electronic Measurement Technology,North University of China,Taiyuan 030051;China Academy of Launch Vehicle Technology,China Aerospace Science and Technology Corporation,Beijing 100000)

机构地区：[1]中北大学电子测试技术国家重点实验室,太原030051 [2]中国航天科技集团有限公司中国运载火箭技术研究院,北京100000

出　　处：《计算机与数字工程》2025年第3期678-683,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:51975541)资助。

摘　　要：针对肺气肿的两种典型特征(喘鸣音和水泡音)采用多输入神经网络对其进行分类,达到判断是否患有高原肺气肿的目的。针对肺音数据,对其进行滤波降噪后采用梅尔谱图(Mel)、恒Q变换(CQT)、小波变换(WT)和短时傅里叶变换(STFT)四种谱图特征提取方式。并采用LBP和Mixup进行数据增强,输入多输入卷积神经网络(Mul-CNN)中进行肺音分类。得到肺音在使用WT和Mel作为输入时准确率为93.6%,特异率为92.3%,灵敏度为94%,ICBHI得分为93.1%。Multiple input neural network is used to classify the two typical characteristics of emphysema(stridor and blister),so as to determine whether high altitude emphysema is present or not.For lung sound data,four spectral feature extraction methods including Mel spectrum(Mel),constant Q transform(CQT),Wavelet transform(WT)and short-time Fourier transform(STFT)are used after filtering and denoising.LBP and Mixup are used for data enhancement,and lung sound classification is performed in Mul-CNN.The accuracy,specificity,sensitivity and ICBHI scores of lung sounds are 93.6%,92.3%,94%and 93.1%when WT and Mel are used as inputs.

关 键 词：肺气肿 Mul-CNN LBP Mixup 

分 类 号：TN912.3[电子电信—通信与信息系统] R563.1[电子电信—信息与通信工程]