基于原始肺音的端到端深度学习肺炎辅助诊断  

End-to-end deep learning for auxiliary diagnosis of pneumonia using original lung sounds

作　　者：肖欣[1] 高玉清 张建敏 XIAO Xin;GAO Yuqing;ZHANG Jianmin(School of Artificial Intelligence,Jianghan University,Wuhan 430056,China)

机构地区：[1]江汉大学人工智能学院,湖北武汉430056

出　　处：《中国医学物理学杂志》2025年第2期274-280,共7页Chinese Journal of Medical Physics

基　　金：武汉市科技计划项目(2020020601012320);老年医学与现代康养学科(群)项目([鄂教研函[2021]5号])。

摘　　要：提出一种基于双路径不同卷积核的DCL-Net的端到端肺炎辅助诊断方法。该方法无需进行特征工程,将原始肺音信号直接输入模型,利用卷积核分别为1∗3和1∗5的双路径卷积网络,每个路径包含3个残差块,以便模型自动学习肺音信号不同尺度的特征,同时避免模型退化问题。为验证端到端方法的性能,将其与信号分析领域常用的梅尔倒谱图、短时傅里叶变换和小波变换这3种特征提取方法进行比较。结果显示,四分类任务(正常、普通、病重、病危)诊断准确率为61.4%,相比3种特征工程方法分别提高1.6%、5.0%和3.7%;二分类任务(正常、异常)诊断准确率为89.7%,相比3种特征工程方法分别提高11.0%、5.1%和11.2%。实验结果表明该方法可为肺炎病情评估提供更有效的诊断工具。An end-to-end auxiliary diagnosis method for pneumonia based on DCL-Net with dual-path of different convolutional kernels is proposed,in which no feature engineering is required,and the original lung sound signal is directly input into the model.The dual-path convolutional network with kernel sizes of 1*3 and 1*5,with each path containing 3 residual blocks,allows the model to automatically learn features of lung sounds at different scales while avoiding model degradation.The performance of the end-to-end method is validated through the comparisons with 3 commonly used feature extraction methods in signal analysis,namely Mel-spectrogram,short-time Fourier transform,and wavelet transform.The results show that the proposed method has a diagnostic accuracy of 61.4%for the 4-class classification task(normal,moderate,severe,critical),which is 1.6%,5.0%,and 3.7%higher than the other 3 feature extraction methods,and the diagnostic accuracy is 89.7%for the binary classification task(normal or abnormal),which is 11.0%,5.1%,and 11.2%higher than the other 3 feature engineering methods,demonstrating that it can serve as an effective diagnostic tool for pneumonia.

关 键 词：原始肺音 肺炎 智能辅助诊断 端到端学习 DCL-Net 

分 类 号：R318[医药卫生—生物医学工程] R563.1[医药卫生—基础医学]