摘要
针对传统肺音识别方法,存在准确率低且难以适应实际需求等问题,提出了一种基于深度学习的肺音辅助诊断方法.将肺音信号转换成具有时频特性的梅尔倒谱图,设计深度学习模型GoogLeNet-LSTM深入分析梅尔倒谱图的空间特征和时间序列特征,对病毒性肺炎患者的病情进行诊断.与VGG16、ResNet等网络模型进行比较,提出的模型性能更优,诊断准确率达到92.06%,相比单模型的准确率提高了至少4%.实验结果表明,该网络模型能有效支撑病毒性肺炎患者病情评估工作,起到辅助诊断作用.
Abstract
Aiming at the problems that the traditional lung sound recognition method has low accuracy and difficulty in adapt-ing to practical needs,a lung sound auxiliary diagnosis method based on deep learning is proposed.The deep learning model GoogLeNet-LSTM is designed to deeply analyze the spatial and time series features of the Mel feature maps to diagnose the condi-tion of patients with viral pneumonia.Compared with network models such as VGG16 and ResNet,the proposed model has better performance,with a diagnosis accuracy rate of 92.06%.The accuracy of the single model is improved by at least 4%.The experimen-tal results show that the network model can effectively support the evaluation of patients with viral pneumonia and play an auxiliary role in diagnosis.
基金项目
武汉市科技局应用基础前沿专项(2020020601012320)
维普
万方数据