Artificial Intelligence Assisted Diagnosis Of Viral Pneumonia Based On Lung Sounds

Lung sounds contain lots of information,which can reflect abnormal conditions of the human body.Auscultation of lung sounds is an effective technique for diagnosing viral pneumonia.Traditional auscultation methods have the advantages of being non-invasive and easy to implement.It mainly relies on doctors’ years of clinical experience,which may have subjective errors.It is necessary to use computer-aided diagnosis technology to improve the diagnosis efficiency.However,the current auxiliary diagnosis method based on lung sounds still has problems such as insufficient accuracy.Therefore,more optimized algorithms are needed to complete the clinical diagnosis and assessment of viral pneumonia.To this end,an intelligent assisted diagnosis method based on lung sounds was proposed in this thesis.The method first analyzed the characteristics of the patient’s lung sound data,and then used a neural network to build a classification model to evaluate the condition of patients,including moderate,severe and critical.The main work of this paper are as follows:Firstly,according to the doctor’s annotation,a dataset of lung sounds was established.The fifth-order Butterworth bandpass filter was used to filter out the noise.And the lung sound signals were re-segmented in units of 10 s.The characteristics of lung sound signals have been studied from multiple perspectives in the time and frequency domains.Secondly,three feature extraction methods were considered separately,including the Short Time Fourier Transform(STFT),Mel-Frequency Cepstral Coefficients(MFCCs)extraction method and Wavelet Transform.Through experimental demonstration,MFCCs extraction were determined as the feature extraction method in this work.Finally,the GoogLeNet network model was used for recognition and classification,and it was improved.The GoogLeNet-LSTM network model was designed to deeply analyze the spatial and time series characteristics of MFCCs.It improved the generalization ability of the network model and the accuracy of disease assessment.In this paper,the GoogLeNet-LSTM network model was used to diagnose and evaluate the condition of patients with viral pneumonia.It reached the best classification accuracy of92.06%,which was 4.45% higher than that of the single network model.The results of related experiments in this paper showed that the GoogLeNet-LSTM network model can effectively support the evaluation of viral pneumonia and play an auxiliary diagnosis role.