Research On Rales Detection Method Based On Deep Learning

The lives of more than 1 billion people worldwide are affected by respiratory diseases.Due to limited medical resources,the treatment rate of respiratory diseases is low.Auscultation is a commonly used screening method for respiratory diseases,which has the advantages of simplicity,non-invasiveness and low cost.Rales are a common pathological signals during auscultation,and its appearance is usually an early sign of respiratory diseases such as asthma and pneumonia.Traditional analog stethoscopes require professional judgments,subjective limitations,and low efficiency;digital stethoscopes have the functions of high fidelity,storage and remote transmission of auscultation signals,and are receiving more and more attention.This article focuses on the research on the rales detection algorithm in the digital stethoscope.The main contents are as follows:(1)The rales signal is affected by many hidden variables,including individual differences,noise,etc.,showing strong non-stationarity.This thesis uses Residual Networks(Res Net)to learn potentially discriminative rale detection features.Res Net has a stacked residual structure,which can adaptively infer hidden variables and avoid network degradation problems.In view of the short duration of rales,the inconsistent degree of interference and the easy concealment by noise,the attention mechanism helps to capture key time-frequency characteristics.Therefore,this paper proposes a rale detection method based on residual network and attention mechanism.A total of325 subjects’ clinical breath sounds were collected by doctors using digital stethoscopes.The data obtained was cleaned and a breath sound database containing 2620 valid samples was established.On this basis,a clinical breath sound detection experiment was carried out to verify the effectiveness of the model.(2)In the existing research on respiratory sounds,extra memory space is needed when data preprocessing transforms features,and fixed transform parameters are not necessarily the optimal problem.The one-dimensional convolution kernel is used to realize the Fourier transform instead of the traditional feature preprocessing and embed it into the model for joint optimization.The introduction of the lightweight and efficient network Efficient Net is more accurate than Res Net.Efficientnet uses a residual link similar to Res Net,and also uses an attention mechanism in its composition.At the same time,in order to capture more rale timing information,this paper integrates the Long Short Term Memory(LSTM)network into the model,and obtains an end-to-end rale detection algorithm based on Efficientnet+LSTM.By adding the same attention mechanism in Res Net and the experimental comparison of LSTM on Efficient Net,it is proved that the introduction of LSTM in Efficientnet has a better effect on the rales detection of the boosted model.Both of the proposed deep-learning-based rhetoric detection algorithms achieve the goal of automated rales detection.The Efficient Net-LSTM-based end-to-end rales detection algorithm is more effective in identifying normal breath sounds and rales,and solves the problems of fixed parameters and extra memory space required for traditional feature transformations,making the network more practical and suitable for migrating the algorithm model to digital stethoscope systems.