Research On Pneumonia Recognition Algorithm Based On Deep Learning

Pneumonia is the disease with the highest mortality rate among infectious diseases in China,it needs to be focused on prevention and control.Medical diagnostic tools for pneumonia often require specialized equipment,such as blood routine examination,chest X-ray,sputum culture,etc.The non-contact pneumonia identification method based on cough sound is realized by distinguishing the difference between normal cough sound and pneumonia cough sound,which has the advantages of low cost,and easy to operate and so on.The characteristics of non-contact diagnosis can block the transmission route of the diagnosis process of infectious pneumonia.At the same time to avoid blindly seek medical treatment,and causing infections in the hospital.As a clinical aid to reduces the doctor’s workload and provide health screening.It provides health screening function as a clinical auxiliary means to solve the problem of lack of medical resources in remote and poor areas.Pneumonia identification studies have previously focused on finding feature combination combining methods,and output by machine learning models.The disadvantage is that complex feature engineering and low recognition effect,or studies around a particular type of pneumonia,there is a lack of universal identification of pneumonia.Therefore,it is important to propose a recognition algorithm that can accurately identify pneumonia.This paper adopts the standardized cough sound acquisition process and processing steps,and studies the pneumonia recognition algorithm based on deep learning.The main research contents are as follows:(1)We propose the research of pneumonia recognition algorithm based on genetic algorithm and recurrent neural network.According to the temporal characteristics of cough sounds,the various time-frequency domain features of cough sounds were extracted.Network model with memory function are used as recognition models for comparative analysis and research,such as recurrent neural networks,long and short term memory networks,gated recurrent units,etc.Genetic algorithm is introduced into the original best model to automatically update the network model and parameters.The accuracy of the best model Max-GA obtained by genetic algorithm update was 80.8%,an improvement of 2.59 percentage points.It is shown that the approach of genetic algorithm and recurrent neural network can identify pneumonia very well.(2)We propose a study of pneumonia recognition based on feature cascade and attention mechanism.We process cough sounds using wavelet packet decomposition and generating spectrograms,respectively.Thus,multi-level and multi-modal cough sound features are extracted.On the improved Le Net-5 network model,we use the feature cascade method to fuse multimodal features and output the identification results.The purpose is to enhance the ability to express the cough sound and avoid the problem of missing information due to single feature.Based on the original model,we propose to use an attention mechanism to differentiate the fused features.In this way,the overall attention of the model is focused on the salient features and suppress the invalid features.The salient features facilitate the identification of the results.The experimental results show that the method is 7.67 percentage points higher than the original model and achieves an accuracy of 84.19%,indicating that the method can improve the recognition ability of the network model for pneumonia.