Research On Diagnosis Of Lung Diseases Via Transfer Learning Based Children’s Breath Sound Recognition

A healthy respiratory system can maintain the metabolism required for life,but compared with adults,children’s lung health is more vulnerable to serious impacts due to their immature immune system and lack of self-prevention means and self-protection awareness.With the continuous development of signal processing technology and artificial intelligence technology,an objective and accurate computer-aided breath sound recognition model has received widespread attention since the effective recognition and analysis of children’s respiratory sounds are crucial for predicting and diagnosing early lung diseases in children.However,the scarcity of medical data limits the development of these data-driven methods toward reliability and high recognition accuracy.The thesis proposes a transfer learning-based children’s respiratory sound recognition method to obtain a superior performance recognition model based on small datasets.The main work is as follows:1.Cooperate with hospitals to collect children’s respiratory sound data,which constitute experimental datasets for subsequent analysis and classification;use band-pass filtering,wavelet thresholding,empirical mode decomposition three denoising methods to filter out low-frequency high-frequency noise and heart sound components in respiratory sounds.The experiment shows that this method can effectively extract pure respiratory sounds.2.Aiming at the problem that the small sample size of children’s respiratory sounds causes difficulty in training and difficulty in adapting to high-standard clinical needs,a dual-model fusion classification model based on deep learning and transfer learning is proposed.Taking three feature maps of respiratory sounds as input,when constructing the network,use Image Net dataset pre-trained Mobile Net V2 model and fine-tune the model;since Res Net has residual connection property,it is suitable as a feature extractor After extracting features,use random forest method for classification.The two models are fused by soft voting method,and finally get three categories of respiratory sound recognition results.The accuracy rate of respiratory sound classification by the model is97.96%,precision rate is 97.83%,recall rate is 97.89%,specificity is 98.89%,F1 Score is 0.98.3.The thesis develops a respiratory sound recognition classification system based on Python language.It realizes the development of PC-side signal classification software based on transfer learning.Users can intuitively see respiratory sound waveforms,features and analysis results on the system.