Lung Sound Recognition Based on Pre-Trained Convolutional Neural Network

Abstract

It has become one of the focuses of respiratory medicine to recognize lung sounds by machine learning methods, and then assist doctors to diagnose patients' pulmonary diseases. Aiming at the problems of model overfitting and low classification accuracy caused by the small size of lung sound dataset in current lung sound recognition, a lung sound recognition method based on the combination of pre-trained convolutional neural network and CatBoost algorithm was proposed. The pre-trained convolutional neural network on the image dataset ImageNet is transferred to lung sound recognition. The channel attention mechanism CBAM is fused to enhance the recognition performance of the network, and the lung sound waveform data is converted into logarithmic MEL frequency spectrum for input and training. Finally, the trained model is used as a feature extractor, and the feature vectors with high-level semantics are input into the ensemble learning algorithm CatBoost to achieve the final classification. After experiment, the result shows that the specificity, sensitivity and ICHBI-score of the proposed method for lung sound recognition in ICBHI-2017 lung sound dataset reach 88.34%, 63.13% and 75.73%, respectively, which is superior to the previous methods. The display has a good application prospect in lung sound recognition.