Classification and Recognition of Lung Sounds Based on Improved Bi-ResNet Model

Abstract

Lung sound classification is an important diagnostic task in the medical field. By analyzing respiratory sounds, doctors can help diagnose various respiratory system diseases. Chronic respiratory diseases worldwide are usually associated with abnormal lung sounds, which are clinically related to conditions such as bronchitis or chronic obstructive pulmonary disease. In recent years, the outbreak of COVID-19 has once again sparked research into lung sound classification. However, due to the environmental noise and heart sounds mixed in abnormal lung sounds, further improvements are still needed for accurate classification. In this paper, an improved Bi-ResNet network structure model is proposed to enhance the accuracy of lung sound classification and fully utilize feature extraction information. The model still processes the extracted lung sound features in parallel, but by introducing skip connections and increasing the use of direct connections, it allows information to be directly transmitted and fully integrates original and processed features within the network. This improved structure enables the model to learn features from the data at a deeper level, enhancing the expressiveness of the features. Additionally, the improved Bi-ResNet model combines convolutional neural networks (CNN) and residual networks (ResNet), and uses two types of features, the lung sound short-time Fourier transform (STFT) and wavelet transform (Wavelet), for model training and analysis. This comprehensive approach captures lung sound data information more comprehensively, differentiating between different types of lung sounds and providing better diagnostic assistance to doctors, thereby promoting early diagnosis and treatment of respiratory system diseases. Through experiments, the proposed model achieved a classification accuracy of 77.81% on the Int. Conf. on Biomedical Health Informatics (ICBHI) 2017 dataset, representing a 25.02% improvement over the Bi-ResNet model, with an F1 score of 71.05%.

Keywords

• Lung sound classification
• Bi-ResNet model
• deep learning
• Fourier transform
• wavelet transform