Multilevel Feature Fusion-based Convolutional Neural Network for Anomaly Classification of Respiratory Sound

Respiratory sound classification has been extensively studied to early detect heart and lung diseases using digital stethoscope. Previous works mainly leveraged convolutional neural network (CNN)-based deep learning methods by transforming sound signals into spectrogram images. The fine-tuned model based on pretrained ResNet (RespireNet) has been proposed and achieved high performance. However, ResNet has a very large network size, which is not suitable for mobile applications. In this paper, we proposed a lightweight deep learning approach based on a CNN architecture by incorporating self-attention and multilevel feature extraction. With considerably fewer number of parameters, we achieved a competitive performance compared to the state-of-the-art model, RespireNet, on the ICBHI dataset for normal-abnormal respiratory sound classification. The smaller model size can give an advantage in the development of internet of things (IoT) applications for automatic respiratory sound classification.