Classification of Abnormal Lung Sounds Using Deep Learning

Lung sound is an important reference factor in determining respiratory diseases. In particular, automatic lung sound classification systems could be of great help in situations where medical professionals are unavailable. In this work, we preprocess the original lung sound signal to remove noise interference from the signal. The processed sound signal is generated as a spectrogram by a short-time Fourier transform. The spectrogram is classified by a deep learning network based on ResNet, thus identifying the respiratory cycle into four types: normal, crackle, wheeze, and both. To address the issue of varying time scales in spectrograms, we extend the respiratory cycles to a uniform fixed time. The official benchmark standards of the ICBHI 2017 challenge and the dataset partitioning scheme have been used to validate the proposed method. Experiments and comparisons show that the proposed method achieves promising results in the classification of the respiratory cycle.