부산대학교 도서관

We propose LungNet; a deep learning architecture for the diagnosis of respiratory pathologies using multi-dimensional representations of lung sounds. The model is comprised of two independent branches, denoted CNN and LSTM. We have worked on three independent databases and conducted two experiments to evaluate the performance of our model. In the first experiment, the databases were merged into a single dataset. In the second experiment, we studied the distribution shift between the databases by training the model on one database and testing it on the other two. Our model outperformed the two baselines for both experiments. Our results suggest that LungNet has the potential to be a valuable tool for diagnosing respiratory pathologies, and could be integrated into clinical practice.