A Deep Neural Network with Triplet Loss for Detecting Anomaly of Respiratory Sounds
* Presenting author
Abstract:
In this paper, we propose a deep learning based system for detecting anomaly of respiratory sounds. The system is separated into two main steps: front-end feature extraction and back-end classification, respectively. In the first step, audio recordings of respiratory cycles collected from patients are transformed into Gammatone-based spectrograms where both temporal and frequency features of respiratory sounds are presented. In the second step, a convolutional neural network (CNN) based architecture, referred to as the baseline, is proposed to detect whether the spectrogram input contains abnormal features or not. To further improve the baseline performance, we then propose a triplet-based network architecture in which parallel CNN baseline networks and triplet loss function are made use. To make our work comparable, we evaluate our systems on the 2017 Internal Conference on Biomedical Health Informatics (ICBHI), which is one of the largest public benchmark respiratory sound datasets. The proposed CNN baseline and triplet-network architecture outperform the ICBHI baseline, improving by 5% and 7%, respectively.