Classification of Respiratory Sounds by scSE-CRNN from Triple Types of Respiratory Sound Images

Bibliographic Information

Other Title
  • scSE-CRNNと3種類の呼吸音変換画像による呼吸音の分類

Search this article


<p>Due to the respiratory diseases such as chronic obstructive pulmonary disease and lower respiratory tract infections nearly 8 million people were died worldwide each year. Reducing the number of deaths from respiratory diseases is a challenge to be solved worldwide. Early detection is the most efficient way to reduce the number of deaths in respiratory illness. As a result, the spread of infection can be suppressed, and the therapeutic effect can be enhanced. Currently, auscultation is performed as a promising method for early detection of respiratory diseases. Auscultation can estimate respiratory diseases by distinguishing abnormal sounds contained in respiratory sounds. However, medical staff need to be trained to perform auscultation with high accuracy. Also, the diagnostic results depend on each staff subjectively, which can lead to inconsistent results. Therefore, in some environments, a shortage of specialized health care workers can lead to the spread of respiratory illness. To solve this problem, an application that analyzes respiratory sounds and outputs diagnostic results is needed. In this paper, we use a newly proposed deep learning model to automatically classify the respiratory sound data from the ICBHI 2017 Challenge Dataset. Short-Time Fourier Transform, Constant-Q Transform, and Continuous Wavelet Transform are applied to the respiratory sound data to convert it into the time-frequency region. Then, the obtained three types of breath sound images are input to CRNN (Convolutional Recurrent Neural Network) having scSE (Spatial and Channel Squeeze & Excitation) Block. The accuracy is improved by weighting the features of each image. As a result, AUC (Area Under the Curve): (Normal:0.87, Crackle:0.88, Wheeze:0.92, Both:0.89), Sensitivity: 0.67, Specificity: 0.82, Average Score: 0.75, Harmonic Score: 0.74, Accuracy: 0.75 were obtained.</p>


Related Projects

See more

Details 詳細情報について

Report a problem

Back to top