２つの解像度スペクトログラムとTF-CRNN による呼吸音の分類

抄録

Respiratory disease is a serious illness that accounts for four of the world's top 10 causes of death in a year and accounts for more than 8 million deaths worldwide. Currently, the diagnosis of respiratory diseases is made by auscultation, however the diagnosis result depends on the proficiency of the doctor. Therefore, a computer-aided diagnosis system that quantitatively classifies breath sounds and outputs the results as a "second opinion" is required. In this paper, we describe the development of an algorithm for automatically classifying large-scale respiratory sound data sets used in the ICBHI 2017 Challenge. The proposed method consists of two stages. First, by preparing two types of window widths for short-time Fourier transform, two types of images, a spectrogram with high time resolution and a spectrogram with high frequency resolution, are generated. Second, T-CRNN that learns the series features in the time direction, F-CRNN that learns the series features in the frequency direction, and respiratory sounds were classified by inputting them into TF-CRNN, which is a composite model of them. In this paper, we apply our proposed method to 920 respiratory sound data, and average score of 64%, harmonic score of 62%, sensitivity of 54% and specificity of 73% are obtained.