事前学習つき深層ニューラルネットワークを用いた肺聴診音識別

―CNN, LSTM, 畳み込みLSTMの性能比較―

抄録

Physicians in the medical field have carried heavy burdens of diagnosis because they need to find various diseases of many patients on the basis of various examinations. Recently, to reduce their burdens, deep learning is enthusiastically applied to medical fields. For example, there have been many applications of deep learning to chest CT and X-ray images. However, there are few studies on deep learning for auscultation. Therefore, we aim to build a lung sound classification system using deep learning. Although a large number of data with annotation are generally required for deep learning, it is difficult to collect a sufficient number of lung sounds data. Therefore, we propose some lung sound classification systems with deep learning for efficiently training neural networks with a small number of data. In detail, 1) Mel-Frequency Cepstral Coefficients are used for feature extraction and 2) some pre-training techniques with three types of neural networks such as a convolutional neural network (CNN), long short term memory (LSTM), and convolutional long short term memory (C-LSTM) are designed to realize efficient learning for a small number of lung sounds data. From the experimental results, it is clarified that the proposed pre-training techniques show effective classification performance, and especially, C-LSTM with pre-training achieves higher accuracy than conventional CNN and LSTM.