深層ニューラルネットワークを用いた肺聴診音の異常検知―DAGMM, Efficient GANの性能比較と改良―

カテゴリ: 論文誌(論文単位)

グループ名: 【C】電子・情報・システム部門

発行日: 2022/12/01

タイトル(英語): Anomaly Detection of Lung Sounds Using Deep Neural Networks―Comparison and Improvement of the Performance of DAGMM and Efficient GAN―

著者名: 若本　亮佑（山口大学大学院創成科学研究科／（株）電通国際情報サービス），間普　真吾（山口大学大学院創成科学研究科），木戸　尚治（大阪大学大学院医学系研究科），呉本　尭（日本工業大学先進工学部情報メディア工学科）

著者名(英語): Ryosuke Wakamoto (Graduate School of Sciences and Technology for Innovation, Yamaguchi University/Information Services International-Dentsu, Ltd.), Shingo Mabu (Graduate School of Sciences and Technology for Innovation, Yamaguchi University), Shoji Kido (Graduate School of Medicine, Osaka University), Takashi Kuremoto (Department Information Technology & Media Design, Nippon Institute of Technology)

キーワード: 聴診音，深層学習，異常検知，DAGMM，Efficient GAN_x000D_　　lung sounds，deep learning，anomaly detection，DAGMM，Efficient GAN

要約(英語): To find lung diseases, physicians need to conduct various examinations. Recently, to reduce their burdens, many applications of deep learning have been proposed to diagnose chest X-ray images. However, there are few studies using deep learning for auscultation, and also, there are only a few small-scale benchmark datasets of lung sounds that are annotated for machine learning. Therefore, we aim to build an anomaly detection system that only uses normal data for the training. When building anomaly detection systems, it is important to capture generalized features based only on the normal data. To solve this problem, first, we propose some algorithms that improve the Deep Autoencoding Gaussian Mixture Model (DAGMM). Second, we propose some algorithms that improves Efficient GAN. Various types of neural networks such as CNN, LSTM, and convolutional LSTM (C-LSTM) are applied to DAGMM, and GMM and C-LSTM are applied to Efficient GAN for effective feature extraction. The experimental results show that each of the proposed methods has effective classification performance for lung sounds, and especially, the combination of convolution and LSTM, and the combination of feature extraction and GMM are effective for any of the models.

本誌: 電気学会論文誌C（電子・情報・システム部門誌） Vol.142 No.12 （2022） 特集：電気・電子・情報関係学会東海支部連合大会

本誌掲載ページ: 1328-1335 p

原稿種別: 論文／日本語

電子版へのリンク: https://www.jstage.jst.go.jp/article/ieejeiss/142/12/142_1328/_article/-char/ja/