Document Type : Full Research Paper

Authors

• davoud saadati ¹
• Sattar Mirzakuchaki ²

¹ M.Sc, Student, Department of Electrical Engineering, Iran University of Science and Technology

² Professor/Department of Electrical Engineering, Iran University of Science and Technology,

Abstract

Analysis and examination of sound of organs can be utilized in order to diagnose various diseases and abnormal conditions. Diagnostic methods based on audio signal processing are non-invasive and inexpensive and can be especially useful in under-developed countries, where inadequate medical specialists and equipment has led to high fatality rates. Development of accessible methods based on machine learning can aid with early diagnosis. we used a convolutional network to attain the advantages of transfer learning. In previous studies, models have been proposed that feed spectrograms with frequency characteristics as inputs to the convolutional network. In this article, we propose a model which additionally employs a recurrent representation (Recurrence plot) that reflects the temporal characteristics of the sound. The audio data sequence is investigated by adding the temporal attention mechanism and the bi-directional recurrent gates for weighting data according to its informational value. Data used in this article is from the ICBHI lung sound database. The presented model was able to classify lung sounds into three categories: healthy, chronic obstructive pulmonary disease (COPD), and other diseases with an accuracy of 97%, which shows the superiority of the proposed method compared to results obtained from previous methods on the same database.

Keywords

• Diagnosis
• Lung disease
• Sound processing
• Recurrence Plot
• Temporal attention
• Recurrent network
• Transfer learning
• machine learning

Main Subjects

• Biomedical Signal Processing / Medical Signal Processing / Biosignal Processing