Detection of abnormalities in imaged lung sounds based on deep learning
Authors: Shogo Matsumoto, Naoya Wakabayashi, Hiromitsu Shimakawa, Humiko Harada
Abstract: Despite the increase in respiratory diseases, the number of respiratory specialists is decreasing. The shortage of respiratory specialists has made the COVID-19 pandemic more serious. The pandemic has revealed the difficulty of controlling transmission, diagnosing, monitoring disease status, and responding to symptoms of infectious respiratory diseases. The global outbreak of the new virus infections has reminded us of the fragility of the conventional healthcare system.The most effective examination in the examination of respiratory disease is auscultation. However, features of abnormal sounds the disease brings are too obscure for doctors who are not specialists in respiratory to distinguish abnormal sounds from normal ones. Furthermore, due to aging, we would suffer from difficulty in hearing high-pitched sounds, which obliges even specialists often make mistakes in diagnosis. Diagnosis by auscultation depends on subjective judgment and the skill of the specialist. Today, when specialists are in short supply, information technology is expected to support even non-specialists to be able to diagnose respiratory diseases with high accuracy based on objective criteria. Utilizing the technologies, we should prepare for new pandemics.Specialists diagnose respiratory diseases by listening for peculiar sounds from the auscultatory sounds of patients who are suffering from lung disease. The study proposed in the paper transforms lung sounds collected by auscultation into a spectral image using the short-time Fourier transform. If auscultatory sounds contain disease-specific sounds, specific features should also appear in the spectral image of lung sounds. Deep learning techniques for analyzing images have made remarkable progress.Images can provide objective judgment criteria even to non-specialists. Analysis of images allows both specialists and non-specialists to diagnose objectively, unaffected by hearing loss due to aging. Doctors have accountability for patients on diseases. Images have comprehensive explanatory power for patients.Only a short-time Fourier transform of the spectral image of auscultatory sounds does not sufficiently highlight features specific to respiratory disease. The proposed method converts auscultatory sounds from the lung into a spectral image that emphasizes the frequency region of the sound recognizable to humans. The study refers to it as a mel-spectrogram, which facilitates finding disease features. The proposed method detects disease-specific features appearing in mel-spectrograms with Yolo, an object detection technique based on deep learning. The proposed method has discriminated auscultatory sounds obtained from actual patients with an accuracy of 0.7 in the F1-Score.Deep learning analysis of images provides evaluation criteria that are objective and independent of the skill of doctors. This study will enable non-specialists in respiratory medicine to examine whether persons are suffering from respiratory diseases, which would eliminate the shortage of specialists. This is diagnostic support for nonspecialists to address the explosion of patients due to respiratory infection outbreaks in the pandemic. It contributes to preventing the collapse of health care.
Keywords: Deep Learning, objective criteria, preventing the collapse of health care, Specialists, recognizable to human
Cite this paper: