Deep Learning-Based Classification of Common Lung Sounds via Auto-Detected Respiratory Cycles

Area of Science:

• Medical Technology
• Artificial Intelligence
• Signal Processing

Background:

• Chronic respiratory diseases are a leading cause of global mortality.
• Early diagnosis via lung sound auscultation is crucial but requires expert auditory skills.
• Artificial intelligence (AI) shows promise in classifying lung sounds (LSs) with high accuracy.

Purpose of the Study:

• To compare various AI-based methods for classifying lung sounds.
• To evaluate different time-frequency representations for LS analysis.
• To determine the optimal AI architecture and feature representation for accurate LS classification.

Main Methods:

• Automated detection and labeling of respiratory cycles from lung sounds.
• Utilized time-frequency representations: spectrograms, scalograms, Mel-spectrograms, and gammatonegrams.
• Employed transfer learning with pre-trained convolutional neural networks (CNNs) and compared CNN, CNN-LSTM hybrid, and support vector machine models.

Main Results:

• Gammatonegrams, with their high fidelity in low-frequency spectral structure and noise resistance, were analyzed.
• Combining gammatonegrams with a CNN architecture yielded the highest classification accuracy of 97.3% ± 1.9%.
• This AI approach demonstrated superior performance compared to other methods evaluated.

Conclusions:

• AI, particularly CNNs combined with gammatonegrams, offers a highly accurate and reliable method for lung sound classification.
• This technology can aid in the early and non-invasive diagnosis of chronic respiratory diseases.
• The study highlights the potential of advanced signal processing and deep learning in respiratory medicine.