Breathing Sound Segmentation and Detection Using Transfer Learning Techniques on an Attention-Based Encoder-Decoder...

Area of Science:

• Medical Informatics
• Biomedical Signal Processing
• Artificial Intelligence in Medicine

Background:

• Pulmonary diseases significantly impact respiratory function, necessitating accurate diagnostic tools.
• Interpreting respiratory sounds is crucial for diagnosing lung conditions, but can be subjective.
• Automated analysis of breathing sounds offers potential for objective and precise diagnosis.

Purpose of the Study:

• To develop and evaluate an attention-based encoder-decoder model for precise segmentation and detection of inspiratory and expiratory breathing sounds.
• To enhance diagnostic accuracy for pulmonary diseases through interpretable respiratory sound analysis.
• To leverage deep learning for objective assessment of lung sound signals.

Main Methods:

• Utilized an attention-based encoder-decoder architecture for respiratory sound analysis.
• Trained the model using spectrograms of lung sound signals with time-segment labels.
• Employed ResNet101 as the encoder for feature extraction.
• Recorded respiratory sounds from 22 participants using digital stethoscopes and anti-noising microphones.

Main Results:

• Achieved a high accuracy of 92.006% for breathing sound segmentation and detection using 0.5-second time segments.
• Demonstrated consistent performance through ten-fold cross-validation experiments.
• The attention mechanism provided interpretable outputs, aiding physician diagnosis.

Conclusions:

• The proposed attention-based model effectively segments and detects breathing sounds with high accuracy.
• This approach offers a promising tool for objective diagnosis and monitoring of pulmonary diseases.
• Interpretable AI in respiratory medicine can significantly improve clinical decision-making.