Lung Sound Classification Using Snapshot Ensemble of Convolutional Neural Networks.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Signal Processing

Background:

• Accurate lung sound classification is crucial for diagnosing respiratory diseases.
• Existing methods face challenges with class imbalance and feature extraction.

Purpose of the Study:

• To develop a robust and efficient lung sound classification system.
• To improve diagnostic accuracy for various respiratory conditions using deep learning.

Main Methods:

• Utilized a snapshot ensemble of convolutional neural networks (CNNs) for feature extraction from log mel spectrograms.
• Employed temporal stretching and vocal tract length perturbation (VTLP) for data augmentation.
• Addressed class imbalance using the focal loss objective.

Main Results:

• The proposed system achieved 78.4% micro-averaged accuracy for four classes (normal, crackles, wheezes, both).
• Achieved 83.7% micro-averaged accuracy for two classes (normal, abnormal).
• Outperformed state-of-the-art systems on the ICBHI 2017 dataset.

Conclusions:

• Snapshot ensembles of CNNs offer a robust approach to lung sound classification.
• The system demonstrates significant potential for improving respiratory diagnostics.
• Effective data augmentation and focal loss are key to handling class imbalance.