Lung Sound Classification Using Co-Tuning and Stochastic Normalization.

Area of Science:

• Computational intelligence
• Medical informatics
• Signal processing

Background:

• Lung sound analysis is crucial for diagnosing respiratory diseases and critical care monitoring.
• Existing computational methods face challenges with data imbalance and recording device variations.
• Deep learning models offer potential for improved lung sound classification.

Purpose of the Study:

• To leverage pre-trained ResNet models for classifying adventitious lung sounds and respiratory diseases.
• To investigate various transfer learning techniques (fine-tuning, co-tuning, stochastic normalization) for optimal model performance.
• To address data imbalance and recording device variability using data augmentation and spectrum correction.

Main Methods:

• Utilized pre-trained ResNet architectures as feature extractors.
• Applied transfer learning strategies: vanilla fine-tuning, co-tuning, stochastic normalization, and combined approaches.
• Implemented time and time-frequency domain data augmentation.
• Introduced spectrum correction to mitigate recording device inconsistencies.

Main Results:

• The proposed systems, particularly those combining co-tuning and stochastic normalization with data augmentation and spectrum correction, demonstrated superior performance.
• Achieved state-of-the-art results in classifying adventitious lung sounds and respiratory diseases on both ICBHI and a multi-channel dataset.
• Effectively handled class imbalance and recording device variations, enhancing model robustness.

Conclusions:

• Transfer learning with ResNet backbones, augmented by advanced techniques, significantly advances lung sound classification.
• The developed computational methods provide a robust framework for computer-aided diagnosis of respiratory conditions.
• This work offers a promising direction for improving critical care monitoring and diagnostic support systems.