A Lightweight CNN Model for Detecting Respiratory Diseases From Lung Auscultation Sounds Using EMD-CWT-Based Hybrid
IEEE Journal of Biomedical and Health Informatics
|December 29, 2020
Summary
A new lightweight convolutional neural network (CNN) accurately classifies respiratory diseases from lung sounds. This automated analysis aids low-resource healthcare by improving diagnostic accuracy for chronic and pathological conditions.
Area of Science:
- Medical Technology
- Artificial Intelligence in Healthcare
- Respiratory Medicine
Background:
- Auscultation is crucial for diagnosing respiratory conditions.
- Automated lung sound analysis offers potential in resource-limited settings lacking specialists.
Purpose of the Study:
- To develop a lightweight convolutional neural network (CNN) for classifying respiratory diseases using lung sound analysis.
- To evaluate the CNN's performance on a patient-independent dataset.
Main Methods:
- Utilized empirical mode decomposition (EMD) and continuous wavelet transform (CWT) for hybrid scalogram-based features.
- Developed a novel lightweight CNN architecture for classifying individual breath cycles.
- Tested the model on the ICBHI 2017 lung sound dataset.
Main Results:
- Achieved 98.92% accuracy for three-class chronic classification.
- Achieved 98.70% accuracy for six-class pathological classification.
- Outperformed the VGG16 model by over 1% in accuracy for both classifications.
Conclusions:
- The proposed lightweight CNN demonstrates high accuracy in classifying respiratory diseases from lung sounds.
- This approach is effective for automated diagnosis, particularly in underserved healthcare environments.
- The model shows superior performance compared to existing lightweight and larger models.
