RespireNet: A Deep Neural Network for Accurately Detecting Abnormal Lung Sounds in Limited Data Setting

Area of Science:

• Medical technology
• Artificial intelligence
• Pulmonology

Background:

• Auscultation is key for diagnosing lung diseases, but manual analysis is subjective.
• Automated analysis with digital stethoscopes can enable remote lung disease screening.
• Deep neural networks (DNNs) show promise but require large datasets, which are scarce in respiratory research.

Purpose of the Study:

• To develop an efficient deep learning model for respiratory sound analysis using limited data.
• To improve the accuracy of automated lung disease classification.

Main Methods:

• Proposed RespireNet, a Convolutional Neural Network (CNN)-based model.
• Implemented novel techniques: device-specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding.
• Evaluated performance on the ICBHI dataset for 4-class classification.

Main Results:

• RespireNet demonstrated efficient use of a small dataset (6898 breathing cycles).
• Achieved a 2.2% improvement over state-of-the-art results for 4-class classification.
• Successfully addressed the data scarcity challenge for training deep learning models in respiratory acoustics.

Conclusions:

• RespireNet offers a viable solution for automated lung disease screening via respiratory sound analysis.
• The proposed techniques enable effective DNN training with limited respiratory datasets.
• This approach has the potential to enhance tele-screening of fatal lung diseases.