RespireNet

RespireNet automates detection of abnormal lung sounds from recorded respiratory audio using deep neural networks, including convolutional neural networks, to support lung disease classification.

Key Features:

• Automated Auscultation Analysis: Uses deep neural networks, specifically convolutional neural networks (CNNs), to analyze respiratory sounds for abnormality detection.
• Efficient Use of Limited Data: Implements strategies to maximize model performance on small respiratory datasets such as the ICBHI dataset containing 6,898 breathing cycles.
• Device-Specific Fine-Tuning: Adapts model parameters to specific recording devices to improve accuracy and generalizability across devices.
• Concatenation-Based Augmentation: Increases effective dataset size by intelligently combining audio segments.
• Blank Region Clipping: Removes non-informative silent or blank regions from recordings to focus training on relevant signal content.
• Smart Padding: Applies padding strategies to ensure uniform input sizes for model training without distorting signal characteristics.
• Performance Improvements: Demonstrates a 2.2% improvement in 4-class classification accuracy on the ICBHI dataset compared with prior state-of-the-art.

Scientific Applications:

• Abnormal Lung Sound Classification: Enables research-grade classification of respiratory audio into multiple pathological classes (4-class classification).
• Tele-screening and Remote Monitoring: Supports integration with digital stethoscopes for remote respiratory screening and monitoring workflows.
• Benchmarking on ICBHI: Provides an evaluated approach for comparison on the ICBHI dataset and other limited respiratory sound corpora.
• Augmentation and Preprocessing Research: Serves as a testbed for evaluating concatenation-based augmentation, blank region clipping, and padding strategies on small audio datasets.

Methodology:

RespireNet employs convolutional neural networks and other deep neural networks with device-specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding, and it was evaluated for 4-class classification on the ICBHI dataset (6,898 breathing cycles).