Respiratory Anomaly and Disease Detection Using Multi-Level Temporal Convolutional Networks.

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Signal Processing

Background:

• Automated analysis of respiratory sounds via Deep Learning (DL) is crucial for early lung disease detection.
• Existing DL methods often analyze spatial and temporal features of respiratory sounds separately, limiting their effectiveness.
• There is a need for advanced DL frameworks that can effectively integrate spatiotemporal information for improved diagnostic accuracy.

Purpose of the Study:

• To propose a novel Deep Learning (DL) framework, Multi-Level Temporal Convolutional Networks (ML-TCN), for enhanced analysis of respiratory sounds.
• To improve the accuracy of detecting anomalous breathing cycles and classifying respiratory conditions using lung sound audio.
• To leverage transfer learning for efficient feature extraction from limited and imbalanced respiratory sound datasets.

Main Methods:

• Developed a novel DL framework incorporating convolution operations for spatial feature extraction.
• Utilized temporal convolution networks to capture spatiotemporal correlations within respiratory sound features.
• Integrated Multi-Level Temporal Convolutional Networks (ML-TCN) for enhanced anomaly detection and classification.
• Employed a transfer learning technique for efficient semantic feature extraction from scarce and imbalanced data.

Main Results:

• The proposed ML-TCN framework demonstrated superior performance compared to state-of-the-art methods on the ICBHI 2017 dataset.
• Achieved improvements of up to 2.29% in the Score metric for binary classification of anomaly breathing cycles.
• Demonstrated improvements of up to 2.27% in average sensitivity and specificity for multi-class anomaly breathing cycle detection.
• Showcased enhanced classification accuracy: 2.69% for healthy-unhealthy binary classification and 1.47% for multi-class diagnosis.

Conclusions:

• The ML-TCN framework effectively integrates spatial and spatiotemporal features for superior respiratory sound analysis.
• The proposed method significantly enhances the accuracy of detecting abnormal breathing cycles and classifying respiratory diseases.
• Transfer learning proved effective in handling limited and imbalanced respiratory sound data.
• The ML-TCN framework presents a promising advancement for respiratory healthcare technology and early disease diagnosis.