Interpretation of lung disease classification with light attention connected module.

Area of Science:

• Medical Diagnostics
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Lung diseases pose significant health risks, exacerbated by factors like the COVID-19 pandemic.
• Accurate diagnosis of lung conditions relies on interpreting respiratory sounds, a process often dependent on clinician experience.
• There is a need for objective, AI-driven tools to supplement traditional stethoscope-based lung disease diagnosis.

Purpose of the Study:

• To develop and evaluate a deep learning-based artificial intelligence model for classifying lung diseases using respiratory sounds.
• To enhance classification performance by incorporating an attention module, specifically the Efficient Channel Attention module (ECA-Net).
• To analyze the model's diagnostic capabilities and compare its performance against expert opinions and existing methods.

Main Methods:

• Respiratory sounds were processed using log-Mel spectrogram MFCC (Mel-frequency cepstral coefficients).
• A VGGish model was improved and integrated with a light attention module (ECA-Net) for classification.
• Lung disease classification performance was assessed using metrics including accuracy, precision, sensitivity, specificity, f1-score, and balanced accuracy.
• Gradient-weighted class activation mapping (Grad-CAM) was employed for analyzing classification decisions.

Main Results:

• The proposed AI model achieved high performance metrics: 92.56% accuracy, 92.81% precision, 92.22% sensitivity, 98.50% specificity, 92.29% f1-score, and 95.4% balanced accuracy.
• The attention module significantly contributed to the model's high performance in classifying normal and five types of adventitious respiratory sounds.
• Analysis using Grad-CAM provided insights into the model's classification reasoning, and its performance was validated against open lung sounds and expert evaluations.

Conclusions:

• The developed AI model demonstrates significant potential for the objective and accurate classification of lung diseases based on respiratory sound analysis.
• The integration of attention mechanisms enhances the model's ability to interpret complex respiratory sound patterns.
• This AI-powered approach can serve as a valuable tool for early disease detection and interpretation, potentially integrated into smart stethoscopes for improved clinical practice.