Feature-Based Fusion Using CNN for Lung and Heart Sound Classification.

Area of Science:

• Medical acoustics
• Biomedical signal processing
• Artificial intelligence in healthcare

Background:

• Lung and heart sound classification is complex due to audio data's dynamic nature and challenges with small, noisy, or imbalanced datasets.
• Data quality significantly impacts deep learning model performance in biomedical acoustics.
• Existing methods struggle with the intricate time and frequency characteristics of physiological sounds.

Purpose of the Study:

• To propose a novel feature-based fusion network, FDC-FS, for enhanced heart and lung sound classification.
• To leverage transfer learning from multiple deep neural network models for improved diagnostic accuracy.
• To address data limitations and noise issues inherent in biomedical audio datasets.

Main Methods:

• Developed the FDC-FS framework for classifying heart and lung sounds using feature-based fusion.
• Transformed audio data into image vectors and fused features from three distinct deep learning models.
• Utilized data augmentation techniques (noise distortion, pitch shift, time stretching) on public lung and heart sound datasets.
• Extracted Spectrogram, MFCC, and Chromagram features and fed them into fused convolutional neural network models.

Main Results:

• The FDC-FS model demonstrated superior performance compared to state-of-the-art methods.
• Achieved a maximum accuracy of 99.1% for Spectrogram-based lung sound classification.
• Reached 97% accuracy for Spectrogram and Chromagram-based heart sound classification.

Conclusions:

• The proposed FDC-FS fusion model effectively addresses challenges in heart and lung sound classification.
• Transfer learning and feature fusion significantly improve deep learning model performance on biomedical audio data.
• The FDC-FS framework offers a promising approach for accurate and robust detection of cardiopulmonary conditions.