Heart Sound Classification Based on Multi-Scale Feature Fusion and Channel Attention Module

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) show promise for intelligent heart sound diagnosis, but their performance is limited by model parameters and structure.
• Existing CNN models for heart sound classification have room for improvement in accuracy and efficiency.
• Addressing the challenge of limited datasets is crucial for developing robust heart sound diagnostic models.

Purpose of the Study:

• To propose CAFusionNet, a novel heart sound classification model that fuses features from different CNN layers.
• To improve the accuracy and efficiency of intelligent heart sound diagnosis using advanced deep learning techniques.
• To leverage transfer learning to overcome the limitations of small datasets in medical applications.

Main Methods:

• Developed CAFusionNet, a model that fuses features from varying resolutions and receptive field sizes across different CNN layers.
• Incorporated a channel attention block to weight critical features for heart valve disease detection at each layer.
• Applied a homogeneous transfer learning approach to mitigate the impact of limited dataset size.
• Utilized a combined dataset of public and proprietary data for model training and evaluation.

Main Results:

• CAFusionNet achieved an accuracy of 0.9323 on a combined dataset, outperforming existing models.
• The transfer learning approach resulted in an accuracy of 0.9665 for the triple classification task.
• Visualized heat maps confirmed the significance of feature fusion from multiple layers.
• The proposed methods demonstrated substantial enhancement in heart sound classification performance.

Conclusions:

• Feature fusion from different layers is critical for improving heart sound classification accuracy.
• CAFusionNet, combined with transfer learning, offers a powerful approach for intelligent heart sound diagnosis.
• The study highlights the potential of deep learning and attention mechanisms in cardiovascular diagnostics.