Attention-Fusion-Based Two-Stream Vision Transformer for Heart Sound Classification

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Heart sound classification (HSC) is crucial for cardiovascular disease diagnosis.
• Existing methods often use single-stream architectures, missing multi-resolution feature benefits.
• Current multi-stream approaches struggle with cross-modal interactions and information loss during fusion.

Purpose of the Study:

• To develop a novel attention fusion-based two-stream Vision Transformer (AFTViT) for enhanced heart sound classification.
• To effectively capture and integrate multi-resolution and cross-modal features in heart sound signals.
• To overcome limitations of conventional fusion methods in existing HSC architectures.

Main Methods:

• Proposed an AFTViT architecture utilizing two-dimensional mel-cepstral domain features.
• Employed a Vision Transformer (ViT)-based encoder for capturing long-range dependencies and multi-scale contextual information.
• Introduced a novel attention block for feature-level integration of cross-context features.

Main Results:

• The AFTViT architecture demonstrated superior performance compared to state-of-the-art CNN-based methods on PhysioNet2016 and PhysioNet2022 datasets.
• Achieved higher accuracy in heart sound classification tasks.
• The attention fusion mechanism effectively enhanced feature representation by integrating cross-contextual information.

Conclusions:

• The AFTViT framework shows significant potential for improving the accuracy of heart sound classification.
• This approach offers a promising tool for the early diagnosis of cardiovascular diseases.
• The study highlights the efficacy of attention-based fusion in multi-stream Vision Transformer architectures for biomedical signal processing.