Chest x-ray diagnosis via spatial-channel high-order attention representation learning

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Artificial Intelligence in Healthcare

Background:

• Chest x-ray analysis is crucial for computer-aided diagnosis.
• Existing methods using CNNs and Transformers have limitations in capturing channel correlations and local context.
• Effective feature representation for chest x-ray images remains a significant challenge.

Purpose of the Study:

• To propose a novel spatial-channel high-order attention (SCHA) model for chest x-ray image representation and diagnosis.
• To address limitations of existing methods by incorporating channel correlations and local context-aware features.
• To improve the accuracy and discriminative power of computer-aided diagnostic systems for chest x-rays.

Main Methods:

• Developed a context-enhanced backbone network (CEBN) using multi-head self-attention for initial feature extraction.
• Introduced a spatial-channel high-order attention (SCHA) module with spatial and channel attention learning branches.
• Implemented a local biased self-attention mechanism for spatial feature extraction and Brownian Distance Covariance for channel feature encoding.

Main Results:

• The proposed SCHA approach demonstrated superior performance on ChestX-ray14 and CheXpert datasets compared to existing methods.
• Experiments validated the effectiveness of integrating spatial and channel high-order attention for enhanced feature representation.
• The model achieved better performance in multi-label diagnosis classification and prediction.

Conclusions:

• The SCHA model offers a more discriminative approach for chest x-ray classification.
• This study provides an effective technique for advancing computer-aided diagnosis in medical imaging.
• The findings highlight the importance of channel correlations and local context in chest x-ray analysis.