LMeRAN: Label Masking-Enhanced Residual Attention Network for Multi-Label Chest X-Ray Disease Aided Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Chest X-ray (CXR) is vital for diagnosing thoracic diseases.
• Current computer-aided diagnosis (CAD) systems often miss global context and label correlations.

Purpose of the Study:

• To develop a novel framework, LMeRAN, for multi-label CXR image classification.
• To address limitations in existing CAD systems regarding feature representation and label dependency modeling.

Main Methods:

• Proposed the Label Masking-enhanced Residual Attention Network (LMeRAN).
• Introduced label-specific residual attention and multi-head self-attention with average pooling.
• Implemented a label mask training strategy to learn label dependencies.

Main Results:

• LMeRAN achieved the highest mean AUC of 0.825 on the ChestX-ray14 dataset.
• Demonstrated performance improvements of 3.1% to 8.0% over advanced baselines.
• Visualized lesion regions for enhanced model interpretability.

Conclusions:

• LMeRAN effectively captures both local and global features in CXR images.
• The framework successfully models label correlations, improving diagnostic accuracy.
• LMeRAN offers a promising advancement in automated thoracic disease diagnosis.