Lesion Location Attention Guided Network for Multi-Label Thoracic Disease Classification in Chest X-Rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Traditional clinical diagnosis relies on identifying abnormalities in chest X-rays (CXRs).
• Attention mechanisms in deep learning have shown promise for improving diagnostic accuracy by focusing on relevant image regions.
• Existing methods may not optimally leverage lesion location information for multi-label thoracic disease classification.

Purpose of the Study:

• To propose a novel lesion location attention guided network (LLAGnet) for multi-label thoracic disease classification in CXRs.
• To integrate region-level attention (RLA) and channel-level attention (CLA) for enhanced feature extraction from lesion locations.
• To improve the accuracy and effectiveness of automated thoracic disease diagnosis using deep learning.

Main Methods:

• Developed LLAGnet, a network incorporating both RLA and CLA modules.
• The RLA module utilizes a weakly supervised attention mechanism for lesion localization and fine-grained feature extraction.
• The CLA module adaptively weights channel-wise features based on interdependencies, focusing on lesion-relevant channels.

Main Results:

• LLAGnet demonstrated superior performance in multi-label thoracic disease classification on the ChestX-ray14 dataset.
• The integrated RLA and CLA modules effectively focused on discriminative features from lesion locations.
• The proposed method outperformed state-of-the-art baseline models in experimental evaluations.

Conclusions:

• LLAGnet offers an effective approach for leveraging lesion location attention in CXR analysis.
• The integration of RLA and CLA enhances the model's ability to capture crucial diagnostic information.
• The proposed method holds significant potential for improving automated diagnosis of thoracic diseases.