Learning Hierarchical Attention for Weakly-Supervised Chest X-Ray Abnormality Localization and Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning advances medical imaging but faces challenges in clinical adoption due to lack of interpretability.
• Physicians require trustworthy AI tools with clear decision reasoning for clinical use.
• Current methods often lack detailed abnormality localization, hindering clinical application.

Purpose of the Study:

• To develop a weakly supervised algorithm for accurate abnormality localization in clinical applications.
• To enhance the interpretability and trustworthiness of deep learning models in medical imaging.
• To reduce the need for expensive, expert-annotated localization data.

Main Methods:

• A novel attention-driven, weakly supervised algorithm utilizing a hierarchical attention mining framework.
• Integration of activation- and gradient-based visual attention mechanisms.
• Implementation of explicit ordinal attention constraints for principled model training and explanation generation.

Main Results:

• Significant improvements in abnormality localization performance on large-scale chest X-ray datasets (NIH ChestX-ray14 and CheXpert).
• Achieved competitive classification performance alongside enhanced localization capabilities.
• Demonstrated the generation of visual-attention-driven model explanations using localization cues.

Conclusions:

• The proposed algorithm effectively addresses the challenge of abnormality localization in medical imaging.
• Weakly supervised learning with attention mechanisms can yield high-performance, interpretable AI tools.
• This approach offers a viable solution for enhancing clinical trust and utility of deep learning in radiology.