Med-ViX-Ray: Enhancing explainable chest X-ray analysis with clinical knowledge graphs.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models excel in chest X-ray interpretation but often lack transparency.
• This opacity hinders trust and clinical adoption of AI in radiology.
• Interpretable AI is crucial for integrating advanced diagnostics into clinical workflows.

Purpose of the Study:

• Introduce Med-ViX-Ray, a knowledge-guided framework for interpretable chest X-ray classification.
• Integrate symbolic clinical reasoning with a vision Transformer backbone.
• Enhance transparency beyond traditional heatmaps by explaining predictions with clinical signs.

Main Methods:

• Utilized a structured graph of radiological signs and conditions.
• Employed a probabilistic soft-matching module to align image attention with domain knowledge.
• Implemented a nudging mechanism to refine classifier outputs based on clinical reasoning.

Main Results:

• Med-ViX-Ray demonstrated improved recall and F1-score compared to a SwinV2 baseline.
• Achieved competitive overall performance with enhanced interpretability.
• Qualitative analysis confirmed clinically relevant region highlighting and sign-activation.

Conclusions:

• Knowledge-guided attention and sign-based explanations improve interpretability and recall in chest X-ray models.
• The framework offers a path towards more trustworthy and explainable AI in medical diagnostics.
• Future work includes extending the framework for report generation and clinical validation.