Evaluating Explainable Artificial Intelligence (XAI) techniques in chest radiology imaging through a human-centered...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning (DL) is increasingly used in radiology for diagnostics.
• Explainable AI (XAI) aims to enhance transparency and trust in DL models.
• Challenges remain in healthcare adoption of XAI due to technical and human-centered evaluation gaps.

Purpose of the Study:

• To evaluate visual XAI techniques for DL-based chest radiography diagnostics.
• To assess user perception and preferences for XAI methods from a human perspective.
• To identify user-driven requirements for effective XAI system integration in clinical practice.

Main Methods:

• Developed DL models for pneumonia and COVID-19 detection in chest X-rays/CT scans.
• Applied Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME) for visual explanations.
• Conducted a user study with medical professionals evaluating XAI outputs for clinical relevance, coherency, and trust.

Main Results:

• DL models achieved 90% accuracy for pneumonia and 98% for COVID-19.
• Participants generally perceived XAI systems positively but showed limited awareness of their practical value.
• Grad-CAM outperformed LIME in coherency and trust, though clinical usability concerns were noted.

Conclusions:

• Medical professionals favor XAI systems in chest radiography, but awareness and practical understanding need improvement.
• User-driven requirements emphasize multi-modal explainability and inclusive design for better clinical integration.
• Enhancing practitioner awareness and addressing usability are crucial for widespread XAI adoption in radiology.