CLARE-XR: explainable regression-based classification of chest radiographs with label embeddings

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Accurate pathology classification in chest X-rays is crucial for clinical decision-making.
• Existing systems often lack transparency, hindering end-user trust and regulatory compliance.
• Interpretability is essential for the adoption of AI in medical diagnostics.

Purpose of the Study:

• To introduce CLARE-XR, a novel methodology for interpretable chest X-ray pathology classification.
• To provide visual explanations for diagnoses by retrieving similar cases.
• To enhance end-user trust and support clinical decision-making.

Main Methods:

• Developed a regression model to map chest X-ray images into a 2D latent space.
• Generated reference coordinates for pathologies via label embedding of ground-truth annotations.
• Inferred classification based on the distance to reference coordinates and retrieved similar training images.

Main Results:

• CLARE-XR demonstrated superior performance compared to a ResNet50 baseline on the NIH ChestX-ray14 dataset.
• The system successfully identified pathologies and provided explanations through case retrieval.
• The methodology offers inherent interpretability and discloses classification rules.

Conclusions:

• CLARE-XR offers an interpretable framework for chest X-ray analysis, moving beyond predictive performance.
• The system mimics clinical practice by comparing current scans with similar past cases.
• This approach enhances trust and aids regulatory compliance in AI-driven medical diagnostics.