SegX: Improving Interpretability of Clinical Image Diagnosis with Segmentation-based Enhancement

Current explainability tools fall short in medical imaging: Conventional AI explanation techniques like Grad-CAM and SHAP often highlight regions that aren't clinically relevant, which could mislead doctors relying on these explanations to understand why an AI system made a diagnosis.

SegX aligns AI explanations with clinical expertise: The proposed method uses segmentation models to ensure that when an AI system explains its diagnosis, it focuses on the anatomically or pathologically significant areas that doctors actually care about, rather than irrelevant image regions.

Uncertainty scores can indicate when to trust AI predictions: The paper introduces SegU, which measures how confident an AI system should be by checking whether its explanation aligns with clinically important areas—predictions with poor alignment receive lower confidence scores that could alert clinicians to be more cautious.

The approach works across different medical imaging types: Testing on both dermoscopic images (skin lesions) and chest X-rays demonstrates that SegX is generalizable, suggesting it could potentially improve interpretability across various clinical imaging applications.

This is a practical, plug-and-play solution: Because SegX works with existing models without requiring retraining, hospitals could potentially adopt this enhancement to make their current AI diagnostic tools more interpretable and trustworthy for clinical decision-making.