CT-xCOV: a CT-scan based Explainable Framework for COVid-19 diagnosis

Explainability matters for medical AI: The paper demonstrates that COVID-19 detection models can be made interpretable by combining multiple explanation techniques (Grad-Cam, LIME, and Integrated Gradients), allowing clinicians to see which parts of CT scans the AI is using to make diagnoses rather than treating it as a "black box."

Different explanation methods have varying reliability: The study found significant performance differences between explanation techniques, with Grad-Cam achieving a 55% match to actual infection areas compared to only 29% for LIME and 24% for Integrated Gradients, suggesting that not all explainability methods are equally trustworthy for medical applications.

High accuracy alone isn't sufficient for clinical deployment: While the detection model achieved 98.4% accuracy, the framework goes further by adding both visual explanations (highlighting infected lung regions) and textual explanations (percentage of infection), recognizing that clinicians need to understand why the AI made its recommendation to safely use it in practice.

Validation of explanations requires ground truth data: The paper proposes evaluating explanation quality by comparing AI-generated visualizations against actual infection patterns, establishing a standard for assessing whether explanations are genuinely reflecting what the model "sees" rather than just appearing plausible.

Responsible AI requires multiple components: The end-to-end framework combines segmentation, detection, and explanation stages, illustrating that trustworthy medical AI systems need careful attention to each step, not just optimizing for prediction accuracy.