Active Globally Explainable Learning for Medical Images via Class Association Embedding and Cyclic Adversarial Generation

The paper addresses a critical gap in explainable AI for medical imaging: Current explainability methods struggle to extract global knowledge about what makes an image belong to a particular disease category, resulting in unclear explanations that lack important context—a significant limitation when AI systems are used to support medical diagnoses.

The proposed method generates "counterfactual" examples to show what would need to change for a diagnosis to differ: By creating synthetic images that preserve a patient's individual characteristics (like body type or imaging angle) while systematically changing disease-related features, the system can show clinicians exactly which visual patterns distinguish one condition from another—making the AI's reasoning more transparent and actionable.

This approach separates "class-related" features from "individual" features in medical images: The technology distinguishes between features that indicate disease (class-related) versus features unique to each patient (individual), allowing explanations to focus on medically relevant differences rather than irrelevant variations between patients.

The method enables active learning where the AI can identify which cases need human review: By understanding the decision boundaries between disease categories, the system can flag uncertain or edge-case diagnoses for clinician review, potentially improving both safety and efficiency in medical AI deployment.

Counterfactual explanations provide intuitive, global understanding of classification tasks: Rather than highlighting individual pixels, this approach shows complete synthetic examples of "what would need to change," which may be easier for medical professionals to understand and validate than traditional saliency maps.