ExplAIn: Explanatory Artificial Intelligence for Diabetic Retinopathy Diagnosis

Explainability doesn't require sacrificing accuracy: ExplAIn achieves the same diagnostic performance as traditional "black-box" AI systems while providing clear explanations for its decisions, demonstrating that transparency and effectiveness are not mutually exclusive in medical AI.

AI can learn to identify disease patterns without explicit instruction: The system automatically learned to recognize and categorize diabetic retinopathy lesions (abnormalities in eye images) from image-level labels alone, without requiring doctors to manually annotate where specific lesions appear—suggesting AI can discover clinically meaningful concepts on its own.

Explanations can be simple and visual: Rather than complex technical outputs, ExplAIn explains its diagnoses through images highlighting problem areas and brief text descriptions, making it potentially more accessible to healthcare providers and patients who need to understand why a diagnosis was made.

Transparent AI may accelerate real-world medical deployment: By combining high performance with built-in explainability, this approach addresses a major barrier to AI adoption in healthcare—the difficulty clinicians face in trusting and validating "black-box" diagnostic recommendations, potentially enabling faster integration into clinical practice.

Self-supervised learning improves diagnostic reliability: The system was trained to distinguish healthy eye tissue from diseased areas by learning what a healthy eye should look like, which improved the accuracy of identifying where lesions are located in images.