CAMEO: A Class-Activation-Mapped Equitable Overlay Framework for Fair and Robust Deep Learning-based Skin Condition Diagnosis

XAI can actively improve fairness, not just audit it: Rather than using explainable AI only to identify problems after a model is built, CAMEO repurposes it as a training tool that actively corrects bias. The framework uses model explanations to identify and remove spurious background cues during training, demonstrating a practical pathway for integrating fairness into model development from the start.

AI skin lesion classifiers were relying on the wrong features: The research found that deep learning models for diagnosing skin conditions were achieving high accuracy by focusing on background elements like skin tone, device artifacts, and rulers rather than actual lesion characteristics. This means models could appear accurate while being fundamentally unreliable for real-world use.

Fairness improvements don't require sacrificing accuracy: CAMEO maintained diagnostic accuracy while reducing background-driven errors by nearly four times on both standard and dark-skin datasets. This challenges the common assumption that making AI systems fairer requires trading off performance, showing these goals can be compatible.

The approach works across different skin tones: By replacing backgrounds with realistic synthetic skin during training, the model became more robust when backgrounds changed and more consistent in its attention patterns regardless of skin tone. This suggests the framework addresses fairness systematically rather than through tone-specific fixes.

Transparency about model reasoning enables better solutions: By understanding why models were making mistakes through their visual explanations, researchers could design targeted interventions. This highlights how explainability serves as a foundation for building more trustworthy AI systems.