Fairness Evaluation Using Augmented Resnet Models

Artificial intelligence (AI) has transformed dermatology by enabling scalable diagnostic tools, but its effectiveness has often been limited by poor performance on darker skin tones. Prior studies have found that underrepresentation of Fitzpatrick Types V and VI in training datasets contributes to systemic diagnostic bias in dermatology AI models. This study examines whether standard image augmentation techniques can reduce this bias without using fairness-specific methods. A ResNet18-based multi-task model was trained on the Fitzpatrick17k dataset to classify both skin disease and Fitzpatrick skin type. The training process included normalization and random image transformations such as flips and rotations. Unexpectedly, the model achieved its highest classification accuracy on darker skin tones, suggesting that even untargeted augmentation may improve performance for underrepresented groups. These results challenge prior assumptions that fairness can only be achieved through complex algorithmic interventions. The findings have important implications for the development of equitable dermatological AI systems. Simple, resource-efficient techniques like data augmentation may provide a practical first step toward reducing algorithmic bias in healthcare settings where fairness infrastructure is limited.