A Novel Approach to Promote Equity in Skin Disease Diagnosis by AI Models

The goal is to increase the accuracy of existing models in diagnosing skin disease across various skin tones within 10% of that obtained in diagnosing fairer skin tones, which is about 95.8%.

AI-based systems are increasingly used to diagnose skin diseases with datasets available on the internet. However, the training data predominantly represents fair skinned people. The goal is to increase the accuracy of existing models in diagnosing skin disease across various skin tones within 10% of that obtained in diagnosing fairer skin tones, which is about 95.8%. The publicly available HAM10000 dataset with 10000 fair skin images was preprocessed and used to fine-tune a Vision Transformer model trained on the ImageNet-21K dataset. The model obtained about 94.9% accuracy in diagnosing fair skin images and about 19.8% accuracy in diagnosing real non-fair skin images. Next, 50% of the training data was transformed into non-fair skin simulation by CycleGAN based on 320 real non-fair images. After several iterations, the model achieved about 78.6% accuracy in classifying any skin tone and 52.1% accuracy in classifying real non-fair images. Though this 68.4% is outside of the desired range, it still falls within the accuracy of dermatologists (48–77%). However, it is the accuracy in classifying non-fair skin that makes this model significant. The model achieved a 28.6% increase in accuracy in diagnosing non-fair images using simulated data thereby narrowing the disparity. A confusion matrix was plotted to visualize the validity of the predictions. The functionality of this model suggests that similar data augmentation techniques could be applied to other AI models to ensure their fairness to all categories of people and correct any biases in data due to historic under-representation.

Explore More!

Research Paper