Towards Equitable Diagnosis: Bias Evaluation and Mitigation in Skin Cancer Classification

Abstract

The introduction of artificial intelligence into health-care has significantly increased diagnosis rates, reduced diagnosis times, and expanded treatment options. However, the employment of AI in disease diagnosis has sparked concerns about its capability to diagnose accurately and fairly for marginalized groups that are underrepresented in datasets. A worrisome example is the discrepancy in skin cancer diagnosis accuracies between patients with darker and lighter skin tones. In this study, we employ the Kruskal-Wallis hypothesis test to quantify biases that may arise during AI model training and assess various strategies to mitigate them. We find that augmentations tailored to address the under-representation of darker skin tones in datasets could significantly alleviate bias in AI models. Additionally, our investigation unveils notable variations in the effectiveness of different augmentation techniques in reducing bias, with Color Jitter emerging as a standout augmentation technique that can effectively eliminate bias. Furthermore, our analysis of the interplay between light and dark skin images during training underscores the importance of incorporating diverse skin tones in the training process to narrow the accuracy gap and mitigate bias across different skin colors.