Addressing Racial Bias in HAM10000 with Fitzpatrick Skin Tone Labeling for Skin Cancer Classification

Skin cancer classification can be addressed with transfer learning and deep learning techniques that are fast and cost-effective, in contrast to conventional diagnostic methods. In this work, three different transfer learning models- ResNet, VGG, and DenseNet were trained on a recent skin cancer dataset HAM10000 through weighted augmentation to evaluate model performance based on weighted average recall and then the best model was evaluated over different skin tones. Past works finetuned pre-trained models without emphasizing on reducing false negatives which is significant in healthcare domain, and so is explored in this work. DenseNet121 was found to be the best performing model with an accuracy of 0.919 and recall of 0.91. A comparison of recent transfer learning models’ performance on HAM10000 showed the DenseNet121 of this paper has better performance considering both accuracy and recall. Skin cancer datasets suffer from racial bias, and so, Fitzpatrick skin tone labeling algorithm was also applied on the HAM10000 to later evaluate the proposed model performance with respect to different skin tones. It was found that model performance degraded from fairer to darker skin type, agreeing with previous contributions in literature. Worst group accuracy was 0.789, coming from the darkest skin tone subgroup of HAM10000. This proved the under-representation of darker skin type in the HAM10000 despite being the most diverse dataset in literature as of now. Therefore, regions with populations whose skin types are not well reflected in current benchmarks must take the initiative to develop their own dataset with proper representation of all skin tones, through a collaboration between AI experts, medical practitioners, and policy-makers.