Exploring the Effect of Race in Automated Skin Cancer Detection

With the rising incidence of skin cancer, its automated detection can help in timely intervention, leading to a reduced burden on the healthcare system. However, majority of related research is focused on Caucasian populations which may limit its application in skin of colour. In this paper, we attempt to show that using AI models trained on predominantly Caucasian population do not perform equally well in Asian population of Fitzpatrick skin types II-III. We trained an EfficientNetB2 deep learning model on ISIC 2019 dataset (predominantly Caucasian) and evaluated it on ISIC 2019 (using cross-validation), PH² (predominantly Caucasian) and our own dataset primarily consisting of Asians. The area under receiver operator characteristics or AUC was 0.90 on ISIC 2019 and 0.91 on the PH² dataset, while it dropped to 0.81 for our dataset and did not improve upon fine tuning.Clinical relevance- Our results show that existing AI-based methods for skin cancer detection may not be optimized for Asian skin. This highlights the importance of specifically studying skin cancer in Asian skin.