Bias in Deep Learning Skin Cancer Detection: Parallel Residual Convolution Network Classification and Racial Bias Quantification

Abstract

Globally, skin cancer remains one of the most popular and lethal forms of cancer, significantly affecting the death rate. Several studies have been carried out regarding the automatic identification and categorization of skin cancer, employing multiple datasets with varying image compositions. The discrepancies in how different skin tones, colors, and attributes are represented in the datasets cause this variation in image compositions, often known as racial bias. The observed variations significantly influence the development of machine-learning algorithms for skin cancer detection. We have successfully detected racial biases in two datasets. In addition , we have presented a deep convolutional neural network (DCNN) model that is intended for the thorough categorization of skin cancer in several classes. Leveraging the HAM10000 and ISIC-2019, we have developed a robust and accurate model. We have attained 98.55% classification accuracy for HAM10000 dataset and 92.71% classification accuracy for ISIC-2019 dataset. Furthermore, the model's performance is evaluated against several pre-trained transfer learning models to boost efficiency.