Melanoma Skin Cancer Classification Using Transfer Learning

Melanoma is one of the most aggressive types of skin cancer as it rapidly spreads to various areas of the body. With the increase and fatal nature of melanoma, it is of utmost importance to establish computer assisted diagnostic support systems to aid physicians in diagnosing skin cancer. In this paper, we make use of deep learning and transfer learning by testing 14 pre-trained models for the classification and detection of skin cancer. Historically, the data in which Deep Convolutional Neural Networks are fed and trained on comes predominantly from European datasets resulting in biased data. To overcome this issue, we first determine the differences of melanoma that lie within people of different skin tones. Thereafter, we make use of the GrabCut segmentation technique to accurately segment the lesion from the surrounding skin tone in order to solely focus on the lesion. The pre-trained CNN, Squeezenet1-1, achieved the best experimental results with an accuracy rate of 93.42%, sensitivity of 92.11% and specificity of 94.74%. The experimental results achieved indicate that there is a possible solution to the underrepresented data of dark-skinned people.