Melanoma Cancer Classification using Deep Convolutional Neural Networks

Abstract

Cancerous melanoma is a relatively rare skin lesion that, if detected, can cause death due to its high mortality rate. The excessive production of melanocytes causes cancerous melanoma in the skin due to high exposure to solar radiation and poor skin care against these conditions. For this reason, we decided to use deep learning models to help detect melanoma without removing skin samples for biopsies. In this work, we proposed a new deep learning model called CNN-2, based on a deep convolutional neural network architecture to successfully classify skin lesions on a data set of 2860 skin lesions taken from the ISIC Archive. The proposed model CNN-2 was trained for 50 epochs, using a three-repeated 10-fold stratified cross-validation scheme. CNN-2 reached an AUC score of 0.915 ± 0.02. Although this model was trained for only 50 epochs, the AUC scored did not represent any statistical differences from other more complex models. Furthermore, the CNN-2 model achieved an AUC score of 0.9626 when used in a test dataset. This CNN-2 model allowed one to distinguish between benign skin lesions and melanoma.