Characterization of Melanoma Using Convolutional Neural Networks and Dermoscopic Images

Abstract

Characterization of melanoma from moles in skin lesions using dermoscopic images, guides to an efficient treatment. In this study a new approach to analyze dermoscopic images and characterize melanoma from other benign skin lesions, using Convolutional Neural Networks (CNN), is implemented. The database employed in this experimental study consists of dermoscopic skin lesion images. This database contains malignant lesions called melanomas and benign skin lesions called moles. In order to optimize the characterization of melanoma, a data augmentation process is used. The best accuracy of melanoma classification obtained in this study is equal to 81.6% comparing to 75.6% obtained using a known CNN model called LeNet-5. These results show that the suggested CNN model works well for melanoma characterization using dermoscopic images.