Generative Adversarial Neural Networks for Pigmented and Non-Pigmented Skin Lesions Detection in Clinical Images

Abstract

Skin cancer is one of the most common malignancies in fair skin population. It can be divided in two main classes: melanoma and non-melanoma skin cancer. This means that pigmented and, also, non-pigmented skin lesions might raise an important risk. Due to the fact that melanoma is more aggressive, pigmented skin lesions gained more attention in terms of automatic diagnosis. One of the most important steps in this procedure is to correctly identify the skin lesion in an image (acquired with a dermoscope or a standard camera). Based on the accurate identification of the lesion, specific automatic algorithms for cancer diagnosis can be developed. In this paper we propose and evaluate an artificial intelligence method for pigmented and non-pigmented lesion segmentation. The method uses generative adversarial neural networks. The network was trained and tested on a large set of images acquired with smartphone cameras. The results show that approximately 92% of the lesions are correctly identified on the test set.