Skin Lesion Segmentation by using Deep Learning Techniques

Skin cancer is a common disease among middle-aged and elderly white-skinned people. It is divided into many types in terms of medical criteria. Malign melanoma is one of the most dangerous and fatal cancer types, it can be treated if detected early. The main focus of this paper is to provide a precise, effective, robust and automated way to segment the lesion in order to facilitate the classification of the lesion with high accuracy during the early diagnosis of skin cancer. This process consists of two stages. In the first stage, image processing techniques (Image Enhancement, Linear Filtering, and Image Restoration) are used to obtain images free of artifacts such as hair and ruler marks. The second stage which is the important part of this paper is to modify a U-Net architecture and propose a 46-layered structure of U-Net to obtain a successful lesion segmentation rate. In this study, experiments were performed on two different U-Net architectures (U-Net 32, and U-Net 46).