Segmentation and Classification of Dermoscopic Skin Images Using U-Net and Handcrafted Features

The most common type of cancer is skin cancer, which accounts for more than half of all cancer diagnoses. Melanoma is the rarest type of skin cancer, but it is the deadliest, with a high risk of spreading and death. On the other hand, melanoma is almost always treatable if detected early. In this regard, computer approaches for processing and analyzing skin lesion images have been investigated and developed. Our approach attempts to achieve this feat by developing a new method of melanoma segmentation and classification. The procedure begins with the preparation of dermoscopic images to remove hair using the Dull Razor algorithm, followed by image segmentation using a U-Net convolutional neural network to separate the skin lesion from the background. The segmented images obtained with the deep learning algorithm U-Net are then used to compute the Gray Level Co-occurrence Matrices (GLCM), followed by extracting Asymmetry, Border, Color, and Diameter (ABCD rule) features of the region of interest. Our approach achieves good results compared to the state-of-the-art.