Skin lesion segmentation with deep learning

Abstract

Skin lesion segmentation is an important process in skin diagnostics because it improves manual and computeraided diagnostics by focusing the medical personnel on specific parts of the skin. Image segmentation is a common task in computer vision that partitions a digital image into multiple segments, for which deep neural networks have been proven to be reliable. In this paper, we investigate the applicability of deep learning methods for skin lesion segmentation evaluating three architectures: a pre-trained VGG16 encoder combined with SegNet decoder, TernausNet, and DeepLabV3+. The data set consists of images with RGB skin lesions and the ground truth of their segmentation. All the image sizes vary from hundreds to thousands of pixels per dimension. We evaluated the approaches with the Jaccard index and the computational efficiency of the training. The results show that the three deep neural network architectures achieve Jaccard Index scores of above 0.82, while the DeeplabV3+ outperforms the other approaches with a score of 0.876. The results are encouraging and can lead to fully-fledged automated approaches for skin lesion segmentation.