An Ensemble of Fully Convolutional Neural Networks for Automatic Skin Lesion Segmentation

Accurate extraction of skin lesion borders in dermoscopic images is a prerequisite step in skin lesion classification. However, it is a challenging task due to the great intraclass variation, the high degree of interclass visual similarity, low contrast between skin lesion and surrounding normal skin, and the existence of extraneous and intrinsic artifacts. To overcome these challenges, an ensemble of fully automatic semantic segmentation method based on fully convolutional neural networks is presented for skin lesion segmentation in this paper. For this purpose, firstly, FCN, U-Net, SegNet and DeepLabV3+ networks are trained in end-to-end manner without any pre- or post-processing steps. Then, majority voting approach, which is one of ensemble method, is applied to the acquired semantic segmentation maps. Furthermore, we utilize inverse class frequency weighting in order to deal with class imbalance problem. The effectiveness and efficiency of the presented methods are evaluated on International Skin Imaging Collaboration (ISIC) 2017 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset. The experimental results show that the proposed ensemble network achieves a Jaccard index of 92.68%, a sensitivity of 98.08%, a specificity of 79.86% and an accuracy of 93.99%.