Skin lesion semantic segmentation is a vital process that aims to identify each pixel in the input image whether it belongs to the foreground (lesion skin) or background (normal skin). Skin lesion image segmentation is an essential step in the medical image

analysis domain for use in radiotherapy to enhance diagnostic radiology. Misclassified border pixels cause a significant reduction in the global accuracy because they maybe belong to the foreground or background. The aim of this paper is to improve the skin

segmentation results at border pixels by building a deep fully convolutional network fed with gradient skin images instead of traditional color images. The proposed segmentation network produces a binary predicted output image with efficient inference at all image pixels while giving extra attention to border pixels. The appropriate gradient components of the input skin image are employed to train one

of the famous deep convolutional neural networks called U-Net with some modifications. The dice loss function is utilized to train the network instead of the cross-entropy network in order to improve the performance segmentation results, especially in the border pixels. Several experiments are conducted using the ISIC 2018 dataset to evaluate the performance of the proposed network compared to other

state-of-the-art approaches.