Skin lesion segmentation using convolutional neural networks with improved U-Net architecture

The location of skin lesions is of particular importance in the diagnosis and monitoring of skin disease. For this purpose, image segmentation could be used, for which various methods, algorithms and approaches have been proposed. Lately, many convolutional neural networks (CNN) with different architectures have been effectively employed for semantic image segmentation. In this paper, a CNN with improved U-Net architecture is introduced; this architecture is used for applying image segmentation to a dermatology image dataset including images of three different skin damage types. In the proposed method, the efficiency of the architecture is significantly improved by employing a pre-trained architecture in the encoding section and replacing some of the pooling layers. Various factors affecting the network such as the function of layers and their effects on network performance are investigated. Compared to existing CNN architectures, the proposed method attains higher stability and efficiency for the given dataset. For training data, %92 accuracy and for testing data %89 accuracy has been achieved.