DU-Net+: a fully convolutional neural network architecture for semantic segmentation of skin lesions

Automatic skin lesion segmentation significantly influences the accuracy of skin cancer detection and classification. In this paper, we proposed a robust convolutional neural network based architecture called double U-Net+ (DU-Net+) for the segmentation of skin lesions. The proposed architecture comprises two modified U-Nets that learn incredible details through the additional skip connections and emphasize the contextual features using atrous spatial pyramid pooling (ASPP). The novel contributions of the proposed architecture are residual connections to optimize the learning process, attention gates to efficiently extract the important features, ASPP to capture multi-scale spatial information, and initialization with pre-trained weights of VGG19 to robustly increase the generalization ability for accurate skin lesion segmentation from dermoscopic images. We evaluate the performance of the proposed architectures on publicly available data ISIC 2018. The quantitative and qualitative analysis exhibit that the DU-Net+ architecture outshone the baseline architecture and other state-of-the-art architectures with a dice coefficient of 0.9293, Jaccard of 0.8813, accuracy of 0.9741, sensitivity of 0.9512, and specificity of 0.9676.