RADU-Net: A Fully Convolutional Neural Network for Efficient Skin Lesion Segmentation

Automatic skin lesion segmentation in dermoscopic images is a challenging task due to the presence of artifacts, fuzzy boundaries, variability in skin lesion appearance, low contrast between lesion skin and normal skin, and limited training data. In this paper, we introduced a robust, fully convolutional neural network for efficient and effective segmentation of skin lesions. Our proposed model, the Residual Attention Double U-Net (RADU-Net), is an encoder-decoder architecture that connects two U-Nets in a cascade manner. It comprises repeated convolutional and deconvolutional layers, connected with skip connections through the attention gates, and incorporates residual learning within convolution layers. Double U-Net captures fine details while persevering contextual information. The attention gates enable the model to preserve essential spatial information while ignoring irrelevant feature representations, enabling the model to focus on target regions. Additionally, the inclusion of residual connections enhances the training of a deeper network by ensuring a smoother gradient glow. We assessed the performance of the proposed model on the publicly available ISIC 2018 dataset, achieving a dice of 0.9074 and Jaccard of 0.8589. The proposed model consistently outperforms the baseline models for skin lesion segmentation.