AttR2U-Net: Deep Attention Based Approach for Melanoma Skin Cancer Image Segmentation

Unless detected and treated in its earliest stages, melanoma skin cancer can grow deeply into nearby skin tissues letting patients with serious and complex conditions. Therefore, a considerable effort has been made to improve the process of melanoma detection and segmentation from the surrounding skin. However, even recent deep learning models that have been used for automatic melanoma segmentation still suffer from many challenges, including the large variations in shape, size, color of skin lesions. Also, these models are usually fed with images, in which lesions occupy only small regions, without incorporating any mechanisms allowing the network to differentiate between important parts of the image from parts that are not. In an effort to address these challenges, we propose AttR2U-Net, a neural net architecture, based on the R2U-Net architecture with an additional attention mechanism. This mechanism allows the network to be guided to concentrate on specific parts of the image for efficient lesion segmentation. In this work, we conducted an experimental study on the ISIC-skin 2018 dataset to show how and where to integrate such mechanism within the R2U-Net architecture. Our experimental results show a noticeable performance improvement.