MaSA-UNet: Manhattan Self-Attention U-Net for Skin Lesion Segmentation

Abstract Melanoma is one of the deadliest forms of skin cancer in the USA, with a survival rate of 23% for delayed diagnosis. However, early detection could extend the survival rate up to 99%. Due to the importance of early analysis of skin lesions, many efforts have been dedicated to implementing end-to-end automated artificial intelligence systems to detect the presence of melanoma. In this study, we introduce the MaSA-UNet, a U-Net-like architecture complemented by the Manhattan Self-Attention mechanism for biomedical image segmentation. Additionally, we propose a set of weighted compound loss (WCL) functions and self-supervision learning mechanisms to improve the segmentation baseline performance of the MaSA-UNet deep learning model for skin lesion segmentation tasks, particularly focusing on melanoma detection. This research utilizes several popular and publicly available skin lesion datasets: the ISIC 2016, 2017, 2018, and PH\((^2)\) datasets. The results show that our proposed MaSA-UNet outperforms the state-of-the-art deep learning architectures for skin lesions segmentation tasks in terms of the Dice coefficient and Jaccard score. The source code will be available in a GitHub repository (available upon publication).