文章基本信息

标题：A Morphological Convolutional Autoencoder for Segmenting Pigmented Skin Lesions
本地全文：下载
作者：Alberto M. Ceballos-Arroyo ; Sergio Robles-Serrano ; German Sanchez-Torres 等
期刊名称：Engineering Letters
印刷版ISSN：1816-093X
电子版ISSN：1816-0948
出版年度：2020
卷号：28
期号：3
页码：855-866
语种：English
出版社：Newswood Ltd
摘要：Pigmented skin lesions can be benign or malignant,such as skin cancer. Malignant melanoma is the most dangerouskind of skin cancer, and it causes 75% of related deaths. Earlydiagnosis can result in preserving the lives of most patients,but in most countries, the analysis is done based on a manualinspection by specialists, which can be inaccurate. Digitaldermatoscopy is a non-invasive methodology that allows in-vivo evaluation of different skin conditions at the macroscopiclevel using histological features, and it can be automatised bymeans of computational tools. In this paper, we propose a U-Net-based architecture including morphological layers, calledMorpho-U-Net, for the automatic segmentation of skin lesions.The output of this architecture consists of binary masks that canbe used to separate the lesions from the rest of the dermoscopyimage and can serve as input for either human-based analysisor other algorithms for skin lesion classification. Our strategywas tested on the ISIC 2017, ISIC 2018, and NH2 data-sets.Our experiments showed that our work is above several state-of-the-art proposals, with an average thresholded Jaccard scoreof 0.93. Furthermore, we believe our architecture could be usedas the basis for addressing other image segmentation problems.
关键词：Melanoma; visual computing; machine learn- ing; auto-encoders; deep learning.