文章基本信息

标题：EMPIRICAL STUDY AND ENHANCEMENT ON DEEP TRANSFER LEARNING FOR SKIN LESIONS DETECTION
本地全文：下载
作者：NOUR ELDEEN M. KHALIFA ; MOHAMED LOEY ; AHMED A. MAWGOUD 等
期刊名称：Journal of Theoretical and Applied Information Technology
印刷版ISSN：1992-8645
电子版ISSN：1817-3195
出版年度：2020
卷号：98
期号：9
页码：1351-1361
出版社：Journal of Theoretical and Applied
摘要：Skin cancer is the most common type of cancer. One in every three cancers diagnosed is a skin cancer according to skin cancer foundation statistics globally. The early detection of this type of cancer would help in raising the opportunities of curing it. The advances in computer algorithms such as deep learning would help doctors to detect and diagnose skin cancer automatically in early stages. This paper introduces an empirical study and enhancement on deep transfer learning for skin lesions detection. The study selects different pre-trained deep convolutional neural network models such as resnet18, squeezenet, google net, vgg16, and vgg19 to be applied into two different datasets. The datasets are MODE-NODE and ISIC skin lesion datasets. Data augmentation techniques have been adopted in this study to enlarge the total number of images in the datasets to be 5 times larger than the original datasets. The adopted augmentation techniques make the DCNN models more robust and prevent overfitting. Moreover, seven accredited performance matrices in deep learning have been used to conclude an optimal selection of the most appropriate DCNN model that fits the nature of the skin lesions datasets. The study concludes that vgg19 is the most appropriate DCNN according to testing accuracy measurement and achieved 98.8%. The seven performance matrices strengthen this result. Also, a comparative result was introduced with related works. The vgg19 overcomes the related work in terms of testing accuracy and the performance matrices on both datasets. Finally, the vgg19 model was trained on a smaller number of images than the related work by 10 times, which proved that the choice of data aug-mentation techniques played an important role in achieving better results. That would reflect on reducing the training time, memory consumption and the calculation complexity.
关键词：Cancer;Skin Cancer;Melanoma;Deep Transfer Learning;Convolutional Neural Network.