Skin cancer is a serious global health concern with high morbidity and mortality rates. Therefore, computer-aided automatic diagnostic systems are essential to diagnose skin lesions successfully. One of the most important ways to achieve that is the segmentation of skin lesions. This study compared different criteria used to improve the performance of lesion segmentation in skin images via deep Convolutional Neural Network (CNN) architectures. In addition, for the segmentation of skin lesions, a new hybrid FCN (Fully Convolutional Network)-based deep learning architecture is proposed, in which AlexNet and ResNet18 deep learning architectures are used in the encoder structure, and three deconvolution layers are included in the decoder part. In the study, ISIC2017 and ISIC2018 data sets, the most used in the segmentation of skin lesions, were used. Results show that the proposed architecture, FCN-ResAlexNet, trained with the ISIC2017 data set, provided 0.35%, 2.73%, and 4.2% higher performance than the FCN-AlexNet architecture in accuracy, dice, and Jaccard index performance metrics, respectively. Also, The FCN-ResAlexNet architecture trained with the ISIC2018 dataset provided 1.79%, 0.95%, and 1.5% higher performance than the FCN-AlexNet architecture in accuracy, dice, and Jaccard index performance metrics, respectively.
Cite this article
S. Barın, G. E. Güraksın, An automatic skin lesion segmentation system with hybrid FCN-ResAlexNet, Engineering Science and Technology, an International Journal, Volume 34, 2022, 101174, ISSN 2215-0986, https://doi.org/10.1016/j.jestch.2022.101174.