Automated Skin Lesion Segmentation in Medical Images Using U-Net Architectures

Le résumé fourni par la source

The biggest challenge, modern methods for laser and surgical treatments face in context of skin diseases is to find the exact boundaries of skin lesions. However, with the integration of deep learning applications with these treatments have shown a great improvement in finding the lesions boundaries. The aim of this study is to comparatively investigate the performances of U-Net and its three improved variations, Residual U-Net, Attention U-Net and Residual Attention U-Net, in skin lesion segmentation. The models were tested on two widely available public datasets, namely ISIC-2016 and ISIC-2017, and the comparison was performed using the same training parameters, image dimensions and evaluation metrics namely accuracy, Dice score and IoU. Attention U-Net model achieved the highest success on ISIC-2016 dataset with 94.4% accuracy, 81.9% Dice score and 81.5% IoU. On the ISIC-2017 dataset, the Residual Attention U-Net model showed superior performance with 92.2% accuracy, 76.9% Dice score and 69.5% IoU. The results show that attention mechanisms and residual structures provide significant contributions to the accurate segmentation of skin lesions and that these architectures have the potential to be used in clinical decision support systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.