SkinXNet: A DoG-based Model for Automatic Detection of Skin Lesion using Deep Learning

Regarding the poor contrast of skin lesions, the enormous intraclass heterogeneity of melanomas, the high rate of a visual resemblance within non-Melanoma and Melanoma lesions, and the appearance of several objects in the images, automatic melanoma classification based on dermoscopy images is a very challenging task. To meet these challenges, we propose a novel approach for Melanoma recognition by leveraging the Difference of Gaussian (DoG) feature enhancement technique with a combination of ResNet structure for feature extraction and classification. First, features were enhanced by DoG edge detection technique. Then, processed images went through the proposed deep residual architecture for selecting the prominent features and classification of Melanoma from Melanocytic nevus. The proposed approach was validated on the well-known ISIC (2019) dataset. In comparison to the previous studies, the proposed framework illustrated a superiority in terms of sensitivity on the Melanoma identification. The experimental results demonstrated that our proposed SkinXNet framework can reliably classify skin lesions with a sensitivity of 0.93, AUC of 0.815, and accuracy of 81%. Obtained results lending support to the expectation that the introduced improved hybrid approach could be useful for providing a second opinion for medical experts.