A Comparative Study of CNN Architectures for Melanoma Skin Cancer Classification

Abstract

Skin cancer is recognized as one of the common and deadliest cancer. Therefore, an easy, early and discriminatory diagnosis of the skin cancer would be crucial to ensure appropriate and e ective treatment for patients. Although there are many computerized methods for classifying skin lesions, Convolutional Neural Networks (CNN) have proven superior to the conventional methods. In the present work, we investigated seven deep neural network classi- ers, namely, baseline CNN, InceptionV3, ResNet50, VGG16, Xception, MobileNetV2 and DenseNet201 to predict the input skin lesion class (melanoma or nevus). The adapted architectures were tested on 1500 ISIC images and the comparisons of the obtained results with the E - cientNetB3 as well as InceptionResNetV2 have demontrated the superiority of the Baseline CNN, DenseNet201 and Xception. These models would provide a strong automated support for the melanoma diagnosis and could be exploited in the classi cation of eczema and psoriasis skin lesions.