Ensemble of Deep Convolutional Neural Network for Skin Lesion Classification in Dermoscopy Images

Abstract

Diagnosis of skin lesions is a very challenging task due to high inter-class similarities and intra-class variations between lesions in terms of color, size, site and appearance. As a result, as is the case with many types of cancer, early detection of skin cancer is vital for survival. Advances in artificial intelligence, in particular, deep learning have enabled to design and implementation of intelligence-based lesion detection and classification systems that are based on visible light images that are capable of performing early and accurate diagnosis of different skin diseases. In most cases, the precision of these methods has reached a level of accuracy that is comparable to that achieved by a qualified dermatologist. This work presents potential skin lesion classification solutions based on the datasets taken from the most recent publicly available “Skin Lesion Analysis Towards Melanoma Detection” grand challenges ISIC 2018. The proposed classification approach leverages convolutional neural networks (CNN) and transfer learning to enhance skin classification. Different pre-trained models were applied, including VGG-Net, ResNet50, InceptionV3, Xception and DenseNet121. Additionally, the heavy class imbalance is examined as a critical problem for this dataset and multiple balancing techniques, such as weight balancing and data augmentation, are considered. Finally, an ensemble approach is evaluated by combining and averaging several CNN architectures to classify the seven different types of skin lesion. The experimental results indicate that the proposed frameworks exhibit promising results when compared with ISIC 2018 challenge live leaderboard.