Abstract
Early detection of skin cancer is vital when treatment is most likely to be successful. However, diagnosis of skin lesions is a very challenging task due to the similarities between lesions in terms of appearance, location, color, and size. We present a deep learning method for skin lesion classification by fusing and fine-tuning three pre-trained deep learning architectures (Xception, Inception-ResNet-V2, and NasNetLarge) using training images provided by ISIC2019 organizers. Additionally, the outliers and the heavy class imbalance are addressed to further enhance the classification of the lesion. The experimental results show that the proposed framework obtained promising results that are comparable with the ISIC2019 challenge leader board.