<p>Automatic skin-cancer detection is an area in computer vision in which convolutional neural networks (CNNs) have shown remarkable performance since they have proved to be very efficient image feature extractors. Nevertheless, developing an efficient skin-cancer detection model in real world scenarios is a challenging task, since skin-cancer image datasets are characterized by instances with a large amount of noise and redundant information. Therefore, it is essential to develop an efficient augmentation method in order to create robust and diverse image representations. For this task, we propose a feature augmentation-based method, which enlarges the dimension of skin-cancer image representations. In this approach different types of augmentation functions mainly composed of noise-based injection functions, such as hair and microscope effect, are injected into the initial representation. Subsequently, they are used to extract different types of features based on pre-trained CNN backbones such as ResNet and DenseNet. All these different types of feature embeddings are concatenated together creating an augmented multi-feature representation of the initial image instance. However, in order to remove redundant information and reduce the large multi-dimension of the CNNs’ output feature embeddings, we incorporate U-Map dimensional reduction method. Finally, the compressed feature embedding is used for training an XGBoost model in order to perform the final classification task. The main findings of this work are that the proposed methodology achieved an accuracy of 0.929 (via 10-fold cross-validation), significantly improving CNN performance in skin-cancer diagnosis and outperforming state-of-the-art image classification and skin-cancer detection CNN frameworks.</p>