A Compact Deep Ensemble for High Quality Skin Lesion Classification

Convolutional Neural Networks (CNNs) are widely employed in the medical imaging field. In dermoscopic image analysis, the large amount of data provided by the International Skin Imaging Collaboration (ISIC) encouraged the development of several machine learning solutions to the skin lesion images classification problem. This paper introduces an ensemble of image-only based and image-and-metadata based CNN architectures to classify skin lesions as melanoma or non-melanoma. In order to achieve this goal, we analyzed how models performance are affected by the amount of available data, image resolution, data augmentation pipeline, metadata importance and target choice. The proposed solution achieved an AUC score of 0.9477 on the official ISIC2020 test set. All the experiments were performed employing the ECVL and EDDL libraries, developed within the european DeepHealth project.