Novel Ensembling Methods for Dermatological Image Classification

In this paper we investigate multiple novel techniques of ensembling deep neural networks with different hyperparameters and differently preprocessed data for skin lesion classification. To this end, we have utilized the datasets made public by two of the most recent “Skin Lesion Analysis Towards Melanoma Detection” grand challenges (ISIC2017 and ISIC2018). The datasets provided by these two challenges differ in multiple aspects: the size, quality and origin of the images, the number of possible target lesion categories and the metrics used for ranking. We will show that ensembling can be surprisingly useful not only for combining different machine learning models but also for combining different hyperparameter choices of these models and multiple strategies for preprocessing the input data at the task of skin lesion detection, outperforming more mainstream methods like hyperparameter optimization and test-time augmentation both in terms of speed and accuracy.