Melanoma Recognition by Fusing Convolutional Blocks and Dynamic Routing between Capsules

SIMPLE SUMMARY: The early treatment of skin cancer can effectively reduce mortality rates. Recently, automatic melanoma diagnosis from skin images has gained attention, which was mainly encouraged by the well-known challenge developed by the International Skin Imaging Collaboration project. The majority of contestant submitted Convolutional Neural Network based solutions. However, this type of model presents disadvantages. As a consequence, Dynamic Routing between Capsules has been proposed to overcome such limitations. The aim of our proposal was to assess the advantages of combining both architectures. An extensive experimental study showed the proposal significantly outperformed state-of-the-art models, achieving 166% higher predictive performance compared to ResNet in non-dermoscopic images. In addition, the pixels activated during prediction were shown, which allows to assess the rationale to give such a conclusion. Finally, more research should be conducted in order to demonstrate the potential of this neural network architecture in other areas. ABSTRACT: Skin cancer is one of the most common types of cancers in the world, with melanoma being the most lethal form. Automatic melanoma diagnosis from skin images has recently gained attention within the machine learning community, due to the complexity involved. In the past few years, convolutional neural network models have been commonly used to approach this issue. This type of model, however, presents disadvantages that sometimes hamper its application in real-world situations, e.g., the construction of transformation-invariant models and their inability to consider spatial hierarchies between entities within an image. Recently, Dynamic Routing between Capsules architecture (CapsNet) has been proposed to overcome such limitations. This work is aimed at proposing a new architecture which combines convolutional blocks with a customized CapsNet architecture, allowing for the extraction of richer abstract features. This architecture uses high-quality [Formula: see text] skin lesion images, and a hyper-tuning of the main parameters is performed in order to ensure effective learning under limited training data. An extensive experimental study on eleven image datasets was conducted where the proposal significantly outperformed several state-of-the-art models. Finally, predictions made by the model were validated through the application of two modern model-agnostic interpretation tools.