Classification of medical images of skin lesions using capsular neural network

Deep networks are a type of learning method that can model high-level relationships in data. One of the most widely used types of deep models are convolutional networks that are able to model spatial dependencies in images using convolutional layers, but do not consider the hierarchical spatial structures within the image. Capsule networks are one of the new ideas proposed for modeling the hierarchical structure of features in the image, which use grouped capsules or neurons with a dynamic routing algorithm. Despite the effectiveness of the idea of ​​capsule networks on simple data sets, the performance of these networks on complex data is still unclear. In this paper, the performance of this network is examined on a complex skin cancer dataset, which has been selected due to the importance of skin lesions diagnosis in medicine, the complexity and huge number of images and the imbalance of categories. In order to better extract the diversity of skin lesions, changes were made in the initial layers of the network. Also, due to the imbalance in the mentioned data set, changes were made in the cost function of the network. The effect of using different activation functions in the network was also investigated. The results show that the idea of ​​a capsule network can be used optimally on complex data sets by making appropriate adjustments