Capsule neural networks in classification of skin lesions

Abstract

Classification of skin lesions is a difficult issue even for highly experienced dermatologists and pathologists because of several reasons such as low contrast between lesions and surrounding skin tissue, high noise, fuddled lesion boundaries, visual similarities of skin lesions. On the other hand, early and accurate classification of lesions is crucial for timely and accurate treatment of skin diseases. Therefore, automated methods have been developed to perform objective, quantitative and re-producible results. Recent methods in pattern recognition and image classification is based on deep networks, particularly convolutional neural networks. However, pooling layers providing down-sampling in these networks lead to data loss and cause low performance in generalization. Also, convolutional neural networks cannot transfer spatial information and instantiation parameters (e.g., pose of low-level features to each other, deformation and texture information). To overcome these problems with dynamic routing, capsule neural networks have been proposed. Capsules can transfer pose parameters and part-whole relationship using likelihood and spatial information between low-level features. In this work, capsule networks applied for skin lesion classification have been explored and their performances have been evaluated. It has been observed that although capsule networks can overcome deficiencies of convolutional neural networks, there are only four techniques based on capsule networks in the literature to achieve automated classifications of skin lesions. Motivated by the lack of articles on this topic, a comprehensive assessment of the capsule networks proposed for skin lesion classification is presented in this paper. Also, strengths and weakness of those four techniques have been presented to help the researchers interested in this area.