MSFE-CapsNet: a Multiscale Feature-enhanced Capsule Network for Skin Lesion Image Classification

Automatic classification of skin lesion images is an important method for learning lesion characteristics from dermoscopic images and determining the category to which they belong, which is crucial for the diagnosis and treatment of skin cancer. However, the large variation in lesion size seriously affects the classification of the images. Therefore, in this paper, we propose a Multiscale feature-enhanced capsule network (MSFE-CapsNet), which employs a large size 31 × 31 convolutional kernel to obtain a larger perceptual domain, and design a multiscale feature enhancement block to augment the local feature information to further learn the semantic information of the lesion image, and finally introduces the Efficient Multi-Scale Attention to efficiently obtain spatial location information of skin lesions on a high-level feature map. The experimental results show that MSFE-CapsNet achieves 95.05% accuracy on the HAM10000 dataset, which is better than the existing methods for skin lesion classification, and has only 1.58M parameters.