Few-Shot Classification with Multiscale Feature Fusion for Clinical Skin Disease Diagnosis

Skin disease is one of the most common diseases and can affect people of all ages and races. However, the diagnosis of skin diseases via observation is a highly challenging task for both doctors and patients, and would benefit from the use of an intelligent system. Building a large benchmark with professional dermatologists is resource-intensive, and we believe that few-shot learning (FSL) methods would be helpful in solving the problem of annotated data scarcity. In this paper, we propose CDD-Net (Context Feature Fusion and Dual Attention Dermatology Net), a plug-in module for FSL clinical skin disease classification.

Current FSL methods used in skin disease classification are limited to nonuniversal approaches and few disease classes. Our CDD-Net has a flexible structure, including a context feature-fusion module and dual-attention module to extract discriminating texture feature and emphasize contributive regions and channels. The context feature-fusion module localizes discriminatory texture details of skin lesions by integrating features from different layers, while the dual-attention module highlights discriminative regions via channel-wise and pixel-wise depictions based on weight vectors and restrains the contributions of irrelevant areas. We also present Derm104, a new clinical skin disease data benchmark that has significant coverage of rare diseases and reliable annotation between primary species and subspecies for better validation of our approach.

Our experiments validated the versatility of CDD-Net for different FSL methods and achieved an improvement in accuracy of up to 9.14 percentage points compared with the vanilla network, which can be considered state of the art. The ablation study also showed that the dual-attention module and context feature-fusion module worked efficiently in CDD-Net.