Classifying Melanocytic Nevus by Using Extracted Depth-Encoded Channel Information

Melanocytic nevi are common benign skin lesions. However, some types of nevi have a higher risk of malignant transformation into melanoma, which is the most aggressive form of skin cancer. To avoid unnecessary biopsies and help reduce medical costs, accurate fine-grained nevi classification methods are important. Different from skin lesion classification, melanocytic nevi classification is more challenging due to the inter-class similarity. Inspired by the physical principles underlying dermoscopy imaging and pathological basis of nevi, we propose a novel classification network to extract depth-relevant channel-specific information and channel-shared information for dermoscopy images. A reconstruction module with a differential loss is used to exploit and enlarge the representations of different channels. Our algorithm is evaluated on a three-class melanocytic nevi dataset with ground truth verified by histological examination. The accuracy of the proposed method is 3.69% higher than state-of-the-art skin lesion classification technique.