Deep metric attention learning for skin lesion classification in dermoscopy images

Abstract Currently, convolutional neural networks (CNNs) have made remarkable achievements in skin lesion classification because of their end-to-end feature representation abilities. However, precise is still challenging the following three issues: (1) insufficient training samples, (2) inter-class similarities and intra-class variations, (3) lack ability to focus on discriminative parts. To address these issues, we propose a deep metric attention learning CNN (DeMAL-CNN) for classification. In DeMAL-CNN, triplet-based network (TPN) first designed based learning, which consists weight-shared embedding extraction networks. TPN adopts triplet samples as input uses loss optimize embeddings, can not only increase number but also learn embeddings robust variations. addition, mixed mechanism considering both spatial-wise channel-wise information integrated into construction each network, further strengthen localization DeMAL-CNN. After extracting layers are used generate final predictions. procedure, combine with hybrid train We compare DeMAL-CNN baseline method, methods, advanced challenge state-of-the-art methods ISIC 2016 2017 datasets, test its generalization PH2 dataset. The results demonstrate effectiveness.