A Multi-Input CNNs with Attention for Skin Lesion Classification

Abstract

Melanoma skin cancer is one of the fastest growing and deadliest cancers in the world. Therefore, the classification of melanoma in dermoscopy images is of great significance for computer-aided diagnosis. Although with the application of convolutional neural networks, skin lesion classification tasks have achieved great breakthroughs, due to similarity between classes, differences within classes, as well as the class imbalance of the dataset, the class imbalance of the foreground and background of the image, the classification of melanoma in dermoscopy images is still challenging. To solve these problems, we propose a deep convolutional neural network for dermoscopy image classification based on multi-input and attention mechanism. We evaluated our model on the HAM10000 dataset, and our results show that our model is more capable of achieving advanced performance.