Melanoma Classification Using Deep Convolutional Neural Networks with Ensemble Scheme

Abstract

Among many skin diseases, melanoma is the most common and deadly malignant skin cancer which seriously threatens people’s physical health. An effective way to treat melanoma is to use dermoscopic images for early diagnosis. With the limited highly-trained experts, deep learning can be an alternative for melanoma classification. In this paper, we adopt the convolutional neural networks (CNNs) model, i.e., EfficientNet [1], along with ensemble schemes to solve the melanoma classification problem. The EfficientNet model can design the proper network architecture to extract features with better accuracy and efficiency than other CNNs. In order to solve the problem of data imbalance and try multiple parameter designs, we construct six different basic EfficientNet model and integrate their predictions. We conduct experiments on dataset released by a Kaggle competition. The experiments results show that our approach achieves 0.950 AUC score on the validation set, outperforming VGG16 and VGG19 [2] by 0.059 and 0.028 respectively, which verify the superiority of our CNN model and the improvement induced by the model ensemble strategy.