Data Enhancement for Melanoma Classification

Melanoma Classification is a popular question in computer vision, with numerous works are proposed. Recent works focus on Convolutional Neural Networks (CNN) in this task. However, the question of insufficient or unbalanced data has not been paid attention to. In this paper, we aim to solve this problem and improve the accuracy of the model. We first compare the effect of the data-enhancement and sampling method in the modeling and then use two types of CNN to construct the model. Finally, we use ensemble learning to try to balance and improve the prediction. Experiments on data enhancement show our method is capable of handling the unbalance dataset with 97.7% accuracy. Finally, our method has a great improvement on the task of Melanoma Classification, which is from 79.9% to 97.8% using unbias training data. Our model decreases the influence of data imbalance and achieves a satisfying performance without label bias.