Classification for Dermoscopy Images Using Convolutional Neural Networks Based on Region Average Pooling

In this paper, a novel melanoma classification method based on convolutional neural networks is proposed for dermoscopy images. First a region average pooling (RAPooling) method is introduced which makes feature extraction can focus on the region of interest. Then an end-to-end classification framework combining with segmentation information is designed, which uses the segmented lesion region to guide the classification by RAPooling. Finally, a linear classifier RankOpt based on the area under the ROC curve is used to optimize and obtain the final classification result. The proposed method integrates segmentation information into the classification task, and in addition, by the optimization of RankOpt, a better classification performance for imbalanced dermoscopy image dataset is obtained. Experiments are conducted on ISBI 2017 skin lesion analysis towards melanoma detection challenge dataset, and comparisons with the other state-of-the-art methods demonstrate the effectiveness of our method.