Automatic melanoma diagnosis framework based on common image feature learning

Abstract

Automatic melanoma diagnosis based on image processing can give more objective results. To facilitate examination for patients at home, we propose a new automatic melanoma diagnosis framework based on common images. Firstly, we use illumination assessment based on variational framework for Retinex (VFR) to filter the images with illumination problem caused by variation of capturing cameras viewpoint and ambient light. Secondly, the GrabCut algorithm based on colour difference is used to segment lesion area. It can complete segmentation automatically and efficiently. Thirdly, we use convolutional neural network (CNN) to extract high-level features and choose support vector machine (SVM) classifier to complete melanoma classification. Compared to hand-craft features, CNN can acquire deep information of images. Because of the lack of medical images, the SVM classifier is better than other classifiers. Finally, we validated our approach from different perspectives and the accuracy is increased by about 5% over other methods.