Hierarchical Clustering for Image Classification in Dermatology: Towards Mobile Deploying

In the context of increasing interest in computer-assisted diagnosis for skin lesion images and mobile applications to be used in real life settings, we propose a combined desktop-smartphone solution for dermatological image classification. Hierarchical agglomerative and divisive clustering are both implemented as methods of cluster analysis, with the RGB color histogram as descriptor for a global image analysis. The cosine similarity is employed for classifying the query image in one of the available clusters, characterized by their centroids. The solution has been tested with a public database of dermoscopic images, with an overall accuracy of 0.73, 95%CI (0.58;0.85).