Abstract: With the rapidly increasing incidence of various types of skin cancer, there is a need for decision support systems to detect abnormalities in the early stages and help reduce the mortality rate. Several computer-aided diagnosis (CAD) systems have been proposed in the last two decades for skin melanoma recognition. Continuous improvements have been made in the accuracy of melanoma diagnosis, but other classes of cancer, such as basal cell carcinoma and squamous cell carcinoma, are not very intact with the non-invasive diagnosis system. In this paper, a generic method of diagnostic system is proposed and is viable to classify the ten classes of a skin lesion. These lesion classes belong to cancer, pre-cancerous, and tumor categories of samples, as shown in a gold standard image dataset. The key idea of the proposed approach is to optimize the bag-of-SURF features by the non-linear Hessian matrix of HSV color descriptors. These features are combined to form a salient bag-of-features, which helps recognize the skin lesion classes more accurately. Experimental results show that the proposed method of skin lesion diagnosis significantly improves the accuracy of recognition up to 89% as compared to the current state-of-the-art accuracy of 81.8%. It does not require any complex pre-processing of images, which affects the performance of the recognition system.

Key words: computer-aided diagnosis (CAD), bag of visual words (BoVW), speeded-up robust features (SURF), hue saturation value (HSV), bag of color (BOC), bag of robust features (BORF), support vector machine (SVM)