Appending global to local features for skin lesion classification on dermoscpic images

Skin cancer is considered the deadliest disease compared with all other kinds of cancer. In this paper, various pre- and posttreatments are proposed for improving automated melanoma diagnosis of dermoscopy images. At first, preprocessing has done to exclude unwanted parts, and a new triple-A segmentation is proposed to extract lesions according to their histogram patterns. Lastly, we suggest appending process with testing many factors for superior detection decision. This paper offers a novel approach with testing different detection rules: first, a system used fuzzy rules based on a different features, and a second test has been done by modeled local colors with bag-of-features classifier. Then, we proposed adding lesion shape on two previous systems as their global form in the first one while distributing it and appending with local color patches in the second system. For each case, different features, various color models, and many other parameters are examined to decide which settings are more discriminating. We evaluated the performance of each method carried out on ISIC2019 Challenge dermoscopic database. The novel processes with their specific parameters raised the classification accuracy to 98.26%.