Representação baseada em superpixels para a segmentação de lesões melanocíticas em imagens macroscópicas

Abstract

Melanoma is a type of skin cancer with high mortality rates due to its ability to form metastasis. It is important that the diagnosis be made in the early stages of the disease. Computational systems have been proposed by researchers in the area of computer vision as an alternative capable of assisting dermatologists in their diagnoses. These systems generally depend on a proper segmentation of the region that delimits the skin lesion. In this work, an unsupervised method for segmenting melanocytic lesions in macroscopic images is presented. Initially, to reduce shaded areas and undesirable artifacts, shadow effects are attenuated in the macroscopic image and a pre-segmentation is obtained using a superpixel-based representation. Each superpixel is then represented by a features set using information from the R, G and B channels of the RGB color space. In the segmentation step, the normalized graph cut method and SVM classifier are used to separate the set of superpixels into two classes, lesion and non-lesion. Finally, post-processing methods based on mathematical morphology are applied in order to improve the obtained segmentation. The results based on a widely used image base demonstrate that the proposed methodology provides satisfactory results in terms of accuracy, specificity and XOR error.