Combinando algoritmos de agrupamento e superpixels para a segmentação de lesões melanocíticas em imagens macroscópicas

Abstract

Melanoma is a skin cancer that can be extremely aggressive in its final stages, and in some cases, it can develop metastasis. Due to these complications, the diagnosis of suspicious melanocytic lesions should be performed as soon as possible. Therefore, proposals to improve the use of computational systems, in this context, have been widely investigated. In the present work, two variations, one supervised and the other unsupervised, of a method for the segmentation of macroscopic images of melanocytic lesions were investigated. Before the segmentation itself, the input image is pre-processed by an algorithm that performs brightness regularization in order to mitigate the effects arising from irregular lighting. Then, the pre-processed image is represented using superpixels. For each of these superpixels, features of the R, G and B channels are extracted. Subsequently, a clustering algorithm is fed with the extracted features and, based on that, partitions the superpixels into two groups, which are expected to predominantly correspond to the regions of lesion (foreground - object of interest) and non- lesion (background - background). Finally, some basic post-processing steps are performed in order to refine and standardize the segmentation results, so that the pixels identified as lesions are represented by white and the others by black. For the preliminary evaluation of the proposed methods, the results produced by several variations of the method were used to compare the variants in terms of accuracy, sensitivity, specificity and XOR errors.