Detección de melanomas a partir de imágenes dermatoscópicas

Abstract

The possibility of surviving melanoma skin cancer is greater if it is detected and diagnosed early. Therefore, this project presents the development of a tool based on image processing, capable of classifying lesions from dermoscopic images. The project was divided into three stages: segmentation, feature extraction and classification. At segmentation stage the Otsu and semantic segmentation methods were implemented with morphological operators as a post-processing stage to enhance the results. From the segmented images, a histogram of 250 visual words was obtained from the most representative images of each class using descriptors of color and shape. Finally, an SVM classifier with a maximum precision of 78% was obtained using the HSV color space and a Gaussian kernel.