Segmentação semântica de imagens dermatoscópicas de lesões de pele utilizando aprendizado profundo

Abstract

Skin cancer is a disease characterized by the formation of malignant cells from the melanocytes, which are cells that give color to the skin. Despite being the most recurrent type of all, skin cancer is extremely treatable in the early stages of the disease. In dermoscopy, there is no tool to help physicians to make an early diagnosis of the disease. Thus, this work proposes a tool that uses machine learning techniques (semantic segmentation) to optimize the work of dermatoscopists. The validation of the tool showed an average result of 0.0971 in the loss dice for the training set and 0.1724 for the validation set, allowing a more accurate analysis of the evolution of skin lesions according to time. Although the tool has good accuracy from the point of view of machine learning techniques, adjustments will still be needed to meet the expectations of an application in the medical field.