Identificação de lesões cutâneas suspeitas de câncer de pele do tipo não melanoma em imagens adquiridas por smartphones

Abstract

Non-melanoma skin cancer is the most common type of cancer in Brazil. Although more frequent, this type of cancer often goes unreported, making it difficult to diagnose the disease early. Several studies on the use of artificial intelligence and neural networks to classify skin lesions, including cancer, have been and are being carried out, mainly using public databases such as the International Skin Imaging Collaboration, an image bank of dermatoscopic examinations. With a focus on mobile applications, there are several software programs on the market that classify skin lesions. In terms of academic work, the PAD-UFES-20, used in this study, a database of images captured using smartphones, has been created and is the subject of studies on the classification of skin lesions. However, no specific studies dealing exclusively with non-melanoma skin cancer were found using this dataset. In this context, the aim of this work is to identify non-melanoma cancer lesions in images from smartphone cameras. To this end, reviews were carried out on skin cancer and its diagnosis, artificial intelligence, neural networks and aspects of their training. In the practical stage, three neural networks were trained using data augmentation techniques, fine tuning and the PAD-UFES-20 database, divided into three different scenarios and their results were compared with similar studies. An application programming interface was developed, split between a client and a server, which can support a neural network module. Finally, the application was tested with real images and images generated by augmenting the PAD-UFES-20 database. As a result, better results were obtained in the scenarios in which training was carried out with a greater amount of data, with a maximum F1 score of 99.3.