Técnicas preditivas para auxílio no diagnóstico de melanomas via imagens

Abstract

Skin cancer is the most common type of cancer worldwide, divided into two main types: melanoma and non-melanoma. Although rarer, melanoma is the most lethal due to its potential to cause metastasis. Non-invasive methods, such as dermoscopy and the ABCDE rule, have been used to avoid unnecessary surgical procedures and have helped in the identification of lesions, contributing to faster diagnoses. With advances in technology, Artificial Intelligence (AI) has gained prominence, proving to be a promising solution for medical data analysis, especially with the use of Convolutional Neural Networks (CNNs), which can recognize patterns in dermoscopic images and help classify lesions as melanoma or non-melanoma in an automated manner. This project proposes an ensemble of classifiers based on Convolutional Neural Networks to classify dermoscopic images as melanoma or non-melanoma, comparing its performance with validated architectures, such as AlexNet and VGG-16, using Transfer Learning techniques The analyses of Precision, Recall, and F1 Score showed that the ensemble of Convolutional Neural Networks outperformed the models using Transfer Learning techniques, with AlexNet showing better performance than VGG-16. The ensemble of Convolutional Neural Networks demonstrated a greater generalization capability, proving to be promising in capturing relevant features from the images, revealing potential for medical applications, although it still needs refinement to meet clinical standards.