Comparativo de modelos de redes neurais convolucionais no diagnóstico e classificação de lesões de câncer de pele através de imagens dermatoscópicas

Abstract

The fundamental purpose of this work is to develop and compare the outcomes of CNN techniques in the classification and recognition of melanoma, a type of skin cancer, through dermoscopic images. Skin cancer is a serious health condition where early diagnosis is crucial for successful treatment. In traditional techniques, the diagnosis relies on visual observation by medical professionals, which may have limitations. To overcome these limitations, the use of deep learning techniques with convolutional neural networks to create an efficient computational model for automated classification of melanoma in dermoscopic images is an alternative.The adopted methodology involves collecting a dataset (HAM10000) and applying data preprocessing and augmentation techniques to optimize the quantity and quality of data available for each class in the model training. The implementation is carried out using five CNN models with specific adjustments for the classification of dermoscopic images. Model evaluation occurs through performance metrics such as accuracy, confusion matrix, and Friedman and Nemenyi tests to ensure efficiency and compare the feasibility of the developed models.As a result, significant differences were found between the compared models, with emphasis on the MobileNet and InceptionV3 models, which showed statistically significant differences with all other models.