Avaliação de arquiteturas convolucionais pré-treinadas na detecção de câncer oral

Abstract

Oral cancer remains one of the main public health challenges, especially in developing countries, where late detection contributes significantly to the high mortality associated with the disease. Automated methods based on Computer Vision and Deep Learning have demonstrated great potential in assisting early diagnosis, providing objective support to clinical evaluation. In this work, we investigated the performance of three widely established pre-trained convolutional architectures, DenseNet121, GoogLeNet, and ResNet18, applied to binary classification of clinical oral cavity images using a public dataset containing 950 samples. The study employs transfer learning and 5-fold cross-validation, allowing an analysis of the models' generalization capacity. The results indicate that all architectures achieved average metrics above 0.92 across all evaluated measures, including accuracy, precision, recall, F1-score, and area under the Receiver Operating Characteristic curve. Although DenseNet121 recorded the highest point estimates among the models (F1-score = 0.9476), the overlap of variability ranges based on standard deviation prevents asserting statistically significant performance differences. These findings reinforce the potential of convolutional neural networks as complementary tools in the screening of oral lesions, supporting the overall feasibility of the proposed approach in clinical scenarios.