Investigação de técnicas de aumento de dados para classificação de lesões da cavidade oral

Abstract

The early diagnosis of oral cancer is crucial in determining the patient's prognosis, as it allows healthcare professionals to take preventive measures, reducing the risk of progression to advanced stages of the disease. However, the evaluation of oral cavity lesions relies on the interpretation of clinical features by specialists, which can lead to differences in opinions and diagnostic challenges. The use of diagnostic support systems can be applied as a tool to assist and improve the classification and decision-making process of specialists. However, these systems still heavily depend on robust training datasets, and obtaining histological images poses several difficulties and challenges. To enhance the diagnostic process of oral cavity lesions, this study proposes the investigation of data augmentation techniques as an alternative to address the scarcity of histological images for the lesion classification stage. Data augmentation techniques involving geometric transformations and Mixup, which combines two or more samples from the training dataset to create a new augmented sample, are explored. For classification, the CNN models ResNet50 and DenseNet201 are utilized. The combination of data augmentation techniques involving geometric transformations and Mixup yielded the best results and demonstrated higher generalization capacity. The strategy achieved an accuracy of 90.6% for the oral dysplasia dataset using both data augmentation models. For the healthy and carcinoma dataset, it achieved F1-scores of 91.7% with the ResNet50 model and 94.9% with the DenseNet201 model. The investigation revealed that the Mixup technique does not replace conventional data augmentation techniques; however, it can complement them for advancements in classification.