Classificação de regiões císticas bucais utilizando Redes Neurais Convolucionais

Abstract

Dental pathologies, such as maxillary lesions and odontogenic cysts, pose diagnostic challenges due to their difficult-to-visualize locations. Additionally, the growth of cystic regions can trigger changes in a patient’s bone structure. Therefore, it is crucial to diagnose these regions as quickly as possible. In response, various artificial intelligence (AI) systems are employed to assist dentists in detecting these regions. This work provides a comprehensive analysis of image processing techniques, convolutional neural network (CNN) architectures, optimization methods, and learning rates to enhance the diagnosis of dental cystic pathologies. The following image enhancement techniques were applied: Laplacian Filter, Prewitt Filter, Roberts Filter, and Black-hat Morphology. Additionally, the Adam, Adagrad, and RMSProp optimizers were tested alongside the AlexNet, InceptionV3, DCNNDan, and CNNGen architectures. Identifying an optimization rate of 0,0001, the Adam optimizer, and the Black-hat filter as the most effective techniques for improving CNN classification performance, the CNNGen network proposed in this research was highlighted as specialized in classifying dental cystic images. The model developed with the DCNNGen network and the Black-hat filter achieved an accuracy of 88,50% and an f1-score of 89,01%, underscoring the importance of selecting appropriate architectures and parameters for model training.