Avaliação de Arquiteturas de Redes Neurais Convolucionais para Predição do Prognóstico de Crescimento Maxilofacial em Pacientes com Fissuras Labiopalatinas

Abstract

Cleft lip and palate are the most common congenital craniofacial deformities worldwide, affecting approximately 1 in every 1000-1500 births. One of the crucial aspects of the treatment and rehabilitation of cleft patients is the analysis of maxillofacial development, performed through the GOSLON YARDSTICK, which guides surgical planning and allows early diagnosis of possible complications associated with cleft. The main objective of this work is to evaluate the performance of convolutional neural networks in predicting the prognosis of maxillofacial growth in cleft patients, as well as to develop an AI model applicable to clinical practice. This work evaluates the performance of the VGG16, VGG19, ResNet50, InceptionV3, Xception, DenseNet201, EfficientNetV2L and MobileNetV3L networks on a dataset containing 597 intraoral scan images divided into three classes, corresponding to the prognosis of facial growth. The results obtained range from 72% accuracy for EfficientNetV2L to 83.5% for VGG19. To add interpretability to the predictions, the Grad-CAM technique was used, which highlights the most important areas for classification. These results highlight the potential of AI in supporting the treatment and rehabilitation of patients with cleft lip and palate.