Redes neurais profundas para classificar imagens de radiografias odontológica panorâmica na perspectiva de grupos etários: um estudo exploratório

Abstract

Estimating chronological age is an essential process in forensic dentistry, as it allows inferring an individual’s age group based on biometric characteristics, particularly those observable in dental radiographs. Although traditional methods, such as those proposed by Nolla and Demirjian, are widely adopted, they often require expert interpretation and may present limitations regarding accuracy and scalability. Deep neural networks have emerged as a promising alternative in this context, offering a non-invasive, automated, and low-cost approach. This dissertation comprises three main parts. The first presents a systematic literature review aimed at identifying the state of the art in using deep learning for age estimation based on panoramic dental radiographs. Following a rigorous selection process, 17 primary studies were included and analyzed. Key f indings indicate that the ResNet and AlexNet architectures stand out among the evaluated approaches, particularly for younger age groups (3 to 14 years), and that accuracy and recall are the most frequently used performance metrics. The second stage involved the construction of a dataset composed of 1,545 panoramic radiographic images obtained through partnerships with dental clinics. All images were anonymized and standardized, and a mask was applied to highlight the anatomical region of interest. This dataset aims to facilitate replication and future expansion in similar studies. Finally, the third part of this work consisted of an exploratory study using seven convolutional neural network architectures (AlexNet, VGG16, ResNet50, DenseNet, MobileNetV2, EfficientNet-B3, and InceptionV3), aiming to evaluate their effectiveness in estimating age based on the constructed dataset. The results support previous findings from the literature and demonstrate the potential of deep learning approaches in real-world forensic scenarios. This research contributes valuable insights for developing intelligent solutions in forensic dentistry, supporting researchers, tool developers, and healthcare professionals.