MT-DWL ViT-KAN: estimação simultânea de sexo e idade a partir de radiografias panorâmicas.

Abstract

Forensic dentistry is a field that applies scientific knowledge to analyze dental elements, such as teeth and dental arches, with the aim of identifying individuals. This analysis is especially valuable in situations such as mass disasters, criminal investigations, and missing persons cases. Traditionally, sex and age determination is done through morphological and metric analyses of dental and bone structures. However, these approaches have limitations, such as data variability and subjectivity of the analyses. With advances in deep learning, it has become possible to apply computer vision to analyze dental radiographs, using classification and regression techniques. This dissertation proposes a Multitask-Dynamic Weighted Loss Vision Transformer-Kolmogorov-Arnold Networks (MT-DWL ViT-KAN) approach that combines multitask learning with self-supervised pre-training using Mask Autoencoders, allowing the model to learn robust latent representations from large volumes of unlabeled data. Furthermore, Kolmogorov-Arnold networks are employed to decompose complex relationships between dental features and sex and age labels, improving model accuracy. The methodology also incorporates a dynamic logarithmic weighted loss function, which automatically adjusts task weights during training, ensuring an optimized balance between sex classification and age estimation. In our experiments, incorporating the Mask Autoencoders strategy, which uses the Vision Transformer architecture, resulted in a significant improvement in model performance compared to the EfficientNetV2-L model pre-trained on ImageNet. Furthermore, by integrating the Kolmogorov-Arnold network into the final multi-task layer, we obtained the best results among all tested configurations. MT-DWL ViT-KAN achieved a mean absolute error of 3.39 years in age estimation and an F1-score of 94.2\% in sex classification. These results highlight the potential of the proposed model in extracting relevant features from dental radiographs and in performing multitask predictions in the forensic dentistry scenario.