Análise comparativa de arquiteturas de Redes Neurais Convolucionais na classificação etária forense a partir de radiografias panorâmicas odontológicas

Abstract

Age estimation based on panoramic radiographs is a well-established practice in forensic dentistry, playing a key role in contexts such as adoption, immigration, criminal responsibility, and victim identification. This study proposed the use of convolutional neural networks (CNNs) applied to forensic age classification from orthopantomograms, exploring the impact of different image preprocessing strategies on model performance. Three architectures commonly employed in the literature were investigated — VGG19, InceptionV3, and EfficientNetV2B0 — all initialized with pre-trained ImageNet weights. The preprocessing pipeline included resizing to 128×128 pixels, maintaining images in RGB format, normalization, application of data augmentation techniques, and the use of eight processing variations, such as edge filters (Sobel, Prewitt, Roberts, Laplacian), morphological transformations (Top-hat and Black-hat), and gamma correction. Training was conducted using 4-fold stratified cross-validation, and performance metrics — accuracy, precision, recall, and F1-score—were computed based on the average across folds. The results revealed performance variations between filters, indicating that certain visual enhancement techniques can either benefit or impair the predictive capacity of CNNs. The comparative analysis among architectures demonstrated differences in sensitivity to preprocessing, underscoring the importance of this stage in clinical and forensic applications.