Jared Murray

TL;DR: Testing the performance of artificial intelligence to classify individuals below and above the legal age thresholds of 14, 16 and 18 years using third molar development found it able to classify male and females below and above the legal age thresholds of 14, 16 and 18 years with high accuracy.

TL;DR: This study uses convolutional neural networks to assess sexual dimorphism in human canines via radiographic morphology, achieving 57-76% accuracy across 17 age categories, with best performance at 12 years and worst at 7 years.

Abstract: Studies have relied on conventional imaging and traditional morphometric analyses of the maxillary sinuses (MS) for sex and age estimation, but little is known about the performance of deep learning models. This study aimed to evaluate the diagnostic accuracy of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) in classifying individuals by sex and age through the radiographic assessment of the MS. Panoramic radiographs of individuals aged 6–22.99 years were sampled. Twenty-one CNNs and two Transformer-based architectures were tested. Tasks consisted of binary sex and age (≤ 15 vs. >15 years) and multiclass (sex + age) classifications. For sex classification, the highest accuracies were achieved by DeiT (0.807), ViT (0.806), and EfficientNetV2M (0.781), while for age classification, YOLOv11 (0.953), ViT (0.949), and DeiT (0.946) showed the best performance. The multiclass task yielded accuracies of 0.754, 0.753 and 0.734 by YOLOv11, DeiT, and ViT, respectively. Transformers consistently outperformed conventional CNNs, while YOLOv11 and EfficientNetV2M also demonstrated competitive performance. The studied artificial intelligence models may be useful as adjuncts for binary sex and age classification, but multiclass applications are still premature needing further research before their use in forensic practice can be recommended.