Sexual dimorphism is a crucial step in the anthropological process of building the biological profile of the deceased [37]. In general, sex-related differences between males and females are expressed as changes in the shape and size of anatomic structures [38]. Puberty is a biological landmark that triggers more evident differences between males and females [39]. Over time, these differences will manifest especially in the pelvic bones and the skull [40]. Teeth, however, are known for their resistance to environmental effects (extrinsic factors) and systemic health conditions (intrinsic factors); and are available for forensic examination in most cases. Moreover, the radiographic visualization of dental anatomy is optimal given the highly mineralized tissues of crown and root(s). This study proposed the use of artificial intelligence for the radio-diagnostic task of sexual dimorphism from human teeth.
A preliminary challenge proposed to test the artificial intelligence in this study was the inclusion of anatomically immature individuals in the sample. This is to say that the human skeleton is not fully influenced by the hormonal changes early in life and that the maxillofacial bones are still similar between males and females in childhood. More specifically, the age limits of the addressed population were 6 and 22.9 years – an interval that covers children, adolescents, and young adults. Deciduous and some permanent teeth, on the other hand, will express full development in childhood. The permanent mandibular first molar, for instance, shows apex closure around the age of 7.5 years. Aris et al. [39], explain that teeth that fully develop long before puberty may have observable dimorphic features that can be explored even before the expression of skeletal dimorphism. Hence, the rationale at this point was to test the performance of the artificial intelligence within a scenario in which the mandible, maxillae, and other skulls bones would not play a major role in sexual dimorphism, giving the chance to teeth to express their dimorphic potential.
The radiographic aspect of the present study differs from the (physical) anthropological assessment of Aris et al. [39], because our study has the preliminary and fundamental scope of screening teeth (or tooth regions) that can play a more important part to distinguish males and females. In a future step, teeth and tooth regions detected as dimorphic in the present study could be tested and validated by means of physical examination (i.e. studies ex vivo). Among the main advantages of the radiographic approach is the visualization of dental anatomy, including the internal aspect of the crown and roots (namely the pulp chamber and root canals, respectively), and the possibility of retrospective dataset sampling from existing databases – which is hampered in observational anthropological/archaeological studies.
DenseNet121 architecture running with TL training approach in 100 epochs led to the best performance for sexual dimorphism. Particularly, the training accuracy maintained high (above 80%) between epochs 19–100, while the validation accuracy was between 70–83% after epoch 31. Consequently, the average accuracy of TL was 82%, with average specificity of 92% in the total sample. Authors claim [41] that when the entire skeleton is available for anthropological assessment, the accuracy of sexual dimorphism can reach 100%. This phenomenon is justified by the contribution of pelvic bones and skull to the analyses. Studies solely based on teeth present much lower estimates. Paknahad et al. [42], for instance, performed a study with bitewing radiographs and reported an accuracy of 68% for sexual dimorphism based on odontometric assessments of the deciduous second molars (mandibular and maxillary). In our study, the higher accuracy rates are possibly justified by the integral assessment of dental anatomy (all the visible bidimensional dental features of the teeth were considered) in the process of sexual dimorphism – instead of specific linear measurements. In the study of Paknahad et al. [42], only the width of the enamel, dentin, and pulp space were considered. Moreover, our study assessed radiographs of 4,003 individuals, while the previous authors [42] sampled only 124 individuals. In practice, a preliminary overall accuracy of 82% (specificity of 92%) corroborates DenseNet121 with TL approach as a proper tool for radiographic sexual dimorphism.
The purpose of the present study, however, was to challenge to artificial intelligence even more. To that end, the sample was divided into males and females below and above the age of 15 years. ROC curves obtained during the analyses per age category showed AUC between 0.90–0.91 for males and females over the age of 15, respectively, while in the younger group the AUC was 0.87 for both the males and females. These outcomes confirm that, in fact, sexual dimorphism is more challenging among children (in this case, between 6 and 14.9 years). In both groups, however, the AUC was considered excellent for diagnostic accuracy tests [43]. Consequently, the features assessed from panoramic radiographs in the present study had enough discriminant power to distinguish males and females with accurate performance.
The Grad-CAM images obtained in our study showed a similar region of activation in both age groups. In general, the activation region was more centralized and horizontal – surrounding the crowns of anterior and posterior teeth. These outcomes are corroborated by studies that show the dimorphic value of canines [44, 45] and incisors [41] between males and females.
This is a preliminary study to understand the discriminant power of dental morphology to distinguish males and females using panoramic radiographs. At this point, these outcomes should not be translated to practice since they currently serve to screen regions of teeth that may weigh more for sexual dimorphism. A few cases in the scientific literature reported the use of postmortem panoramic radiographs for human identification [46, 47]. In these cases, the current findings could have a more tangible application. For anthropological practices in single cases and mass disasters, more comprehensive knowledge of radiographic sexual dimorphism is needed, especially when it comes to the effects of age on dental morphological features.