Radiographic morphology of canines tested for sexual dimorphism via convolutional-neural-network-based artificial intelligence

Abstract

The permanent left mandibular canines have been used for sexual dimorphism when human identification is necessary. Controversy remains whether the morphology of these teeth is actually useful to distinguish males and females. This study aimed to assess the sexual dimorphism of canines by means of a pioneering artificial intelligence approach to this end. A sample of 13,046 teeth radiographically registered from 5838 males and 7208 females between the ages of 6 and 22.99 years was collected. The images were annotated using Darwin V7 software. DenseNet121 was used and tested based on binary answers regarding the sex (male or female) of the individuals for 17 age categories of one year each (i.e. 6–6.99, 7.7.99… 22.22.99). Accuracy rates, receiver operating characteristic (ROC) curves and confusion matrices were used to quantify and express the artificial intelligence's classification performance. The accuracy rates across age categories were between 57–76% (mean: 68% ± 5%). The area under the curve (AUC) of the ROC analysis was between 0.58 and 0.77. The best performances were observed around the age of 12 years, while the worst were around the age of 7 years. The morphological analysis of canines for sex estimation should be restricted and allowed in practice only when other sources of dimorphic anatomic features are not available.