A machine learning approach for age prediction based on trigeminal landmarks

The aim of this study was to estimate the chronological age (CA) of a growing individual using a new machine learning approach on Cone Beam Computed Tomography (CBCT).

The dataset included 48 CBCT and hand-wrist radiographs of growing individuals. 12 landmarks related to trigeminal trajectories were plotted on each CBCT and principal component analysis was applied for dimensionality reduction. The estimated CA was obtained using a decision tree. Finally, a genetic algorithm was implemented to select the best set of landmarks that would optimize the estimation. The age was also assessed following Greulich and Pyle's (GP) method on hand-wrist radiographs. The results (GP and Machine Learning) were then compared to the true CA.

Among the 12 landmarks, the genetic algorithm selected 7 optimal features, and 12 principal components out of 36. The best results for age prediction were obtained by a combination of genetic algorithm, principal component analysis, and regression tree where the Mean Squared Error (MSE) and Mean Absolute Error (MAE) were respectively 1.29 and 0.92. These outcomes showed improved accuracy compared to those of the hand-wrist method (MSE = 2.038 and MAE = 1.775).

A numerical application on a dataset of CBCT showed that the proposed machine learning method achieved an improved accuracy compared to conventional methods and had satisfying performance in assessing age for forensic purposes. Validation of the presented method on a larger and more diverse sample would pave the way for future applications in forensic science as a tool for age prediction.