The development of an age estimation methodology using MR images from a Scottish population

Recently, new methodologies for age estimation in the living have been proposed using magnetic resonance imaging and the implementation of machine learning algorithms. The present study analysed 2-dimensional T1- weighted and Proton density magnetic resonance images for the age estimation of males and females between 10 years and 26 years of age. Four different datasets were included in this study, female T1-weighted (FT1) with a total of 125 individuals, male T1-weighted (MT1) with a total of 79 individuals, female proton density (FPD) with a total of 178 individuals, and male proton density (MPD) with a total of 99 individuals. The epiphyseal gaps and the epiphyses of the radius and ulna were segmented using pixel segmentation, and the areas of these features calculated from the generated masks using a MATLAB script. The areas were organised by slices and by individuals, and traditional machine learning techniques were implemented to carry out the age estimation. The machine learning algorithms implemented were Decision Tree Regressor and Random Forest Regressor, for a total of 32 Experiments. Between the Decision Tree Regressor models and the Random Forest Regressors models, the results showed that the Random Forest Regressors were the models that best performed on the test data. Among these, the highest R2 and lowest mean absolute error values were obtained for the Experiments carried out using a combination of epiphyseal gaps areas and epiphyseal areas of radius and ulna (FT1 – R2 0.80, MAE 1.60 years; FPD – R2 0.63, MAE 2.17 years; MT1 – R2 0.72, MAE 1.57 years; MPD – R2 0.69, MAE 1.95). Results of the Feature Importance Analysis and the Pearson’s Correlation Analysis showed that the feature most important for the estimation of the age, and most correlated with the age was the epiphyseal gap of the radius. Further studies are necessary to improve the methodology and investigate further the implementation of machine learning algorithms using 2D magnetic resonance images.