Legal age estimation using orthopantomographies

This is an example of an image used to train our model. The model has been trained with data from 12 countries and 4 continents so it can generalize to data from never seen samples. It gets a mean absolute error of 1.07 years and an accuracy of 88.38% (for the 18 years threshold) in test. Regarding never seen samples, it gets a mean absolute error of 1.21, 1.45, 1.36 and 1.51 years and an accuracy of 92.04, 81.97, 85.03 and 86.08% in four samples from Russia, Ethiopia, Egypt and Australia, respectively.

Confidence Percentage of minors correctly classified Percentage of adults correctly classified Accuracy
Original model 82.42% 92.30% 88.38%
60% 93.90% 81.14% 86.20%
80% 96.72% 70.67% 80.99%
90% 98.24% 62.05% 76.39%
95% 98.59% 54.12% 71.75%

These are the results of our model when estimating age as the minimum of a prediction interval for different confidence values. This app will return prediction intervals for all the shown confidence values so the minimum age concept can be applied when required.