Research On The Application Of Deep Learning To Age Estimation With Panoramic Radiographs

Section 1:Comparison of conventional and deep learning algorithms in age estimation from mandibular canines in panoramic radiographsObjectives:To compare conventional and convolutional neural network algorithms in human age estimation from mandibular canines in panoramic radiographs.Materials and methods:Panoramic images of 360 female and 360 male patients were divided into a training set that contained 600 patients' data for the establishment of models and a test set containing 120 patients' data for the test.The mandibular canines were segmented as the region of interest(ROI).A logarithmic regression model was established based on the manual feature of the pulp/tooth area.Meanwhile,one end-to-end convolutional neural network(CNN)model was established by Five-fold cross-validation.Both two models were subsequently tested with the same data in the test set.The mean absolute errors were calculated and the chronological ages and the estimated ages were compared for each age group.Results:The logarithmic regression model was established as:AGE=-82.819-45.3471n(pulp/tooth area).The coefficient of determination(R2)was 0.823,and the regression was statistically significant(p=0.000<0.01).The mean absolute error(MAE)obtained from the logarithmic regression model and the convolutional neural networks were 6.209 and 5.358,respectively.However,when the chronological ages and the estimated age were compared,only the ages estimated from the logarithmic regression model were not significantly different from the corresponding chronological ages in the age groups of 12-21,32-41,and 41-52.Conclusions:Although the model established by the deep learning algorithm provides a low MAE for age estimation,the estimation accuracy still needs further investigation.Section 2:Comparison of age estimation by deep learning from different teeth in panoramic radiographsObjectives:To compare convolutional neural network algorithms for age estimation with different teeth in panoramic radiographs.Materials and methods:Panoramic images of 360 female and 360 male patients were divided into a training set that contained 600 patients' data for the establishment of models and a test set containing 120 patients' data for the test.Different teeth(11/21,31/41,13/23,33/43,14/24,34/44)in the maxilla and mandible were segmented as the region of interest(ROI).Meanwhile,the transfer learning models were established by Five-fold cross-validation according to the different teeth.All the models were subsequently tested with the corresponding data in the test set.The mean absolute errors were calculated and the chronological ages and the estimated ages were compared for each age group.Results:The average absolute error(MAE)of the different models obtained from the corresponding test set are:11/21(9.98),31/41(10.40),13/23(6.33),33/43(5.36),14/24(8.74),and 34/44(8.88).Conclusions:The deep learning model based on mandibular canines has the minimum average absolute error and the deep learning model based on the mandibular central incisors has the maximum average absolute error.Section 3:Age estimation using panoramic radiograph by transfer learningObjectives:There have been many studies using panoramic radiographs for age estimation.Traditional methods rely on manually labeled features,which are complicated and time-consuming.A few studies used deep learning to estimate age automatically,but most of the experimental samples are unequally distributed and a number of samples fall in low-age groups.This study aims to assess the accuracy of deep learning in age estimation from equally distributed permanent dentition in panoramic radiographs and provide a new method for age estimation.Materials and methods:Equally distributed 3000 panoramic radiographs were divided into three parts:a training set(2400),a validation set(300),and a test set(300).The transfer learning model was established by Five-fold cross-validation based on the panoramic radiographs.The model was subsequently tested with the data in the test set.The mean absolute error was calculated and the chronological ages and the estimated ages were compared for each age group.Results:The average absolute error(MAE)of the transfer learning model used for age estimation in the validation set and test set are 2.93 and 2.95,respectively.Conclusions:The transfer learning model can extract different features in different age groups and can be used for age estimation in panoramic radiograph.