Age estimation from left-hand radi̇ographs with deep learning methods

Age estimation is required in many different circumstances s, in the medical field monitoring the growth and development in children, in forensic medicine describing criminal liability cases according to age. Additionally, marriage, military recruitment, competitive sports and employment situations, and discrimination between children and adults in refugees may require age estimation. In this thesis, a computer-based decision support system which will speed up the workflow of radiologists for pediatric bone age estimation has been developed by using deep learning technology from left hand wrist radiographs. For this purpose, the performances of IncepitonV3, MobileNetV2, EfficientNetB4 and modified IncepitonV3 deep learning methods were compared. RSNA dataset which is publicly distributed in the literature and a newly formed dataset from the left-hand radiographs of children from the Kütahya province is used in this study. The data set was obtained from Kütahya University of Health Sciences with the permission of the ethics committee. For this essay, trials for two different scenarios: over 18 years old and under 18 years old were carried out. The results for below the 18 years age group with the modified IncepitonV3 transfer deep learning method on new data set are as MAE=4,3, RMSE=5,76, R2=0,99. Additionally with the proposed InceptionV3 transfer learning model on the publicly distributed data set results are as MAE=5,75, RMSE=7,42 and R2=0,96. These are one of the best results among the studies on this data set In addition, this research is one of the first deep learning studies conducted on a dataset that contains images from turkish people. Furthermore for this study, a new data set was created for bone age estimation from hand-wrist radiographs, and also a new deep learning model was recommended. Another part of this research is for bone age determination with hand-wrist radiographs for adults. Upto our knowledge there is no research in the literature for adult bone age determination with deep learning methods from left hand wrist radiographs. Deep learning methods which are DenseNet201, InceptionV3 and ResNet50V2 are used for bone age estimation of subjects over the age of 18. Performance results on the newly created data set are as following, for females MAE=1,666, RMSE=2,665, R2=97,32, for males and MAE=2,044, RMSE=3,069, R2=96,84 With these settings the method used and the results found are the first in this specific field. In the light of these results, we can conclude deep learning methods can be used as an assistant tool for determining bone age in individuals over the age of 18.