Age prediction with hybrid deep learning methods using panoramic, cephalometric, and hand-wrist radiographs

The use of deep learning architectures in the solution process of many real-world problems in engineering science, including the field of health, is becoming increasingly widespread. Today, the most frequently and reliably used methods for age prediction are the methods that evaluate tooth and bone development. Bone age prediction is medically important for the investigation of growth, development, and endocrine disorders in children, while it is legally critical in the determination of criminal liability and international adoption processes, and in fields such as anthropology and forensic medicine. When estimating age on individuals, the development of teeth, skull, and wrist bones, which provide the most reliable results of the human body, is examined and age prediction is conducted. Traditionally, age prediction from teeth is carried out in 2 different periods. The development stages and eruption times of temporary and permanent teeth in children are evaluated. In this study, a deep learning-based hybrid system aimed at determining the age of an individual with panoramic, cephalometric, and hand-wrist radiographs is proposed. Seven different data sets consisting of two or three group classes were used in the study. The classes in the datasets are generally in the age range of 2-6, 6-13, 13-21, 2-13 and 13-21. The reliability and performance of the proposed system were analyzed by applying the same methods to panoramic, cephalometric, and hand-wrist radiographs collected from the same individuals. First, seven different feature extraction architectures were applied to all datasets in the study. The obtained feature map was set as input to the proposed system and five different machine learning methods, three different deep learning methods, and two different hybrid deep learning methods were applied. The system performance was tested by comparing the obtained performance evaluation metrics. The highest 91% accuracy, precision, sensitivity, and F1-score performance values were obtained with InceptionV3 and CNN+RNN hybrid model.