Comparison of Deep Learning Methods for Age Determination in Forensic Clients Aged 5-18 Years

Age determination of individuals, a critical aspect of enforcing criminal laws worldwide, is a key responsibility of forensic medicine. This study aimed to design an expert system to assist in determining the age of individuals aged 5-18 years as a medical decision supporting tool. The study was conducted in three phases in 2022, focusing on forensic patients in the northern Iran. In the first phase, a literature review was performed to identify deep learning methods relevant to age determination. In the second phase, real data were collected from forensic patients in Gilan province, Rasht, Iran, with the age range of 5 to 18 years. The average actual age was 10.86 years, with a standard deviation of 2.61. In the final phase, the expert system was designed and implemented using Deep Belief Networks (DBN) and Convolutional Neural Networks (CNN), including DenseNet121, ResNet50 (with and without freezing layers), Xception (with and without freezing layers), and VGG19 (with and without freezing layers). All methods were evaluated using 70 radiographs of the left wrist of forensic clients with known actual age. Performance was compared across methods using evaluation metrics including R2-score, Mean Squared Error (MSE), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). The findings provide insights into the most effective approaches for automated age determination in forensic applications.