End-to-end vs. human-defined feature extraction: comparing deep learning approaches for age classification using...

Accurate age classification using mandibular third molar radiographs is crucial for legal and forensic applications. This study evaluated different methods for classifying age as under or over 18 years in a Thai population. We compared three approaches: (i) a traditional human-based method using a modified Demirjian classification adapted for mandibular third molars, (ii) an end-to-end deep learning model in which a convolutional neural network (CNN) directly predicts age group, and (iii) a human-defined feature extraction approach, where a CNN estimates tooth developmental stages that are subsequently used for age classification. The dataset included 3,407 images of individuals aged 14-23 years. The results indicated that the traditional human-based method achieved high specificity (0.99) and a strong Bayes' post-test probability (0.99), but it exhibited low sensitivity (0.45). In comparison, the end-to-end deep learning models showed higher sensitivity (0.65 to 0.74) than the traditional method, along with a specificity of 0.91 to 0.95 and Bayes' post-test probability of 0.93 to 0.95. The human-defined feature extraction approach, which used developmental stages for age determination, achieved an accuracy of 0.88 to 0.92 in developmental stage classification. For age classification, the models demonstrated higher specificity (0.95 to 0.97) and Bayes' post-test probability (0.95 to 0.97) than the end-to-end deep learning method, along with sensitivity ranging from 0.51 to 0.56. Our results indicate that although traditional methods excel in specificity, the human-defined feature extraction approach provides a balanced solution with high specificity and interpretability, suggesting its potential value in clinical practice for age estimation.