Conformalized Deep Regression for Age Estimation from Panoramic Radiographs: An Uncertainty-Aware Approach

Background/Objectives: Accurate age estimation is a critical task in forensic and legal contexts, where reliable and objective assessment is essential for age-dependent decision-making. This study aims to explore the application of convolutional neural network-based regression models for age estimation from panoramic radiographs, incorporating conformal prediction to generate prediction intervals representing uncertainty. The estimated ages were also evaluated for their application in legal age classification at thresholds of 10, 15, 18, and 20 years. Methods: A total of 1734 panoramic radiographs of individuals aged 8 to 23 years were used in this study. We employed six models, including ResNet-50, EfficientNet-B0, DenseNet-121, MobileNetV3-Small, VGG-16, and ResNeXt-50, to perform regression and provide point estimates of chronological age. Conformal prediction using a split conformal framework with absolute residual scores was implemented to construct statistically valid prediction intervals with finite-sample coverage guarantees. Results: The results revealed that ResNeXt-50 achieved the lowest mean absolute error for point estimates (0.92 years) and produced among the narrowest conformal margins, ranging from 1.71 to 3.72 years across non-coverage probabilities from 0.20 to 0.01. In legal age classification, the ResNeXt-50 model achieved its highest classification accuracy using point estimates, whereas the application of the conformal lower bound improved specificity and positive predictive value. This shift facilitates conservative decision-making, aligning with the forensic minimum-age principle. Conclusions: The proposed framework offers a practical solution for automated age estimation by quantifying uncertainty through prediction intervals, allowing for more informed assessments aligned with forensic protocols for legal age determination.