Multi-Task Deep Learning for Sex and Age Estimation from Panoramic Radiographs in a Brazilian Young Population

Accurate estimation of age and sex is crucial in forensic and clinical contexts, however conventional methods are subjective and time-consuming. Panoramic radiographs offer valuable data for automated analysis. Therefore, the aim of this study was to present and evaluate a multi-task deep learning framework based on ForensicNet for simultaneous estimation of chronological age and classification of sex using panoramic radiographs of the Brazilian young population aged 5-15 years.

A total of 2200 high-resolution panoramic radiographs were retrospectively collected, balanced by age and sex. After applying strict inclusion/exclusion criteria, the images were randomly split into training (1320), validation (440), and test (440) sets. A multi-task DL model based on EfficientNet-B3 was implemented with task-specific branches incorporating Convolutional Block Attention Modules (CBAM) to predict age and sex. The model was trained end-to-end using a weighted multi-task loss (α = 0.3 for age, β = 0.7 for sex) and evaluated against five benchmark architectures. Grad-CAM was used for model interpretability.

The proposed ForensicNet outperformed all baseline models, achieving the lowest mean absolute error and highest coefficient of determination in age prediction, and highest accuracy and area under the curve in sex classification. Grad-CAM visualisations confirmed the model's focus on anatomically relevant areas. Ablation studies showed that removing CBAM or altering task weights reduced performance.

The proposed ForensicNet-based multi-task deep learning model demonstrated robust performance in both chronological age estimation and sex classification using panoramic radiographs from young Brazilian individuals, supporting its potential forensic and clinical applicability.

This framework may assist forensic experts and clinicians by providing fast, objective and reproducible estimations of age and sex from routinely acquired panoramic radiographs, potentially improving identification processes in forensic and pediatric contexts.