Related Experiment Video
Updated: Aug 5, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Automated MRI-based framework combining deep learning localization and machine learning classification for pubertal
Yuting Luo1, Xiaofei Zhang2, Shinong Pan3
1Department of Radiology, Shengjing Hospital of China Medical University, 36 Sanhao Street, Heping District, Shenyang, 110004, China.
Objectives:
This study developed a magnetic resonance imaging (MRI)-based artificial intelligence (AI) framework for ordered pubertal olecranon bone age (BA) staging and evaluated internal performance and preliminary cross-centre transportability.
Materials And Methods:
This retrospective two-centre study included a development cohort of 197 elbow MRI examinations from Shengjing Hospital of China Medical University and an independent external-validation cohort of 72 elbow MRI examinations from Handan Central Hospital. The development cohort comprised 77 girls aged 9.5-13 years and 120 boys aged 11.5-15 years, whereas the external-validation cohort comprised 30 girls aged 9.5-13 years and 42 boys aged 11.5-15 years. Manually delineated olecranon ROIs were used for preprocessing and YOLO-v8m detector training. Localized images were processed using ResNet-50 feature extraction, Fisher score-based feature selection, and an NN classifier. The locked development-centre pipeline was applied unchanged to the external-validation cohort without retraining, repeated feature selection, hyperparameter tuning, or checkpoint selection based on external data.
Results:
Automated localization achieved a mean average precision of 1.0. The final NN classifier reached internal-validation accuracies of 94.12% in girls and 93.75% in boys, with stage-mapped mean absolute error (MAE)/root mean square error (RMSE) values of 0.029/0.121 and 0.062/0.250 years, respectively. In the external cohort, girls achieved an accuracy of 86.67%, MAE of 0.233 years, and RMSE of 0.658 years, weighted kappa of 0.922, and ICC of 0.969, whereas boys achieved an accuracy of 90.48%, MAE of 0.095 years, RMSE of 0.309 years, unweighted kappa of 0.890, weighted kappa of 0.963, and ICC of 0.991.
Conclusion:
The MRI-based framework provided stage-consistent internal performance and preliminary evidence of cross-centre transportability within the studied pubertal age window. It should be interpreted as a supplementary maturity marker rather than a stand-alone forensic age-estimation tool.
Trial Registration:
Chinese Clinical Trial Registry (ChiCTR2500109735), retrospectively registered on September 24, 2025.