Machine learning and deep learning enabled age estimation on medial clavicle CT images

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• The medial clavicle epiphysis is a key indicator for bone age estimation (BAE) after hand maturation.
• Current BAE methods often rely on hand-based assessments, which may be less effective after skeletal maturity.

Purpose of the Study:

• To develop and evaluate machine learning (ML) and deep learning (DL) models for BAE using medial clavicle CT images.
• To compare the performance of linear, ML, and DL models on both normal and variant clavicles.

Main Methods:

• Retrospective collection of 1049 patient CT scans, split into normal and variant training/test sets.
• Development of linear and support vector machine (SVM) models for BAE.
• Automated segmentation of medial clavicle CT slices for training DL models, including CNNs and CoAt Net.
• Evaluation using Mean Absolute Error (MAE) and classification accuracy.

Main Results:

• The SVM model achieved the lowest MAE of 1.73 years for BAE.
• Deep learning models, particularly SE Net 18, demonstrated comparable performance to SVM on normal clavicles and achieved an MAE of 2.08 years on an external variant test set.
• All models showed strong performance for age classification at 18, 20, 21, and 22 years, with limitations at the 16-year threshold.

Conclusions:

• Both ML and DL models offer desirable performance for bone age estimation based on medial clavicle CT.
• These AI-driven approaches show potential for accurate BAE, even with variant clavicles.