Age Estimation of the Cervical Vertebrae Region Using Deep Learning.

Area of Science:

• Medical Imaging Analysis
• Deep Learning
• Forensic Anthropology
• Radiographic Age Estimation

Background:

• Traditional methods struggle to detect subtle age-related bone changes in adulthood.
• Deep learning shows promise for medical image-based age estimation.
• Cervical vertebrae in lateral cephalometric radiographs (LCR) are valuable for age assessment.

Purpose of the Study:

• To systematically investigate the impact of different cervical vertebral representations on age estimation accuracy.
• To compare the performance of Contour (C), Mask (M), Cervical Vertebrae (CV), and Cervical Vertebrae Region (SR) input modes.
• To evaluate age estimation across different age groups (4-40 years) using deep learning.

Main Methods:

• Developed four distinct input modes (C, M, CV, SR) for deep learning models.
• Utilized a large-scale dataset of 20,174 LCRs from subjects aged 4-40 years.
• Evaluated performance using Mean Absolute Error (MAE), analyzing individual and combined vertebrae.

Main Results:

• The Cervical Vertebrae Region (SR) mode achieved the lowest overall MAE, outperforming CV, C, and M modes.
• SR and CV modes maintained MAE below 10 years for the 26-40 age group, unlike C and M modes.
• Combining vertebrae improved accuracy, with continuous combinations (e.g., C1-2 + C3) showing better results than discontinuous ones.

Conclusions:

• Incorporating peripheral soft tissue and comprehensive vertebral context (SR mode) is crucial for accurate age estimation.
• The SR mode offers superior performance, especially for older age groups, compared to methods focusing solely on bone structure.
• Deep learning models leveraging advanced vertebral representations can effectively estimate age from LCRs across a broad spectrum.