Adult age estimation using multi-domain mandibular CT features from postmortem head CT: Exploratory internal validation with nested cross-validation and age-band-specific error analysis.
Authors
Affiliations (4)
Affiliations (4)
- Department of Forensic Dental Science, Unit of Social Medicine, Nagasaki University Graduate School of Biomedical Sciences, Nagasaki University, 1-12-4 Sakamoto, Nagasaki 852-8523, Japan.
- Department of Forensic Pathology and Science, Unit of Social Medicine, Nagasaki University Graduate School of Biomedical Sciences, Nagasaki University, 1-12-4 Sakamoto, Nagasaki 852-8523, Japan.
- Department of Forensic Pathology and Science, Unit of Social Medicine, Nagasaki University Graduate School of Biomedical Sciences, Nagasaki University, 1-12-4 Sakamoto, Nagasaki 852-8523, Japan. Electronic address: [email protected].
- Department of Forensic Medicine, Faculty of Medicine, Kagawa University, 1750-1 Miki, Kita, Kagawa 761-0793, Japan.
Abstract
Adult age estimation remains difficult because post-maturational skeletal changes are gradual and heterogeneous. We evaluated multi-domain mandibular features extracted from postmortem computed tomography (PMCT) of the head and quantified overall, age-band-specific, sex-specific, and scanner-specific error patterns. This single-institution retrospective study included 1723 individuals aged 20-100 years, with one examination per individual. The mandible and teeth-related structures were segmented with TotalSegmentator version 2.12.0; mandibular pose was standardized; and dental metal- and tooth-affected regions were excluded. HU-related features, cortical-thickness proxies, apparent trabecular parameters, medullary-expansion and porosity proxies, and morphometric features were extracted. Candidate feature sets were selected within each outer training set by three-fold inner cross-validation, and HistGradientBoostingRegressor was evaluated by five-fold outer cross-validation. Predicted and chronological age correlated at r = 0.628; MAE was 10.34 years, RMSE 13.17 years, bias -0.02 years, and p90_abs 21.67 years. MAE was 9.10 years in females and 10.97 years in males. Bias was +24.38 years at ages 20-29 and - 15.18 years at ages 90-100. The multi-domain model reduced MAE by 1.04-5.22 years relative to single-domain models. Coverage of age-band-specific p90_abs ranges was 0.806 when bands were assigned from predicted age, showing inadequate calibration for individual prediction intervals. Multi-domain mandibular analysis from head PMCT may support reproducible forensic screening when accompanied by explicit error profiles and expert assessment; external validation and calibrated intervals remain necessary.