Three-dimensional deep-learning age estimation from the fourth rib on high-resolution CT: An evaluative likelihood-ratio approach.
Authors
Affiliations (2)
Affiliations (2)
- Department of Forensic Medicine, Faculty of Medicine, Van Yüzüncü Yıl University, Van, Türkiye. Electronic address: [email protected].
- Department of Computer Engineering, OSTIM Technical University, Ankara, Türkiye.
Abstract
For four decades, age estimation from the fourth rib has rested on İşcan's visual phase system, which requires the physical bone and returns an ordinal category, not a calibrated probability. We developed a three-dimensional deep-learning pipeline, automated downstream of AI-assisted segmentation, that estimates age from the whole rib on computed tomography (CT) and reports a likelihood ratio. Right fourth ribs from thin-section (1 mm) chest CT of 453 living adults (225 men, 228 women; 18.0-99.8 years) were segmented, posed canonically and encoded as four representations: metric multi-view RGB-D renderings and a point cloud of the sternal end, a point cloud of the complete rib, and twenty morphometric scalars. Sex-specific networks trained by label-distribution learning were evaluated on a frozen hold-out set (three seeds). The selected estimators reached a mean absolute error of 10.6 years in men and 10.4 in women (45-46% below a mean-age predictor; 9.0 and 8.4 years between ages 18 and 70). Likelihood ratios were informative at every threshold from 30 to 70 years (Cllr < 1), sex was classified with an area under the ROC curve of 0.966, and intra-observer reproducibility was ICC 0.872, limited chiefly by the canonical pose. The age signal was not confined to the sternal extremity examined by phase scoring: in women the complete rib outperformed the sternal end (10.4 versus 14.6 years); in men the two were equivalent, consistent with sex-specific standards. Because 1 mm sections cannot resolve İşcan's fine relief, higher-resolution surface capture is a priority for future work.