Whole-axis versus odontoid process CBCT radiomics for forensic sex estimation: a retrospective paired-ROI...

The second cervical vertebra (C2, axis) is included in head-and-neck cone-beam computed tomography (CBCT) and may provide a target for forensic sex estimation. Whether odontoid-only segmentation preserves whole-C2 radiomic information remains unclear.

To develop and internally validate paired CBCT radiomics models using whole-C2 and odontoid-only regions of interest (ROIs), and to compare classification performance while describing feature-level reproducibility for each ROI.

This retrospective single-device study included 140 adult CBCT examinations (68 female, 72 male). Each ROI yielded 107 original-image radiomic features. Shape-only, texture-only, and combined LASSO models and a combined elastic-net model were assessed using training-set-only preprocessing, a balanced 80/20 test split, 100 repeated paired validations, five-fold nested cross-validation, and label-permutation testing.

Across 100 repetitions, mean AUCs were 0.929 for both whole-C2 combined models and 0.924, 0.918, and 0.905 for odontoid texture-only, combined LASSO, and combined elastic-net models, respectively. Nested-cross-validation AUCs for combined LASSO were 0.918 for whole-C2 and 0.914 for odontoid-only; permutation null distributions were centered near 0.50 (both p = 0.003). The highest primary-split AUCs were 0.985-0.995, but these estimates were derived from 28 test subjects and were interpreted as optimistic. Both ROIs showed acceptable descriptive reproducibility, but the small reliability subsets did not support comparative conclusions.

C2 CBCT radiomics showed promising internal discrimination. Whole-C2 retained more shape information, whereas odontoid-only texture models showed similar internal performance with a smaller target. High dimensionality, possible feature-selection instability, uncertain biological meaning of texture descriptors, homogeneous imaging conditions, and absent external validation limit casework applicability.