Semi-automated sex estimation from the cochlea on routine clinical CT: limited incremental value of curvature-based shape descriptors.
Authors
Affiliations (7)
Affiliations (7)
- IUSTI, Aix Marseille University, CNRS, Marseille, France.
- Department of Otolaryngology, Hospital Fondation Adolphe de Rothschild, 29 Rue Manin, 75019, Paris, France.
- MED-EL Corporation, Canada, 123 Edward Street, Suite 908, Toronto, ON, M5G 1E2, Canada.
- ENT-HNS Department, APHM, CNRS, IUSTI, Aix Marseille University, La Conception University Hospital, 147 Boulevard Baille, 13005, Marseille, France.
- MED-EL Medical Electronics, Fuerstenweg 77a, 6020, Innsbruck, Austria.
- MED-EL FRANCE, 400 Avenue Roumanille, Bât. 6 - CS 70062, 06902, Sophia-Antipolis, France.
- ENT-HNS Department, APHM, CNRS, IUSTI, Aix Marseille University, La Conception University Hospital, 147 Boulevard Baille, 13005, Marseille, France. [email protected].
Abstract
Sex estimation is fundamental in forensic anthropology and bioarchaeology, yet classical methods rely on skeletal elements that are often absent or poorly preserved. The bony labyrinth, enclosed within the petrous bone, is a promising alternative owing to its early maturation and taphonomic resistance, and cochlear shape has been reported to carry a sex-specific signal. However, most approaches depend on manual segmentation, micro-CT, or complex shape-analysis frameworks that limit routine applicability. To evaluate the feasibility of a semi-automated method for sex estimation from the cochlea using standard clinical CT and automated OTOPLAN segmentation, and to quantify the contribution of shape descriptors beyond cochlear size alone. We analyzed 139 cochleae from 87 adults using clinical CT scans (slice thickness ≤ 0.6 mm). The three-dimensional lateral wall path was extracted with OTOPLAN; after normalization and resampling, curvature- and torsion-based descriptors and scalar cochlear metrics were computed to train a supervised classifier. Preprocessing, feature, and hyperparameter choices were optimized by Bayesian optimization within a nested cross-validation, so that model selection and performance estimation used separate data. Six feature configurations were compared, with all partitions made at the patient level and the whole procedure repeated over five random partitions. In nested cross-validation, the best configurations reached an area under the ROC curve (AUC-ROC) of approximately 0.71. A model using cochlear diameter as its sole predictor did not differ detectably from those combining size and shape descriptors (AUC-ROC 0.71, 95% CI 0.62-0.80) and gave the highest balanced accuracy (0.70), with a sensitivity of 0.71 and a specificity of 0.70. Configurations restricted to shape descriptors reached a median AUC-ROC of 0.59, with lower confidence limits at or just above chance level (0.50-0.51). Adding curvature to diameter changed the median AUC by + 0.011 (95% CI - 0.064 to + 0.107). Sex-related information is recoverable from cochlear geometry on routine clinical CT using a semi-automated pipeline, but no incremental value of shape descriptors beyond cochlear size was detected. Although accuracy remains moderate, the approach is non-destructive, scalable, and clinically compatible, offering a potential adjunct for forensic anthropology and bioarchaeology when conventional methods are unavailable. This is not an identification method and is not validated as a stand-alone forensic tool: it was evaluated by internal validation only, on in vivo clinical CT.