Scale-free sex estimation from the bony pelvis using deep learning: A proof of concept

Abstract

Aim: Sex estimation from the pelvis is a cornerstone of forensic anthropological identification, but the established methods require physical access to the bone together with either metric measurement or trained morphological scoring. We tested, as a proof of concept, whether sex can be estimated automatically from a single, unscaled image of the whole bony pelvis, as a step toward a scale-free, photograph-based field tool. Materials and Methods: From computed tomography (CT) examinations of 400 adults (200 female, 200 male; mean age 56.9 years), the pelvis was isolated and rendered as two-dimensional projections from five standard orientations (anterior, posterior, superior, inferior and lateral), yielding 2,000 images. After segmenting the bone and discarding absolute scale, an EfficientNet-B0 convolutional neural network was trained by transfer learning with strict subject-level partitioning, and evaluated by subject-grouped five-fold cross-validation and on an internal holdout split. Calibrated posteriors were converted to likelihood ratios (LRs); saliency mapping and a blur-based confound control identified what drove classification. Results: Under cross-validation, sex was classified with an accuracy of 0.934 (95% confidence interval [CI] 0.922 – 0.946; area under the receiver-operating-characteristic curve [AUC] 0.982) for a single image and 0.993 (AUC 0.9998) when the five views were combined; the internal holdout split gave 0.865 and 0.975. All five orientations exceeded an AUC of 0.97 (range 0.975 – 0.989), the anterior view being modestly but significantly better than the posterior (ΔAUC 0.014; p = .02). Combining any two views already raised specimen accuracy to a mean of 0.931. The image-level log-likelihood-ratio cost was 0.420 (calibration-minimised value 0.341). Discriminability was preserved under a blur-based confound control (specimen AUC ≥ 0.996), and saliency was concentrated on bone at 1.77 times its area share. Conclusions: Scale-free, image-based pelvic sex estimation is feasible; validation on real photographs and across populations, and a calibrated multi-view fusion rule, are required before forensic use.