Introduction
Sex estimation is a cornerstone of forensic anthropology, helping to build a biological profile from skeletal remains. While much research emphasizes the pelvis and skull, the sternum offers an underexplored source of sexual dimorphism that may be particularly useful when other bones are damaged or unavailable. This article reviews how various machine learning methods and deep neural networks can improve sex estimation from the sternum in Turkish populations, with implications for forensic practice and research design.
Why the sternum and population specificity matter
The sternum comprises the manubrium, corpus, and xiphoid process, each contributing potential sex-related differences in size, shape, and angular features. In the Turkish population, morphological and metric sternum traits may differ from those in other groups due to genetic, environmental, and developmental factors. Consequently, population-specific models that leverage sternum measurements and imaging data can offer more reliable sex classification than generalized models. The sternum is also often preserved in cases where long bones are missing, making it a practical resource for forensic analyses.
Data and feature types used for sternum based sex estimation
Studies typically utilize radiographs or computed tomography CT scans to extract both metric measurements (lengths, widths, heights, and angles) and geometric morphology data. In addition, three dimensional models and landmark configurations enable shape analysis that captures subtle dimorphic traits in the sternum. A Turkish population study would assemble a dataset with known sex and age, ensuring representation across age ranges while mitigating potential biases from sampling. Feature extraction usually involves standardizing measurements, handling missing data, and applying quality controls to ensure consistent landmark placement.
Machine learning approaches for sternum based sex estimation
Classical machine learning methods
Traditional machine learning algorithms, including logistic regression, random forest, support vector machines, and gradient boosting, can be trained on sternum derived features. These models offer interpretability through feature importance scores and decision boundaries, allowing forensic scientists to understand which sternum metrics drive sex classification. Model selection relies on cross validation, hyperparameter tuning, and careful handling of class imbalance that may arise in forensic datasets.
Deep neural networks
Deep neural networks (DNNs) bring the ability to learn complex, non linear patterns directly from imaging data. Convolutional neural networks can process sternum images to predict sex, while 3D CNNs can operate on CT volumes to exploit internal osseous textures. Graph neural networks can model relationships among sternum landmarks, offering a balance between 2D and 3D representations. Transfer learning, using pre trained networks on related osteological datasets, can enhance performance when Turkish population specific data are limited. Model interpretability tools, such as saliency maps and feature attribution, help validate which sternum features are most informative for sex estimation.
Evaluation and validation considerations
Key metrics for assessing sternum based sex estimation include accuracy, area under the ROC curve (AUC), sensitivity, and specificity. Robust evaluation requires cross validation within the Turkish data and external validation on independent Turkish samples to gauge generalizability. It is crucial to report uncertainty, confidence intervals, and potential biases arising from age, health status, preservation state, and sampling methods. Ethical considerations include respecting consent (where applicable), anonymization of data, and transparent communication with forensic stakeholders about the limits of sternum based models.
Implications for forensic practice
In practice, sternum based ML and DNN approaches can complement traditional methods based on the pelvis and skull, especially when other bones are compromised. Automated pipelines may accelerate analysis while preserving interpretability through feature importance and visualization. Population specific models tailored to Turkish individuals can improve accuracy and reduce misclassification risk, though they must be maintained with ongoing validation across diverse Turkish sub populations and updated as new data emerge.
Conclusion
Integrating machine learning methods and deep neural networks for sternum based sex estimation in Turkish populations holds significant promise for forensic anthropology. By combining metric and geometric features with advanced algorithms, researchers can achieve more reliable sex classification while addressing population specificity and forensic practicality. Ongoing data collection, validation, and ethical implementation will be essential to translate these methods from research to routine forensic use.
