Development of a sex estimation method for skulls using machine learning on three-dimensional shapes of skulls and skull parts

Objectives Sex estimation from the skull plays an important role in the identification of skeletal remains. A novel sex estimation method from three-dimensional shapes of the skull using machine learning technology is presented. Materials and Methods A total of 100 skull shapes were obtained from post-mortem computed tomography data. Homologous models of the whole skull, the cranium only, and the mandible only were created and the coordinates of 40,962 vertices of the models were reduced in dimensionality by principal component analysis (PCA) and partial least squares regression (PLS). Known sexes and the scores of the obtained components were supplied to the support vector machine, and the accuracy rates of sex estimation were obtained by a 10-fold double-looped cross-validation procedure., Lastly, the developed method was applied to six casework skulls with known sexes to validate the power of the estimation. Results and Conclusion By reducing dimensionality with PLS, remarkably high accuracy rates were obtained in the estimation. The rates reached 100% in all parts examined, while the rates of parts reduced with PCA remained around 90%. Virtual shapes created from very large and small scores of the first and second components of PLS showed the clear sexual dimorphisms that have been proposed by many researchers. The validation results on actual casework skulls were less acceptable than expected, suggesting the necessity of further study with greater numbers of samples for machine learning. The sex estimation method developed here enables us to perform objective identification of skeletal remains required in forensic anthropology field.