A Parametric Model of Cervical Spine Geometry and Posture

Abstract

Computational models of the cervical spine are useful tools for research on the biomechanics of neck injury and the evaluation of countermeasures. This report describes the development of a parametric model of cervical spine geometry that is intended to provide input to computational modeling. Two-dimensional landmark data describing the outlines of the bones of the cervical spine and important head landmarks were obtained from lateral radiographs of volunteers in a seated posture taken in a previous study. After imputation and a scaling adjustment, principal component analysis was performed on neutral-posture landmark data from 140 men and women with a wide range of age and body size. The first principal component was primarily related to spine curvature, whereas the second was associated with overall size. A regression analysis predicting principal component scores from subject covariates found significant effects of stature, age, and the ratio of sitting height to stature. Sex was not a significant predictor of principal component scores after accounting for overall body size. A three-dimensional bone shape prediction was created using bone geometry from 38 women extracted from CT studies. A principal component analysis performed on each bone level from C1 to C7 demonstrated that the dominant mode of variation was overall size for C2 but not for the other levels. A detailed examination of vertebra dimensions showed that length, width, and height of the vertebrae and the vertebral bodies at C3 through C7 were not correlated. Two methods for generating 3D geometry to match the 2D predictions were created. First, a method was developed to compute the optimal vector of principal component scores such that the side-view projection of landmarks on the 3D bone best matched the 2D targets. Second, a method based on least- squares alignment and uniform scaling was developed. The method based on principal component scores aligned to the landmarks in the 3D dataset with average root-mean-square (RMS) errors below 1 mm, and matched the mesh with average RMS errors less than 2 mm. When aligning to landmarks generate from the 2D model, RMS errors were less than 2 mm. The RMS errors for the scaling method were only slightly larger. Both the PCAR 2D model and the scaling method for generating 3D bone models were implemented in the Python language for use by other researchers.