Poster R6 - A Robust Method for Automated Landmark Detection and Measurements on 3D Spine Surface Models

Introduction: The spine is a multisegmented structure in a multiplanar orientation which inherently rotates about all three anatomical planes. Consequently, viewing the spine in conventional radiographic imaging as a projection or in a global plane of a 3-dimensional (3D) reconstruction is insufficient to fully capture the spine’s complex geometry and the morphological changes found in spinal deformities such as scoliosis which involve simultaneous lateral bends, axial rotation, and sagittal curvatures. To remove measurement errors due to geometric distortions, such as parallax caused by using a global plane, and intra- and inter-observer variability found in manual methods, the objective was to establish a method to automatically identify local coordinate systems, landmark points, and morphological measurements for vertebrae and intervertebral disc (IVD) spaces of 3D spine models.

Materials and

Methods: The proposed 3D spinal morphology analysis method must apply to medical conditions that alter vertebral and IVD shape. Thus, retrospective data of a porcine model of progressive scoliosis from a collaborating institution was used to validate the robustness of the method. 3D spine reconstructions with removed posterior elements were made using 3D Slicer software from serial computed tomography (CT) scans of three Yorkshire piglets over a 16-week period of induced scoliosis progression which resulted in spinal curvatures with cobb angles ranging from 54 to 65 degrees. Using custom MATLAB code, a novel automated technique of iteratively refining temporary coordinate systems was used to identify endplate surfaces. From these surfaces, local coordinate systems were then defined for each vertebral body with the following definitions: the origin at the midpoint between the endplate centroids, the inferior-superior axis aligned through these centroids, and posterior-anterior and right-left axes derived from averaging the axes of endplate projections obtained from principal component analysis on the local transverse plane (Figure 1). A similar approach was taken to define local coordinate systems for the IVD spaces. On the resulting local axes, landmark points were automatically identified (Figure 2) to calculate coronal and sagittal vertebral wedging angles and widths as well as anterior, posterior, concave, and convex vertebral heights.

Results, Conclusions, and Discussions: Using only vertebral body geometry, the proposed 3D spinal morphology analysis method successfully led to obtaining vertebral endplate surfaces, local coordinate systems, landmark points, and morphological measurements for the thoracolumbar spines of all three piglets across all seven time points of scoliosis progression. Immediate next steps to ensure the method's applicability and reliability will involve applying this method to both normal and scoliotic human spines, and validating the accuracy of the automatically identified landmark points against those placed manually by an expert. Facilitated by the advent of low-radiation-dose biplanar radiography technology, such as EOS imaging, the use of 3D reconstructed spine models is becoming more commonplace in studies of spine deformity and kinematics. The described robust method may be used to standardize such studies going forward with repeatable, objective landmark points and morphological measurements.

Acknowledgements (Optional): We would like to thank the Wyss-Campbell Center for Thoracic Insufficiency Syndrome at the Children’s Hospital of Philadelphia for partially funding this study.