An Automated Method for Geometric Reconstruction of Vertebrae from

Clinical CT Scans

Yifei Dai10/19/2006

Introduction• Finite element models of biomechanics of spinal

fusion– Stress shielding in stabilization Goel et al 1988– Effect of bone graft characteristics Zander et al 2002– Stress investigation of fusion systems Adam et al 2003

• Generic model used to provide insight into global trends

• Cross-sectional experimental studies are used to investigate surgical techniques and devices Boden et al 1995, Erulkar et al 2001, Sandhu et al 1996, Hojo et al 2005

• Models could provide means to perform longitudinal studies– In vivo µ-CT or clinical CT scanners can provide data– Use subject-specific meshes to simulate behavior

Mesh generation• CT is the most common source for generation of

finite element meshes– Must be converted to geometric format compatible with

analysis software

• Voxel based meshes– Neglect or approximate apophyseal joints Weinans et al

2000, Crawford et al 2003 2004, Adam et al 2003

– Large number of elements are required to ensure accuracy Crawford et al 2003

• Geometric generation techniques – Border tracing Testi et al 2001

– Marching cubes Viceconti et al 1998, 1999

Objectives

• Fast and accurate mesh generation is needed for application to clinical evaluation or longitudinal studies

• Objective: develop a semi-automated technique to obtain geometric models of lumbar vertebrae from clinical CT scans– Develop an iterative curve fitting technique– Construct least-squares B-spline contours of one vertebra

from clinical CT scans – Develop an atlas based semi-automated geometry

reconstruction technique– Apply technique to reconstruct a different vertebra

Methods• CT scan data of vertebra• Image-processing:

– Smoothing with 5×5 median filter – Edge detection Canny 1986

– Remove spurious edges

Least-squares B-spline fitting

• Parametric coordinates from closest projection to initial curve

• Edge ordering • B-spline fitting results in

smooth curve • Convergence based on

movement of control points

• Implemented in Matlab on1.4 GHZ AMD Opteron workstation

Iterative fitting

Iteration 1 Iteration 2 Iteration 3

• Edge ordering– Sort points in parametric coordinates

– Points that are out of order in Euclidean space are excluded from fit

• Estimate of parametric coordinates improves with iteration

Definition of initial atlas curves

• Defined from CT scan of an L2 vertebra– Initial approximating curves defined manually

– 5 curves representing different geometric features from vertebra

– 3 simple geometric shapes

Atlas based fitting

• Interactively choose initial curve with appropriate geometric features

• Reorient initial curve to edge points using the axes of the 2-D inertia tensor

• Scale based on a minimax bounding box• Applied to CT images of an L3

vertebra from a different subject

ResultsAtlas definition

Maximum of six iterations were requiredMaximum of 10 min user time were needed for one slice

Atlas based fitting

Discussion

• A semi-automated method was developed to construct geometric models from clinical CT scans– The technique could be further automated to select appropriate

initial approximating curves from atlas using artificial intelligence

• Robust to missing data– Missing edges are interpolated

Discussion

• Automatically and objectively smooth noisy or pixilated edge data– Fidelity to original data determined by number of B-spline points

Smoothness of fit• Curvature fluctuation determine by number

of control points

Discussion

• Stacked contours can provide geometric

input for generation of finite element models

• Meshes of atlas shapes in parametric space could be distorted to automatically generate meshes

• Can be applied to a variety of orthopaedic structures by construction of a corresponding atlas

Dr. Glen L. Niebur

Memorial Lighthouse Medical Imaging Center, Mishawaka, IN

Center for Disease Control CE000789

Acknowledgements