Variability and Accuracy of Spine Segmentation


Patient-specific computational models of bony anatomy taken from medical CT imaging offer significant advances for biomechanics research; these include the ability to perform non-invasive investigations of joint mechanics and interactions between medical devices and the human body using Finite Element (FE) models. This study involved testing the effectiveness of manual segmentation of lumbar vertebrae, taking into account the proximity of articulating surfaces and degenerative joint changes. Simpleware ScanIP was used to generate models suitable for measurement and analysis in MATLAB®.


  • CT scans carried out on lumbar spine segments
  • Visualisation of 3D image stacks in ScanIP and image processing to segment individual vertebrae
  • Model generation of each vertrebral surface
  • Export to MATLAB® for analysis using image registration

Thanks to

Allegheny General Hospital & Drexel University College of Medicine, Pittsburgh, PA 
D.J. Cook • D.A. Gladowski • H.N. Acuff • M.S. Yeager • B.C. Cheng

Image Processing

Eight cadaveric lumbar spine segments (T12-sacrum) were cleaned of muscle and loose connective tissue, before being CT scanned in a 64-slice machine (Somatom, Siemens, Munich, Germany) with a slice thickness of 0.6mm. The 3D image stacks were then imported to Simpleware ScanIP, where a custom window width and level, as well as a curvature anisotropic diffusion noise filter were applied. Interactive thresholding and floodfill tools generated a mask of the bony anatomy, which was then segmented into individual vertebrae and manually edited to deal with connectivity across the facet joint. Morphological closing filters were applied, before smoothing was carried out, and a model generated in STL format for export to MATLAB®.


Once exported to MATLAB®, the quality of the segmentation was tested based on two observers performing the procedure twice on each specimen. Intra-and-inter-observer differences were calculated by registering the surface models with an iterative closest point algorithm. A map of the nearest neighbour distances between each model was created for each vertebra, and the distribution of their distances analysed. Five vertebrae were then disarticulated from their neighbours, cleaned of all soft tissue, rescanned, and segmented without adjacent vertebrae and soft tissue to evaluate the original segmentation routine.


Analysis of the segmentation indicated root-mean-square (RMS) errors, or the distance between neighbouring points on registered surfaces, below 0.39 mm for accuracy based on comparisons between models before and after disarticulation. The 95th percentile of distances was sub-millimetre in all instances, as were all RMS measurements from comparison of models prior to disarticulation. The test therefore demonstrated the accuracy of models generated from CT images of the lumbar spine in Simpleware software, indicating the potential of future studies involving patient-specific modelling of the lumbar spine.