Ales Neubert1,2, Jurgen Fripp1, Kaikai Shen1, Craig Engstrom2, Raphael Schwarz3, Lars Lauer3, Olivier Salvado1, Stuart Crozier2
1The Australian E-Health Research Centre, CSIRO, Brisbane, QLD, Australia; 2Department of Biomedical Engineering, University of Queensland, Brisbane, QLD, Australia; 3Siemens Healthcare, Erlangen, Germany
Automated segmentation of lumbar spine vertebrae and intervertebral discs is presented. Statistical shape models are built from a training database and applied to segment new cases by matching gray level profiles along normals to shape vertices. Automated initialisation is done by computing 3D spine curve and axial vertebral symmetries. Intensity profile along the spine curve is analysed to detect centres of vertebral bodies where the mean shape is placed. Approach was tested on a set of axial T1w and 3D Space T2w images with mean Dice scores of 0.85 (vertebrae) and 0.83 (discs).