

A computer segmentation method is developed to detect automatically the contour lines of the vertebral pedicles in digital radiographs of the scoliotic spine. These contour lines are used to identify stereocorrespondent anatomical landmarks for the 3D reconstruction of the scoliotic spine using a stereoradiographic technique. Automation of the digitization of the landmarks should reduce the variability of the actual manual digitization method and improve the 3D reconstruction precision. First, a knowledge based model is used to generate anatomical markers near the pedicle centers. These markers are then used to control a morphological segmentation method allowing the automatic detection of the pedicle contour lines. Next, an ellipsis is fitted to the pedicle contour lines and stereocorrespondent landmarks are estimated in the stereoradiograhs to finally obtain their 3D location. The preliminary results demonstrated a reduction of the 2D standard deviation for the digitization of the pedicles from 2.3±1.6 mm (manual) to 0.39±0.37 mm (automatic) and an improvement in the 3D reconstruction precision by an amelioration of the stereocorrespondence estimation (reduction of the DLT error from 42.3±3.4 mm (manual) to 1.9±2.3 mm (automatic)). The automation of the stereocorrespondent landmarks digitization using computer vision methods seem to effectively reduce the variability of the method and improve the landmark correspondence identification.