An automatic method for landmark identification of the 3D vertebrae.

Accurate spinal landmarks are essential for registration and three-dimensional (3D) reconstruction of the vertebrae. To avoid operator variability of manual landmarks, we propose an automatic method based on geometric constraints. First, the 3D convex hull of the spine in relation to landmark contours is automatically detected. With the hull, the vertebral geometric symmetry plane and pedicle vertical plane are extracted to split the vertebrae into three geometrical components. Based on the geometric components, 38 landmarks for each vertebra are identified automatically. To evaluate the proposed approach, the manual identification by two specialists is compared with the automatic identification. The results show that the mean Euclidean distance between the automatic and manual landmarks are 1.86±1.15 and 2.07±1.06 mm, respectively. The mean Euclidean distance of two specialists is 2.13±1.22. Results demonstrate that our method is accurate and reliable enough for assisting surgeons to detect spinal landmarks. It is especially valuable for individuals who are not familiar with anatomical structures.