Three-Dimensional Spine Reconstruction from Radiographs - Spinal Imaging and Image Analysis - page 170

used preoperative 3D models from CT or MR images to register with 2D X-ray or
fl fluoroscopic images from gradient amplitudes [ 20 , 41 , 60 ], Fleute and Lavallee

have used statistical a priori knowledge of the 3D geometric shapes in order to
model the 2D vertebral shapes by applying point distribution models (PDM) [ 17 ].
Similar approaches introduced by Lorenz et al. [ 43 ] and Vrtovec et al. [ 63 ] have
used PDM methods from training statistical shape models, thus automatically
capturing the geometrical knowledge of the principal modes of variation to isolate
3D vertebrae from tomography images. A method proposed to use a priori
knowledge of the vertebral shape using eight morphologic descriptors of the ver-
tebral body to accurately estimate the geometrical model [ 50 ]. The obtained model
would be manually re

ned by projecting the spine s silhouette on the X-rays.

'

Inference-based optimization re nements were subsequently presented to obtain an

accurate estimate of the vertebra

s orientation and 3D locations [ 14 ]. Still, these
approaches remain highly supervised by an operator to manually identify land-
marks. Benameur et al. [ 3 ] proposed a 3D-2D registration method for vertebrae of
the scoliotic spine. In this case, the geometric knowledge of isolated normal ver-
tebrae is captured by a statistical deformable template integrating a set of admissible
deformations, and expressed by the

'

first modes of variation in a Karhunen-Loeve
expansion. However, none of these methods have attempted to integrate a statistical
model taking into account the set of admissible deformations for the whole scoliotic
spine shape. Another drawback from most methods is that each vertebra is treated
individually instead of as a whole articulated model which may include the global
3D deformation of the spine. Hence in order to account for the global geometrical
representation of scoliotic deformities, a variability model (mean and dispersion) of
the whole spine allowed increasing the accuracy of the 2D-3D registration algo-
rithm by incorporating knowledge-based inter-vertebral constraints [ 4 ]. Klinder
et al. [ 35 ] has transposed these 3D inter-vertebral transformations to accomplish the
segmentation of the spinal cord from CT-scan images. In fact, 3D spinal curve
analysis where a model of the curvature of the vertebral column describing the
relationship between vertebrae has been particularly useful for 3D medical image
analysis of the spine. Because of the intricate and tortuous 3D nature of scoliosis,
automated curved planar reformation (CPR) techniques have been presented [ 62 ]in
order to increase visualization of the deformity by transforming the orthogonal and
transverse references to a spinal coordinate system. Furthermore, CPR has been
used to assist in the spine segmentation problem using a reformed 3D spinal cen-
terline [ 33 ] or by exploiting the approximate proximity of vertebrae along the
centerline [ 18 ].

4.4 Personalized 3D Reconstruction of the Spine

Given the signi cant challenges in the reconstruction of scoliotic spines, which

exhibit high variability not only on the global pose due to changing inter-vertebral
transformations, but also within the local appearance of scoliotic vertebrae (rotation,