Information Technology Reference
In-Depth Information
Table 2 Results on point-to-
surface comparisons with
mean, root-mean-square
(RMS) and maximum errors
of 8 scoliotic vertebrae with
3D models obtained from
MRI data
Vertebra
N
Mean (mm)
RMS (mm)
Max. (mm)
T10
1
1.0
±
0.7
1.3
±
0.4
4.2
±
0.9
0.9
±
0.8
1.3
±
0.3
3.4
±
1.0
T11
2
T12
2
1.3
±
0.9
1.8
±
0.3
4.4
±
1.2
0.9
L3 1 1.3
±
1.2 1.7
±
0.3 4.5
±
1.1
N Denotes the total number of vertebrae at different vertebral
levels
L2
2
1.1
±
1.0
1.4
±
0.2
4.3
±
5.2.2 Ground-Truth Comparison to MRI Data
To evaluate the overall accuracy of the 3D reconstruction system on real patient data,
8 vertebrae (4 lumbar and 4 thoracic) from two scoliotic patient who were scanned
with an MRI device (AVATO, Siemens Medical Solutions, Germany) were used for
comparison. Both patients also had their biplanar X-rays taken prior to surgery. Each
slice of 1 mm thickness was taken with a resolution of 256
256 pixels and 12 bits
per pixel with no interspacing. The results of the comparisons are expressed as point-
to-surface distances, i.e., each point of reconstructed vertebra is projected onto the
surface on the corresponding scanned vertebra and the point-to-surface Euclidean
distance is computed. For an appropriate clinical comparison, the same six land-
marks from the personalized 3D reconstruction (endplate midpoints, pedicle
extremities) were identi
ed on each vertebra using an interactive graphical computer
tool in order to rigidly register both 3D vertebra models.
The overall point-to-surface comparison results (mean, RMS, and maximum)
between the reconstructed 3D vertebral models issued from the proposed image-
based stereo-radiography method and from MRI scans are presented in Table
2
. The
mean point-to-surface errors are 1.2
0.8 mm
for thoracic vertebra. Visual comparisons between the 3D reconstruction using the
proposed reconstruction technique and the reference model (MRI) are presented for
thoracic vertebrae in Fig.
13
. The results show the patient-speci
±
1.1 mm for lumbar vertebra and 1.1
±
c vertebral models
obtained from stereo-radiography offers an adequate correspondence with the
ground-truth 3D representation given by MRI, speci
cally in the vertebral bodies
and pedicle regions. The validation results presented above show that the accuracy
of the statistical image-based 3D reconstruction method is comparable to ground-
truth 3D reconstructions obtained from MRI data.
5.3 Classi
cation of Spinal Deformities
Using the proposed methods for calibrating and generating a 3D spine model from
X-ray image, a cohort of 170 AIS preoperative patients with right thoracic defor-
mations was processed, classi
ed as Lenke Type-1 by members of the 3D Scoliosis
Committee of the SRS, in order to uncover potential 3D subclasses within this
