Information Technology Reference
In-Depth Information
From the clinical results shown in this section, it can be observed that the image-
based self-calibration system used for the 3D reconstruction of the spine compares
very well with models obtained with the previous system using low-level primitives
(landmarks) which were identi
ed manually by an expert operator. From a clinical
perspective, it demonstrates that the image-based self-calibration method yields
results with insigni
05) with regards to a set of standardized
3D measurements used for the clinical evaluation and the assessment of adolescent
scoliosis, without relying on manual landmark identi
cant differences (p
0
:
\
cation.
5.2 Personalized 3D Reconstruction of the Spine
5.2.1 Clinical Validation
The proposed 3D spine reconstruction method was applied to scoliotic patients
recruited at the scoliosis clinics of Sainte-Justine Hospital (Montreal, Canada). The
selection of the patients included in this group was based on the availability of the
images needed to compute 3D reconstructions of the spine, and that all patients had
12 thoracic and 5 lumbar vertebrae. Twenty pairs of biplanar X-ray images taken
from scoliotic patients with mild deformities (Cobb angle range 15
°
-
40
°
) were used
to evaluate the 2D and 3D differences of the proposed method. For each case,
comparisons between results obtained with the proposed method and those from a
radiology expert were established.
To assess the precision of the image-based similarity measure used for the
optimization procedure, Fig.
11
shows results with the retro-projection of the
deformed 3D vertebra contours (high-level primitive)
fitting adequately to the bony
edges of the corresponding vertebra in the coronal and sagittal X-ray image. The
qualitative evaluation of the global method also shows the projected anatomical
landmarks obtained from the optimized 3D model, and yield better accuracy in
terms of epipolar geometry to the 2D locations manually identi
ed by a radiology
expert on each vertebra. Figure
12
presents a box-whisker diagram with the overall
representation of differences and errors for the group of patients. The overall sum of
squared differences
(method vs. observer)
for
the selected cases was of
0.9
0.9 mm for the vertebral
contours. However the root-mean-square (RMS) epipolar geometry error (distance
of the landmarks to the epipolar line) yields signi
±
0.7 mm for the 2D point landmarks and of 1.8
±
cantly lower errors (p
\
0
:
05) for
the proposed method compared to a manual technique (1.5
3.2 mm).
The point-to-point mean difference between the 3D spine models issued from
the proposed technique and from a manual
±
1.2 vs. 4.7
±
identi
cation yielded a 3D mean
±
±
difference of 1.8
1.6 mm for thoracic
vertebra. Differences are slightly higher in the thoracic region due to extrapolation
errors and lower visibility, thus offering less image-based information on the X-ray
images.
1.5 mm for lumbar vertebra and 2.2
