Information Technology Reference
In-Depth Information
4 Experimental Results
4.1 Experimental Design
We validated the present method on a single lateral
fluoroscopic image of one
cadaveric lumbar spine segment (there are totally four vertebrae in this segment but
only three of them are visible in the image). All the binary volumes of the lumbar
vertebrae contained in the test spine segment were semi-automatically segmented
from the associated CT datasets using the commercially available software package
Amira 5.0 (TGS Europe, Paris, France). To evaluate the reconstruction accuracy, a
surface model derived from the binary volume of each test lumbar vertebra was
used as the ground truth. As we only reconstructed a scaled surface model of the
lumbar vertebra from each lateral
fl
rst recover the
unknown scale factor of the reconstructed model with respect to the ground truth
surface model (or vice versa) before we could evaluate the reconstruction accuracy.
For this purpose, we proposed to estimate the unknown scale factor of the recon-
structed models by performing surface-based registrations [
30
]. After the registra-
tion, the open source tool MESH [
31
] was used to compute the distances between
the reconstructed surface models and their associated ground truth surface models,
which were regarded as the reconstruction errors. We adapted this tool to include
the computation of different error statistics.
Two studies were conducted to evaluate the robustness and the accuracy of the
present technique. Due to the reason that all 4 surface models of the lumbar ver-
tebrae contained in the test spine segment were part of the training database for
constructing the PDM, we named the
fl
fluoroscopic image, we had to
first study as the leave-all-in study. In this
study, each time the PDM as described in Sect.
2
was used together with the single
lateral
fluoroscopic image of the test spine segment to reconstruct a scaled surface
model of a test vertebra. In the second study, all 4 aligned training surface models
corresponding to the lumbar vertebrae in the test spine segment were removed from
the training database and a PDM constructed from the rest 35 training surface
models was used to reconstruct a scaled surface model of each test vertebra. We
thus called the second study as the leave-four-out study. In both studies, two
different surface-based registration techniques, i.e., a surface-based anisotropically-
scaled rigid registration and a surface-based isotropically-scaled rigid registration,
are used to estimate the unknown scale factors between the reconstructed surface
models and the associated ground truth surface models.
For all experiments, we used an Intel Duo Core 2.4-GHz laptop with 4 GB of
RAM. All programming was done using Visual C++ 2005 on Windows Vista.
fl
