Information Technology Reference
In-Depth Information
Fig. 5 An example of
“
over-discriminative
”
vertebrae detectors.
“
Filled circle
”
and
“
Plus sign
”
denote the highest response of T
9
and T
10
detectors, respectively
or T
9
-
T
11
. At run-time, bundle vertebrae detectors are expected to return multiple
peak responses at vertebrae centers belong to the bundle. The locations of these
responses will be further veri
ed based on anchor vertebrae and the local articulated
spine model (see Sect.
5
). The speci
c vertebrae labels are assigned in the same way.
4.4 Inter-vertebral Discs
To detect an inter-vertebral disc, not only its center but also its orientation and size
should be determined. Compared to vertebrae center detection (3-D hypothesis
space), disc detection has a higher dimensional hypothesis space with 9 spatial
parameters. In [
15
], this problem is tackled by marginal space learning (MSL),
which detect the center, orientation and sequentially. In principle, MSL treats the
inter-vertebral disc as a whole and aims to sequentially determine the spatial
parameters through sub-space learning. The sequential nature, however, might
in
uence the robustness of the algorithm. For example, a gross error of disc center
detection may not be corrected by the following orientation/size detection.
Instead of treating the disc as a whole, we propose to formulate the disc detection
as a voxel-wise classi
fl
cally, for each voxel, the disc detector
aims to predict the likelihood of this voxel belonging to the inter-vertebral
disc. Therefore, in the training stage, each voxel is considered as an individual
training sample. On-disc and off-disc voxels are used as positives and negatives,
respectively. At runtime, the learned disc detector will derive a response map. The 9
spatial parameters of the disc is then derived by
cation problem. Speci
fitting disc response maps with an
elliptical cylinder using principal component analysis. This strategy brings
robustness in twofold. First, since each voxel is labelled independently, the clas-
si
uence any others. Second, since principal
component analysis is robust to outliers, sporadic classi
cation errors of one voxel will not in
fl
cation errors at voxel-level
will not dramatically change the derived disc positions.