Cross-Modality Vertebrae Localization and Labeling Using Learning-Based Approaches - Spinal Imaging and Image Analysis - page 305

It is worth noting that this learning scheme follows the same trend from anchor
to bundle vertebrae. As previously discussed, neighboring vertebrae are ”

since they are not easy to be distinguished. Hence, the bundle vertebrae detectors
are not as distinctive as anchor vertebrae detectors and have multiple peaky
responses at runtime. Since voxels on the same disc are even more indistinguish-
able, all of them are

“ bundled

in the training stage. Hence, the learned detectors
becomes even less distinctive and expected to have high responses at any voxels
located on the disc.
To further enhance the robustness, we leverage the information from vertebrae
centers in learning disc detectors. Speci

“ bundled ”

cally, the training images of a disc are
aligned by the line connecting its two neighboring vertebrae centers. Since the inter-
vertebral discs are roughly perpendicular to the line connecting its, this alignment
will effectively remove the variations of the appearance features resulting from
different disc orientations. (Note that Haar-like features are not rotation invariant.)
At runtime, based on the previously detected vertebrae centers, the subject image is
aligned in the same way before applying the disc detectors.

4.5 Summary

The differences in training anchor vertebrae, bundle vertebrae and inter-vertebral
discs detectors primarily exist in the selection of positive/negative samples and
image alignment before feature extraction (see Table 1 ). Moving down the table
from anchor vertebrae to inter-vertebral discs, as the targeted anatomies become
less and less distinctive, more positive samples are extracted, i.e., voxels around the
center of a speci c anchor vertebra ! voxels around centers of several neighboring

vertebrae c inter-vertebral disc. This results in less

discriminative detectors that are expected to return high responses at more voxels.
On the other hand, the image alignment becomes more and more sophisticated.

! any voxels located at a speci

Table 1 Training scheme of detectors for anchor vertebrae, bundle vertebrae and inter-vertebral
discs
Detector Positive samples Negative samples Image alignment

Anchor
vertebrae

Voxels close to the
center of the spe-
ci c vertebrae

Remaining voxels in the
entire volume image

No alignment

Bundle
vertebrae

Voxels close to the
centers of any
vertebrae within the
bundle

Remaining voxels in the
local volume image
covering neighboring
bundles

Aligned by anchor
vertebrae

Inter-vertebral
Discs

Voxels located on
the disc

Remaining voxels in the
local volume image cover-
ing the two neighboring
vertebrae

Aligned by two
neighboring
vertebrae