Information Technology Reference
In-Depth Information
to detect spinal canal, intervertebral disks and three reference regions for landmark
initialization. Final landmarks and labels are selected by Markov Random Field-
based matches of a 3-disk models. Wu et al. [
20
] proposed a method to distinguish
thoracic and lumbar vertebrae in CT images. To exploit the local context informa-
tion, e.g., if a rib is attached to the vertebra, a dictionary-based classi
cation method
is designed. Speci
cally, a cascade of simultaneous orthogonal matching pursuit
(SOMP) classi
ers are applied on 2D vertebral regions extracted from the maximum
intensity projection (MIP) images.
In general, spine detection algorithms are moving from heuristic rules/
lters-
based to learning-based approaches due to two reasons. (1) Learning-based
approaches are able to effectively extract appearance/geometry/shape characteristics
of spine anatomies. In learning-based approaches, spine detection is usually for-
mulated as a classi
cation [
13
,
16
] or regression problem [
17
], in which image and
geometry features are selected and combined to distinguish spine anatomies with
others. Thanks to the development of machine learning technologies, learning-
based algorithms are often able to
find and optimally combine low level features
that may not be easily designed by researchers to identify spine anatomies. (2)
Learning-based approaches provide the scalability for extending the same algorithm
to different imaging modalities, in which the appearance characteristics of spine
may vary dramatically. Since learning-based approaches treat spine detection as a
general classification/regression problem, they are purely data-driven and thus
transparent to highly different image appearances. Given enough training dataset of
an imaging modality, learning-based approaches are able to adapt themselves by
selecting the most distinctive features from the speci
c imaging modalities.
In real clinical settings, patients with severe diseases may appear quite frequently
and the imaging artifacts are sometimes unavoidable because of special patient
conditions, e.g., metal implants (see Fig.
1
). Thus, an auto-spine detection algo-
rithm to be deployed in real clinical environment has to be highly robust to both
imaging artifacts and spine diseases.
In this chapter, we introduce a spine detection method that is highly robust to
severe imaging artifacts and spine diseases. In principle, our method also use
machine learning technologies to capture the appearance and geometry character-
istics of spine anatomies. Similar to [
11
,
13
,
15
], i.e., low-level appearance and
high-level geometry information are combined to derive spine detection. In par-
ticular, our method leverages two unique characteristics of spine anatomies. First,
although a spine seems to be composed by a set of repetitive components (vertebrae
and discs), these components indeed have different distinctiveness. Hence, different
anatomies provide different levels of reliability and should be employed hierar-
chically in spine detection. Second, spine is a non-rigid structure, in which local
articulations exist in-between vertebrae and discs. This kind of articulation can be
quite large in the presence of certain spine diseases, e.g., scoliosis. An effective
geometry modeling should not consider vertebrae detections from scoliotic cases as
errors just because of the abnormal geometry.
Our method includes two strategies to leverage these two characteristics effec-
tively. First, instead of learning a general detector for vertebrae/discs, we use a
