Cross-Modality Vertebrae Localization and Labeling Using Learning-Based Approaches - Spinal Imaging and Image Analysis - page 296

proposed to recognize vertebral landmarks by comparing vertebral features in two
adjacent fluoroscopic frames. Booth et al. [ 5 ] designed a system to construct a 3-D

spinal image from axial MRI cross-sections. Key technologies in this system
include vertebrae detection using anatomical symmetry estimation followed by a
deformable template. Peng et al. [ 6 ] proposed to detect inter-vertebral discs in 2D
MR slices with a polynomial function-based template followed by local postpro-
cessing. Zheng et al. [ 7 ] combined congruency and hough transform to localize
lumbar vertebrae in digital video

fl

uoroscopy (DVF). Deschenes et al. [ 8 ] proposed
to segment vertebrae in digital radiographs using multi-resolution wavelets. Naegel
[ 9 ] proposed to segment the spine in CT images using mathematical morphology.
Speci

fl

cally, it consists of finding markers inside the vertebrae and computing the

watershed from markers.
Compared to other human anatomies, spinal structures has some unique attri-
butes, e.g., the repetitive appearance patterns of vertebrae and a characteristic
geometry of spinal cord. To leverage these attributes, researchers start to use model-
based approaches for spine detection. In the comprehensive spinal column
extraction system proposed by Yao et al. [ 10 ], a four-part vertebra model is
designed to separate vertebral region with surrounding anatomies in CT
images. Alomari et al. [ 11 ] proposed a vertebrae labeling method for 2D lumbar
MR scans. A two-level probabilistic model is designed to incorporate pixel-level
(appearance) and object-level (geometrical) priors. Klinder et al. [ 12 ] proposed a
method to detect and identify vertebrae in CT images, in which a set of models are
constructed to encode shape, gradient and appearance priors.
With the success of machine learning technologies in medical imaging applica-
tions, learning-based approaches gain more attentions in spine detection as
well. Schmidt et al. [ 13 ] proposed one of the

first 3D MR whole spine detection
methods. In their method, local appearance cues are learned by random trees. Ver-
tebrae are localized by combining the responses of the random trees with non-local
geometrical priors modeled by a parts-based graphical model. In Ma et al. ' s[ 14 ]

work for thoracic vertebrae identi er is

trained to detect vertebrae edges and the shape of thoracic vertebrae are learned to
identify their labels. In Kelm et al.

cation in CT images, a discriminative classi

' s[ 15 ] method, intervertebral disc detection in MR

images is formulated as a classi cation problem in a nine dimensional transforma-

tion spaces. Iterative marginal space learning is proposed to generate candidates
comprising position, orientation, and scale of the discs, which are further pruned by
an anatomical network. Huang et al. [ 16 ] proposed a statistical learning approach
based on AdaBoost algorithm to detect vertebrae centers in MR images. The
detected locations are further re s[ 17 ] work

aims to detect vertebrae in CT images using regression forest. The visible part of the
spine are

ned to fit a spine curve. Glocker et al. '

first roughly detected by a trained regression forest. Accurate localization

and identi cation of individual vertebrae is then obtained through a generative shape

and appearance model. The robustness of spine detection to pathological cases are
further improved by using a discriminative centroid classi er using local and

contextual features [ 18 ]. Major et al. [ 19 ] proposed an algorithm to label spine in
both full and partial body CT scans. They employed probabilistic boosting trees