Desktop version

Home arrow Health

Bone

Shouhei Hanaoka

Vertebrae and Ribs Segmenting the vertebrae and ribs presents a special challenge. Usually human beings have seven cervical (C1-C7), 12 thoracic (T1-T12), and five lumbar (L1-L5) vertebrae, as well as 12 pairs of ribs. Therefore, an ideal spine segmentation algorithm needs to detect all vertebrae, identify the number (anatomical name) of each vertebra, and, finally, segment them. The terms “detection,” “identification,” and “segmentation” were used first by Klinder [161], who developed the first fully automatic segmentation method for all 24 vertebrae from CT volume data.

Some detection, identification, and segmentation methods are listed in Table 3.1. In the rest of this section, these three processes are separately discussed, followed by a brief review of segmentation methods for ribs.

Detection of Vertebrae The detection phase may include preprocessing to locate anchor structures such as the pelvis [47] or the spinal canal [321]. For example, Yao et al. [321] used a watershed algorithm to segment the spinal column. Another method is the generalized Hough transform (GHT) to find the spinal canal [235] or the pelvis [47].

Several strategies can be considered for detecting vertebrae. They can be detected directly by GHT [161], template matching [116], feature extraction, or machine learning-based methods [124, 327].

Another strategy is to detect the intervertebral discs. For instance, Kim et al. [154] and Hanaoka et al. [112] applied a ray-casting search algorithm, where intervertebral discs are searched as low-density structures sandwiched between two high-density endplates (the upper and lower surfaces of the vertebral body). Kim also used a fence-like deformable model to segment neighboring vertebrae precisely. Kelm et al. [193] used a marginal space learning method in which positions, rotations, and scales of the discs are hierarchically estimated. First, candidates of target object positions are estimated, followed by estimation of position-rotation, and, finally, of position-rotation-scale. They expanded this marginal space learning

Table 3.1 List of bone detection, identification, and segmentation papers

Authors

Objects of interest

Modality

Pathology

Detection

target

Detection

method

Success

ratio

№>

Identification

method

Success

ratio

(%>

Segmentation

method

Mean

distance

error

(mm)

[161]

Klinder

etal.

(2009)

Spine

C,T,L

CT

(+) incl. scoliosis

Vertebrae

GHT

92

Appearance

model

registration

95

Surface

mesh

1.12

[154]

Kim et al.

(2009)

Spine

lower T, L

CT

Discs

Ray

search

3D

deformable

fence

[47]

Buerger

etal.

(2013)

Spine

T, L

MR

(T1WI

Dixon)

Pelvis,

lungs

GHT,

mesh

model

fitting

100

Counting

from

sacrum

100

Surface

mesh

1.69

[268]

Stem

etal.

(2011)

Spine

T, L (v.

body

only)

CT MR (T2WI)

<+)

Parametric

shape

model

  • 1.17
  • (CT)
  • 1.85
  • (MR)

(Landmark

to

landmark

distances)

[52]

Carballido-

Gamio

etal.

(2004)

Spine

L (V- body only)

MR

(T1WI)

Normalized

cuts

6.225f

(6.64

voxel of

mean

error *

0.9375

mm/

voxel)

Table 3.1 (continued)

Authors

Objects of interest

Modality

Pathology

Detection

target

Detection

method

Success

ratio

№>

Identification

method

Success

ratio

(%>

Segmentation

method

Mean

distance

error

(mm)

[124]

Huang

etal.

(2009)

Spine

С, T, L (v. body only)

MR

  • (T2WI,
  • 2D)

Vertebrae

Wavelet features + modified AdaBoost

97.78

Normalized

cuts

[138]

Kadoury

etal.

(2013)

Spine

T, L

CT

MR(TIWI)

Vertebrae

Artificated deformable model and high-order

MRF (**)

The same method as

  • 2.2 (T),
  • 2.8 (L),
  • 2.9 (T), 3.0 (L)

[308]

Whitmarsh

etal.

(2013)

Spine

L2, L3

CT

Multi-atlas

0.3

[126]

Ibragimov

etal.

(2013)

Spine

L

CT

Landmark

detection

and

atlas-based

registration

0.76

[327]

Zhan

etal.

(2012)

Spine

С, T, L

MR

(scout

scan)

(+) incl. scoliosis

Vertebrae and discs

Wavelet features + AdaBoost cascade

97.7

<-)

[193]

Kelm

etal.

(2013)

Spine

T, L(CT)

C,T,L

(MR)

CT MR (T2WI)

(+)<-)

iterated

marginal

space

learning

  • 98.04
  • (CT)
  • 98.64
  • (MR)

Graph cuts

[184]

Ma et al. (2010)

Spine

T

CT

Vertebrae

Marginal

space

learning

Mean shape fitting

73 ~ 91

Surface

mesh

0.95

[116]

Hayashi

etal.

(2011)

Spine

T2-12, L (v. body only)

CT

Vertebrae

Template

matching

By rib detection

<-)

3.6-5.5

(Mean of

Hausdorff

distances

evaluated

on cross

sections)

[208]

Naegel

(2007)

Spine

T, Lf

CT

*

Vertebrae

Morphology

filtering

*

By detecting the lowest ribs

100

Watershed

algorithm

*

[112]

Hanaoka

etal.

(2011)

Spine

C3-7, T, L

CT

(-)

Vertebrae

Ray search

95

Counting from sacrum

60

Graph cuts with Rieman- nian metrics

  • 1.11(C), 1.43 (T),
  • 1 11 (L)

[194]

Mirzaalian

etal.

(2013)

Spine

C3-7f, T, L

CT

Iterated

marginal

space

learning

Surface

mesh

1.37

(Symmetric point to mesh)

[321]

Yao et al. (2006)

Spine

T, L

CT

( + ) incl.

metas-

tases

Spinal

canal,

discs

Watershed algorithm, intensity profile, etc.

97.2

Four-part deformable model fitting

(continued)

Table 3.1 (continued)

Authors

Objects of interest

Modality

Pathology

Detection

target

Detection

method

Success

ratio

(%>

Identification

method

Success

ratio

№>

Segmentation

method

Mean

distance

error

(mm)

[160]

Klinder

etal.

(2007)

Ribs

CT

Ribs

Ray

search

94.4

ICP registration of rib

centerlines

94.1

Surface

mesh

0.36

[267]

Staal

etal.

(2007)

Ribs

CT

Ribs

Ridge detection and spin- grass classifier

Heuristic

algorithms

98.4

Seeded

region

growing

[167]

Lamecker

etal.

(2004)

Pelvis

CT

Surface

mesh

1.8

[251]

Seim

etal.

(2008)

Pelvis

CT

Surface mesh + graph cuts

0.7

[320]

Yao et al. (2003)

Pelvis

CT

Flexible

mesh

template

matching

[323]

Yokota

etal.

(2009)

Pelvis

and

femur

CT

<+)

Hierarchical

SSM

1.2

*No information provided in the paper f Not clearly described; see the original paper to detect multiple discs. The method showed a detection success rate of 98% for both CT and MR images.

Identification of Vertebrae Identification of all 24 vertebrae would seem to be a very trivial task, as long as the following conditions are met: (1) the field of view of the given image includes the whole spine, (2) the target spine has the normal number of 24 vertebrae, and (3) all 24 vertebrae are successfully detected. Unfortunately, these conditions are rarely satisfied. In fact, to the best of our knowledge, the work by Klinder et al. [161] is the only one which can identify and segment all 24 vertebrae in a CT image. The method uses appearance model registration to distinguish each vertebra. The success rate in their identification method was 95%. This method successfully handled images with various field-of-view sizes, including those encompassing only a part of the spine, and with various pathologies, e.g., scoliosis.

Ma et al. [184] introduced a method employing mean shape fitting to identify and segment thoracic vertebrae. The success rate was 73% if only one vertebra was segmented, but it increased to 91% when seven continuous vertebrae were simultaneously identified.

Another method (that of Hanaoka et al. [112]) could detect 22 vertebrae, excluding C1 and C2. The success rate of identification was no more than 60%, because their identification method was based on a “counting-from-the-bottom” strategy. However, it should be noted that one of the failures was due to an anomalous number of lumbar vertebrae (the patient had six lumbar vertebrae). According to a survey by Carrino et al. [54], no fewer than 8.2% of people have an anomalous number of vertebrae.

Hanaoka et al. also introduced a whole-body landmark detection method [113] which identifies landmarks for each of the 24 vertebrae. This particular method will be discussed later.

Segmentation of the Vertebrae After detection and identification, segmentation of each vertebra must be performed. The most popular segmentation method is surface mesh fitting [47,161,184,194]. For example, Ma et al. used learning-based edge detection and a coarse-to-fine deformable surface mesh model to segment the thoracic vertebrae. Klinder et al. [161] also used a surface mesh-based statistical shape model (SSM) with an image gradient-based external energy term.

Kadoury et al. [138] recently reported a unique method in which detec- tion/identification and segmentation are simultaneously processed. They assume that global deformations of the spinal column will manifest similar local deformations of the vertebrae because of the same type of pathological deviation. Under this assumption, they modeled scoliotic spinal columns as an articulated deformable model, embedded it into a local linear embedding (LLE) manifold, and tried to represent local shape appearances of the vertebrae as a linear combination of shapes of neighbor samples on the manifold. The segmentation problem is formulated as a high-order Markov random field (MRF) and solved as a single optimization problem.

S. Hanaoka et al.

160

Stern et al., among others, modeled pathological changes in vertebral body shapes by a parametrical shape model with 25 parameters. Carballido-Gamio et al. [52] and Huang et al. [124] used a normalized cut method to segment vertebral bodies. Recently Whitmarsh et al. reported a multi-atlas-based segmentation method [308] for lumbar vertebrae with a mean distance error of no more than 0.30 mm.

Segmentation of Ribs Staal et al. [267] reported an automatic rib segmentation and labeling method. After 1D ridges are extracted from the given image, line elements are constructed, classified, and then grouped as rib centerlines. These centerlines are used in the final region-growing algorithm. They reported an identification success rate of 98.4%, excluding the first ribs. Klinder et al. [160] reported another method with an identification success rate of 94.4% and a mean distance error of 0.36 mm. Note that a rib detection method sometimes serves as an identification method for thoracic vertebrae (e.g., in [116] or [208]).

Spinal Landmark Detection In this section, a landmark detection framework for whole spinal anatomical landmarks is introduced and discussed. The framework was developed by Hanaoka et al. [113].

The framework can determine the positions of over 100 landmarks concurrently, taking spatial correlations of all landmark pairs into account. The outline of the framework is illustrated in Fig. 3.1. First, a set of landmark candidate lists is generated by sensitivity-optimized single-landmark detectors. Each landmark detector will detect its target landmark and output approximately 100 candidate positions. Then, a Markov chain Monte Carlo (MCMC)-based combinatorial optimization algorithm will find the most probable combination of candidate positions through maximum a posteriori (MAP) estimation.

Outline of the landmark detection framework (Cands = candidates)

Fig. 3.1 Outline of the landmark detection framework (Cands = candidates)

Spine with six lumbar vertebrae

Fig. 3.2 Spine with six lumbar vertebrae

The unique feature of the framework is that it can handle subjects with segmentation anomalies of the spinal column. As mentioned, no fewer than 8.2% of people have an anomalous total number of vertebrae. Such anomalies are very problematic in both defining and detecting vertebral landmarks (Fig. 3.2). To overcome this, a series of anomaly landmark position set converters is introduced. One converter can convert any landmark position set in a subject with a certain type of anomaly into a virtually normalized landmark position set (Fig. 3.3). Because of this converter, the posterior probabilities can be calculated even in subjects with an anomalous number of vertebrae. The framework can determine the type of anomaly in a given unseen image by (1) applying all anomaly type hypotheses sequentially and (2) adopting the hypothesis with the largest posterior probability.

162

S. Hanaoka et al.

Examples of spinal bone number anomalies (four or six lumbar vertebrae) and their anomaly conversion results

Fig. 3.3 Examples of spinal bone number anomalies (four or six lumbar vertebrae) and their anomaly conversion results

Landmark detection result in a subject with 13 (i.e., one more than normal) thoracic vertebral bodies

Fig. 3.4 Landmark detection result in a subject with 13 (i.e., one more than normal) thoracic vertebral bodies

Through an experiment with artificial detector outputs, the framework achieved a 97.6% success rate in anomaly type determination. The results show the potential of this framework to detect an anomalous number of vertebrae by trying several anatomical variant hypotheses sequentially (Fig.3.4).

Hip Joint The hip joint consists of the femoral head (which is a part of the femur) and the acetabulum (a part of the pelvis), with their respective articular cartilages. The acetabulum is formed by the ischium, ilium, and pubis. SSMs and related methods have been utilized for reconstruction and segmentation of the hip joint from two-dimensional (2D) and three-dimensional (3D) medical imaging data.

Main applications of early studies were on 3D shape reconstruction from sparse or incomplete data such as 3D point data acquired in the operating room [234], 2D X-ray images [320], and 3D ultrasound data [29]. These applications of the SSMs were intended to statistically interpolate and extrapolate sparse and incomplete data in order to reconstruct an approximate 3D shape when 3D CT/MR data were not available, and they are still actively studied [330]. Even when 3D CT/MR data are available, however, accurate 3D reconstruction from them is still not an easy task, and SSMs play an important role, especially for automated segmentation. Furthermore, 3D CT/MR data have more detailed information, and thus it is worth investigating more complex statistical models of the hip joint, rather than a single SSM of the pelvis or femur, to fully utilize these data. In this section, SSM-based segmentation of the hip joint from CT/MR data is reviewed.

One of the earliest works on application of SSMs to CT/MR data segmentation was done by Lamecker et al. [167]. In their study, an SSM of the single pelvis was used, and the SSM was initialized by placing the mean shape manually, which raised several issues. One issue is how the SSMs were automatically initialized, including landmark detection, for this purpose. In addition, the particular issue in the hip joint is how the consistency of the geometric relationship between the acetabulum and femoral head was maintained. Therefore, multi-structure modeling is a key problem. Especially for diseased hips, keeping the consistency becomes more difficult due to joint space narrowing as well as severe deformation. In the following paragraphs, methods for the SSM initialization are described, and then several approaches for keeping the shape consistency are discussed.

Regarding the initialization, GHT has been successfully applied by using the mean shape as the template to automatically determine the initial pose of the pelvis SSM [251]. The problem of using GHT is the trade-off between computational cost and parameter range/resolution. Another approach is to manually provide anatomical landmarks, [106, 248]. Because the landmarks are well-localized features, interoperator variability of the initial pose and shape parameters estimated from the specified landmark points is expected to be small even if the landmarks are provided manually. Therefore, more objective initialization will be possible compared with manual specification of the pose itself. Because determination of the pelvic coordinate system is clinically important, some methods [164, 324] use automatic methods specific to the pelvis to determine the coordinate system, which can be the reference frame of the SSM. Similarly, the hip joint center has been used for initialization for a small field-of-view (FOV) MR dataset [249]. As described in the previous section, initialization will be improved by using automatic landmark detection in the future.

To maintain the consistency between the acetabulum and femoral head, multiobject integrated modeling of the pelvis and femur has been investigated. The simplest integration is simply to avoid overlap of the two SSMs of the pelvis and femur during their fitting to the 3D data [139]. However, it overcomes only one of the inconsistencies (although it is still effective). A more elaborate approach is to model joint-specific motion in addition to bone-specific shapes. Both analytical and statistical approaches have been proposed for incorporating the articular motion. In the analytical approach, one composite SSM of the pelvis and femur is constructed, in which the shape and motion parameters are separately represented by assuming that the hip joint is a ball-and-socket joint, that is, three degrees of freedom rotation centered at the hip joint [140]. Therefore, variability of hip joint motion can be added by the minimum numbers of parameters without reducing the ability of shape representation by the SSM. The motion of diseased hip joints will not be a simple rotation. In the statistical approach, the combined shape of the pelvis and femur of each patient is regarded as a single shape, and one composite SSM of the pelvis and femur is constructed [323]. This composite SSM includes not only shape variations but also motion variations of the joint. Its shape and motion parameters are not separated. One drawback is that the ability of shape variability representation is reduced. Therefore, hierarchical modeling is combined so as to realize a coarse-fine fitting as shown in Fig. 3.5a. Coarse fitting is first performed using the composite pelvis and femur SSM to provide initialization for the subsequent stages, and then finer fitting is performed by using the divided SSMs to gradually increase fitting accuracy as well as to provide initialization for the next stage while keeping the consistency. Furthermore, a conditional SSM are applied for further improvement of segmentation accuracy [325]. This method was shown to be particularly effective for CT segmentation of diseased hip joints even in the presence of joint narrowing and severe deformation. Figure 3.5b shows a typical segmentation result.

Segmentation of the pelvis and femur from CT images using a hierarchical SSM (H- SSM) of the hip joint. (a) H-SSM of the hip joint. (b) Typical segmentation result

Fig. 3.5 Segmentation of the pelvis and femur from CT images using a hierarchical SSM (H- SSM) of the hip joint. (a) H-SSM of the hip joint. (b) Typical segmentation result

  • 3 Understanding Medical Images Based on Computational Anatomy Models
  • 165
 
Source
< Prev   CONTENTS   Source   Next >

Related topics