Você está na página 1de 10

This copy is for personal use only. To order printed copies, contact reprints@rsna.

org
Imaging

Vertebral Body Compression


Fractures and Bone Density:
Original Research  n  Musculoskeletal

Automated Detection and


Classification on CT Images1
Joseph E. Burns, MD, PhD
Purpose: To create and validate a computer system with which to
Jianhua Yao, PhD
detect, localize, and classify compression fractures and
Ronald M. Summers, MD, PhD
measure bone density of thoracic and lumbar vertebral
bodies on computed tomographic (CT) images.

Materials and Institutional review board approval was obtained, and in-
Methods: formed consent was waived in this HIPAA-compliant ret-
rospective study. A CT study set of 150 patients (mean
age, 73 years; age range, 55–96 years; 92 women, 58
men) with (n = 75) and without (n = 75) compression
fractures was assembled. All case patients were age and
sex matched with control subjects. A total of 210 thoracic
and lumbar vertebrae showed compression fractures and
were electronically marked and classified by a radiolo-
gist. Prototype fully automated spinal segmentation and
fracture detection software were then used to analyze the
study set. System performance was evaluated with free-
response receiver operating characteristic analysis.

Results: Sensitivity for detection or localization of compression


fractures was 95.7% (201 of 210; 95% confidence interval
[CI]: 87.0%, 98.9%), with a false-positive rate of 0.29 per
patient. Additionally, sensitivity was 98.7% and specificity
was 77.3% at case-based receiver operating characteris-
tic curve analysis. Accuracy for classification by Genant
type (anterior, middle, or posterior height loss) was 0.95
(107 of 113; 95% CI: 0.89, 0.98), with weighted k of
0.90 (95% CI: 0.81, 0.99). Accuracy for categorization
by Genant height loss grade was 0.68 (77 of 113; 95%
CI: 0.59, 0.76), with a weighted k of 0.59 (95% CI: 0.47,
0.71). The average bone attenuation for T12-L4 vertebrae
was 146 HU 6 29 (standard deviation) in case patients
and 173 HU 6 42 in control patients; this difference was
statistically significant (P , .001).

1
 From the Department of Radiological Sciences, University Conclusion: An automated machine learning computer system was
of California-Irvine School of Medicine, Orange, Calif created to detect, anatomically localize, and categorize
(J.E.B.); and Imaging Biomarkers and Computer-Aided vertebral compression fractures at high sensitivity and
Detection Laboratory, Radiology and Imaging Sciences, with a low false-positive rate, as well as to calculate verte-
National Institutes of Health Clinical Center, Building 10, bral bone density, on CT images.
Room 1C224, MSC1182, Bethesda, MD 20892-1182 (J.Y.,
R.M.S.). Received September 8, 2016; revision requested q
 RSNA, 2017
November 4; revision received December 5; accepted
December 21; final version accepted January 12, 2017.
Address correspondence to R.M.S. (e-mail: rms@nih. Online supplemental material is available for this article.
gov).

Supported by the Intramural Research Programs of the Na-


tional Institutes of Health Clinical Center (1Z01 CL040004).

q
 RSNA, 2017

788 radiology.rsna.org  n  Radiology: Volume 284: Number 3—September 2017


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

T
he societal burden of spine com- risk is additionally linked to compres- developed in the 1990s to assess
pression fractures is substan- sion morphology, lending importance vertebral compression deformities
tial, with an estimated 700 000– to determination of shape type (11). on two-dimensional radiographic im-
750 000 fractures per year in the Finally, increased risk for vertebral ages (3). We have created a system
United States (1–3). Compression fractures is independently linked to that updates the schema by using the
fractures are reportedly underdiag- osteoporosis, with the combined risk full power of three-dimensional (3D)
nosed due to clinically silent fractures, factors of osteoporosis and prevalent computed tomographic (CT) images
nonspecific back pain, and unreported fractures acting as risk multipliers with lateralization of the compression
fractures (4–7). Early diagnosis is im- (20–23). Early diagnosis enables one fracture patterns. When coupled with
portant, as partial compression of one to initiate administration of bisphos- metrics such as bone density, this
vertebral body portends increased risk phonates and selective estrogen re- system has potential for use in pro-
for progressive fracturing of that ver- ceptor modulators to decrease the risk spective clinical trials, in the gener-
tebra, compression fracturing of one of subsequent fractures, versus man- ation of risk factors for progressive
vertebra positively correlates with agement with an intervention such as vertebral body compression fractur-
subsequent compression fracturing of kyphoplasty (7,17–19). A computer ing, and in the generation of statis-
other vertebrae, and vertebral com- system that could be used to sensi- tical profiles for patients who may
pression increases the risk profile for tively detect compression, classify the benefit from intervention rather than
hip fractures (3,5,8–12). Mild com- degree of height loss and pattern of conservative care.
pression fractures increase patient compression, and assess bone mineral The purpose of this study was to
risk for new fractures, even as de- density would be useful in clinical de- create and validate a computer system
tection sensitivity is lower and inter- tection and classification. This system with which to detect, localize, and
observer variability is higher, lending could be integrated into an opportunis- classify compression fractures and to
importance to sensitive and uniform tic detection framework to assess CT calculate bone density of thoracic and
detection (6,13–16). Furthermore, in- images obtained for indications other lumbar vertebral bodies on CT images.
creasing severity of prevalent fractures than vertebral fracture. It could also be
increases risk for incident fractures; used in research studies on risk pre-
thus, severity grading is important for diction in clinical decision making. Materials and Methods
risk stratification (8,10). Assessment The widely used semiquantita-
of severity is also important, as com- tive Genant classification system was Study Subjects
pression of more than 50% increases Our study was compliant with the
the risk of segmental instability, and Implications for Patient Care Health Insurance Portability and Ac-
compression of more than 80%–90% nn This proof-of-function computer countability Act and was conducted
may obviate vertebroplasty and ky- system depicts compression with institutional review board ap-
phoplasty (2,17–19). Incident fracture fractures of the thoracic and proval. Informed consent was waived
lumbar vertebrae with a high for this retrospective study in which we
Advances in Knowledge degree of sensitivity, devoid of analyzed previously obtained images.
nn A fully automated machine interobserver variability, poten-
learning software system with tially enabling early detection https://doi.org/10.1148/radiol.2017162100
which to detect, localize, and and treatment to mitigate the
Content code:
classify compression fractures risk of fracture progression or
and determine the bone density additional fractures. Radiology 2017; 284:788–797
of thoracic and lumbar vertebral nn This system enables one to dis- Abbreviations:
bodies on CT images was devel- cern the severity and pattern of CI = confidence interval
oped and validated. compression of the fractured ver- FROC = free-response receiver operating characteristic
3D = three-dimensional
nn The computer system has a sen- tebra and bone density, and it
sitivity of 95.7% in the detection has potential for use in risk strat- Author contributions:
of compression fractures and in ification for further fracturing to Guarantor of integrity of entire study, J.E.B.; study
the localization of these fractures help guide patient care. concepts/study design or data acquisition or data
analysis/interpretation, all authors; manuscript drafting or
to the correct vertebrae, with a nn Potential benefits of this system manuscript revision for important intellectual content, all
false-positive rate of 0.29 per include automated classification authors; approval of final version of submitted manuscript,
patient. of fractures with an extended all authors; agrees to ensure any questions related to the
nn The accuracy of this computer three-dimensional Genant work are appropriately resolved, all authors; literature
system in fracture classification classification schema and stan- research, J.E.B.; clinical studies, J.E.B.; statistical analysis,
J.Y.; and manuscript editing, all authors
by Genant type was 95% dardized reporting of compres-
(weighted k = 0.90). sion fractures. Conflicts of interest are listed at the end of this article.

Radiology: Volume 284: Number 3—September 2017  n  radiology.rsna.org 789


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

Reports from CT examinations Figure 1


performed with the spine protocol
were drawn from the imaging data-
base of the University of California-Ir-
vine medical center by using the radi-
ology information system application
Radiology Report Search (RadNet;
Cerner Millennium, North Kansas
City, Mo). The date of service for ex-
aminations selected ranged from 2012
to 2015. Figure E1 (online) shows a
Standards for Reporting Diagnostic
Accuracy Studies chart for the report
search method.
Reports dictated as positive for
thoracic or lumbar compression
fractures within the search region
from the T1 to L5 vertebrae were
flagged, and associated CT images
were reviewed with an Impax pic-
ture archiving and communications
system, (AGFA, Mortsel, Belgium).
During this review, exclusion criteria
(Fig E1 [online]) for case selection
Figure 1:  Axial CT image of a vertebra shows zone division of the vertebral
were applied. A case set of 75 consec-
body. The axial section of the vertebral bodies was divided into zones for
utive patients with one or more com- determination of subsection of height loss. This zone division was designed to
pression fractures was assembled. conform to the original Genant classification schema, with allowance for lateral-
Additionally, images in 75 patients ization and development of an enhanced Genant classification schema. The two
without fractures who were matched lateral zones created by the sagittal plane partitions are defined as R (right) and
by age and sex to patients in the case L (left). Three zones were created by partitioning the central sagittal region into
set were assembled to constitute the three central coronal zones A (anterior), C (central), and P (posterior).
control set. The mean age of patients
was 73 years 6 11 (standard devia-
tion), with an age range of 55–96 approximate centroid of each of the data set for percentage height loss,
years. There were 46 women and 29 210 thoracic and lumbar vertebrae that the reference point used was the near-
men in both the patient group and the showed vertebral body compression est normal-appearing vertebra, which
group of age- and sex-matched control fracture deformities was electronically was chosen with qualitative visual
subjects. In the female cohort, the marked and recorded (J.E.B., a board- inspection.
mean age was 76 years 6 10 (range, certified radiologist with 10 years of The axial profile of each verte-
56–92 years). In the male cohort, the experience). bral body was divided into five zones
mean age was 69 years 6 11 (range, Additionally, an enhanced Genant (Fig 1). This zone division was designed
55–96 years). The age difference be- classification of vertebral body com- to conform to the original Genant
tween the male and female cohorts pression fractures by type (wedge, classification schema, with allowances
was significant (P , .001). concave, or crush) and grade of height for lateralization and development of an
loss (grade 1, ,25%; grade 2, 26%– enhanced Genant classification.
Compression Fracture Identification and 40%; grade 3, .40%) was performed The reference standard for system
Description for 113 vertebrae in the case set. performance was a manually annotated
CT images from the picture archiving Classification was restricted to stud- data set of compression deformities.
and communication system were saved ies with no more than two contiguous Percentage height loss for each zone
with soft-tissue reconstruction kernel fracture levels to mitigate the percent- was manually determined by comparing
in uncompressed Digital Imaging and age height loss scaling inaccuracies in it with the same zone on a neighboring
Communications in Medicine format. the assessment of patients with multi- uncompressed vertebra by using stan-
Section thickness of the images ranged ple contiguous compressed levels due dard picture archiving and communica-
from 1.0 to 2.0 mm, and in-plane res- to the lack of a standardized uniform tion system measurement ratio tools on
olution ranged from 0.31 to 0.45 mm. anatomically proximate comparator. sagittal images (Table 1). Genant mor-
In the 75-patient CT case set, the In creating the reference standard phology classification was determined

790 radiology.rsna.org  n Radiology: Volume 284: Number 3—September 2017


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

Table 1 Figure 2
Statistics of Genant Grades
Grade Zone C Zone A Zone P Zone L Zone R

0 26 38 94 67 67
1 16 18 8 14 11
2 30 31 10 20 23
3 41 26 1 12 12

Note.—Data are number of vertebrae. Genant grades of


height loss are as follows: grade 0, less than 15% height
loss; grade 1, 15%–25% height loss; grade 2, 26%–
40% height loss; and grade 3 more than 40% height
loss. Lateral zones are created by the right (R) and left
(L) sagittal plane partitions, and the central sagittal
region was partitioned into three coronal zones: A
(anterior), C (central), and P (posterior). See Figure 1 for
a map of zone locations.

by manual image review of central sag-


ittal subzones.

Quantitative Image Analysis Method


Automated segmentation of the spine
and vertebral body partitioning was per-
formed first, with detection of superior
and inferior endplates of each vertebra
(Fig 2a) (24). A height compass (height
distribution across the vertebral body)
was computed by partitioning the axial Figure 2:  Height compass and compression fracture detection and characterization. The geometric
cross-section of a vertebral body into arrangement of the height compass is a compasslike structure, with a central circular sector surrounded by
17 sectors (Fig 2c–2e). Uniform and two ring-shaped finite thickness concentric bands. Each band is divided by eight radii of common central
systematic averaging of sectors in the angles into eight sectors. All images were obtained in an 86-year-old woman with compression fractures.
three vertebrae cranial to and in the A, Sagittal CT section shows vertebral column segmentation and partitioning. B, Stacked height compass of
three vertebrae caudal to the level of the entire vertebral column. C, Height compass of a grade 2 concave fracture at T3. D, Height compass of a
grade 3 wedge fracture at T7. E, Height compass of a normal vertebral body at L2.
interest was used to determine relative
height loss. Features such as mean ver-
tebral height, height relative to neigh- The characteristics of the fractures are partitioning, the cortex was removed
boring vertebrae, and bone density can then graded by using a support vector from the underlying trabecular bone
be assigned to each sector or zone. regression technique, which is a super- via a process of adaptive erosion by
The height compass can be stacked to- vised machine learning technique that up to 5 mm of the vertebral body pe-
gether to form a global view of height associates vertebral height distribution riphery. Bone density was then esti-
distribution within the entire spinal col- with the fracture grades and types. mated by averaging the attenuation
umn (Fig 2b). The pattern of the height Fully automated bone densitome- (in Hounsfield units) within the re-
compass (17 sectors, with three fea- try computation was then performed maining 3D medullary space (26,27).
tures for each sector) was analyzed and for T12-L4 vertebrae. Bone density Analysis of each case lasted approx-
used to differentiate between fractured algorithms had an origin in previ- imately 5 minutes and was performed
and normal vertebrae. Fifty-one fea- ous software developed to determine with a high-performance desktop com-
tures from these 17 sectors were con- bone mineral density in the spine at puter with a dual 2.30-GHz central pro-
sidered, and all were used in our model. CT (26). Vertebrae marked positive cessing unit, 16.0-GB memory, and a
The technical aspects of this technique with the fracture-detection algorithms 64-bit Windows 7 operating system (Dell
are discussed in more detail in the lit- were excluded from densitometry Precision T7600; Dell, Round Rock, Tex).
erature (25). The sector resolution of calculations because of alterations in The false-positive and false-negative
the system is then decreased to match native bone density inherent in the compression fracture detections were
the simpler enhanced Genant system compressive process. With previously reviewed and were decided in consensus
classification five-zone model (Fig 1). computed vertebral segmentation and by two board-certified fellowship-trained

Radiology: Volume 284: Number 3—September 2017  n  radiology.rsna.org 791


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

Figure 3 Figure 4

Figure 3:  FROC curve of support vector machine performance in the detection of vertebral bodies with
compression fractures shows 95.7% sensitivity (95% CI: 87%, 98.9%) with a false-positive rate of 0.29 per
patient. Error bars represent 95% CIs.

radiologists (J.E.B., R.M.S.; the latter body, 19 had a fracture in two verte-
had 19 years of experience). bral bodies, 13 had fractures in three
vertebral bodies, four had fractures
Statistical Analyses in four vertebral bodies, and 14 had
System performance for the CT data set fractures in five or more vertebral
was gauged by using free-response re- bodies for a total of 210 fractured
ceiver operating characteristic (FROC) thoracic and lumbar vertebral bodies
Figure 4:  Sagittal CT section shows false-nega-
curve analysis with 10-fold cross val- with more than 10% height loss in
tive findings in compression fracture detection with
idation. When we varied the support the case set. An average of 2.8 ver- the computer system in a 77-year-old male patient.
vector machine probability threshold tebral body compression fracture de- Missed concave fractures (arrows) are visible at T8
in the determination of compression formities per patient and a range of and T11.
fracture versus a nonfractured verte- one to nine compressed vertebrae per
bra, it resulted in the generation of the patient were identified in the case set.
FROC curve. Confidence intervals (CIs) A total of 1275 vertebrae were evalu- Figure 4 shows examples of false-
were obtained by resampling the test ated with the computer system in the negative findings in compression
set 1000 times (bootstrapping) and av- combined set of 150 CT examinations fracture detection. Examples of true-
eraging the performance. (case group, n = 75; control group, n positive findings are shown in Figure 5
The target computer-aided detection = 75). and Figure E2 (online). Examples of
sensitivity in compression fracture de- Sensitivity in the detection of ver- false-positive findings are shown in
tection was set at 90% or greater, which tebrae with compression fracture de- Figure E3 (online). A total of 43 false-
was in balance with a clinically reason- formities was 95.7% relative to the ref- positive findings were generated with
able false-positive rate, when we chose erence standard (201 of 210 vertebrae, the system, of which nine (21%) were
the FROC analysis operating point. 95% CI: 87.0%, 98.9%), with a false- due to vertebra fusion, nine (21%)
positive rate of 0.29 per patient (Fig 3). were due to degenerative height loss,
Additionally, at case-based receiver op- six (14%) were due to an L5 vertebra
Results erating characteristic curve analysis, variant, four (9%) were due to end-
Of the 75 patients in the case set, sensitivity was 98.7% and specificity plate deformation, four (9%) were due
25 had a fracture in one vertebral was 77.3%. to partitioning error, three (7%) were

792 radiology.rsna.org  n Radiology: Volume 284: Number 3—September 2017


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

due to osteopenia, and one (2%) was fracture) with the system relative to morphology was not included in the
due to Scheuerman disease. Seven radiologist classification (Table 2). curves because of small statistics.
(16%) of the initially categorized false- The calculated weighted k coeffi- System performance was also deter-
positive findings were found to be mild cient was 0.90 (95% CI: 0.81, 0.99), mined relative to the manual annotation
compression fractures at secondary which was consistent with the high- data set for the severity grade of com-
review; thus, these were true-positive est level agreement category (ie, very pression deformities. There was 68%
detections omitted from the reference good agreement). FROC curves for overall agreement (77 of 113 vertebrae,
standard data set. compression morphology are shown 95% CI: 0.59, 0.76) for categorization
Fracture detection in the case set in Figure E4 (online); crush-type of compression severity (grade 1, 2, or 3)
showed a bimodal distribution, with
local peaks in the number of fractures
at the L1 and T7 levels. The largest Figure 5
distribution of fractures was about the
thoracolumbar junction (Fig 6).
The most common morphologic
compression type found in the case set
was the wedge-type anterior compres-
sion fracture. There were 80 wedge-
type fractures, three crush fractures,
and 30 biconcave fractures. The most
common grade of fracture was grade
3 (.40% height loss). There were 35
grade 1, 31 grade 2, and 47 grade 3
fractures detected with the system.
System performance was deter-
mined relative to the manual annota-
tion data set for the morphology of
compression deformities. There was Figure 5:  Sagittal CT sections show examples of compression fracture grading with the computer system.
95% overall agreement (107 of 113 These are true-positive findings. (a) Image shows the T6 vertebra in an 86-year-old female patient. The
vertebrae, 95% CI: 0.89, 0.98) for radiologist grade was a grade 3 wedge fracture; the computer grade also was a grade 3 wedge fracture. (b)
categorization of compression mor- Image shows the T10 vertebra in a 79-year-old male patient. The radiologist grade was a grade 1 concave
phology (wedge, concave, or crush fracture; the computer grade was a grade 2 concave fracture.

Figure 6

Figure 6:  The distribution of vertebrae in the thoracic and lumbar spine with compression fracture deformities. This
graph is annotated in standard anatomic fashion, with T1 indicating the first thoracic vertebra, L1 indicating the first
lumbar vertebra, and so on. At the top of each bar is the number of compression fractures at that anatomic level in the
case set. The expected bimodal distribution of the frequency of fractures is seen, with a peak in the midthoracic spine
(at T7 here) and a peak in the upper lumbar spine (at L1).

Radiology: Volume 284: Number 3—September 2017  n  radiology.rsna.org 793


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

Table 2 Table 3
Confusion Matrix of Genant Fracture Morphology: Radiologist versus Computer Confusion Matrix of Genant Severity
Assessment of Fracture Type Grade: Radiologist versus Computer
Assessment
Computer Classification
Computer
Radiologist Classification Wedge Fracture Concave Fracture Crush Fracture
Radiologist Grade 1 Grade 2 Grade 3
Wedge fracture 80 0 0
Concave fracture 5 25 0 Grade 1 23 10 2
Crush fracture 0 0 3 Grade 2 3 26 2
Grade 3 2 17 28
Note.—Data are number of cases with concordance (diagonal values) versus discrepancy (off-diagonal values) in the
categorization of Genant fracture morphology by radiologist and by computer system. Of the 210 fractured vertebrae, 113 were Note.—Data are number of cases with concordance
assessed for type and grade of height loss and are listed in this table. (diagonal values) versus discrepancy (off-diagonal
values) in the categorization of Genant severity grade by
radiologist and by computer system. Genant grades of
height loss are as follows: grade 0, less than 15% height
with the system relative to radiologist case and control sets, both at individ- loss; grade 1, 15%–25% height loss; grade 2, 26%–
classification (Table 3). The calculated ual vertebral levels from T12 to L4 as 40% height loss; and grade 3 more than 40% height
weighted k coefficient was 0.59 (95% well as in summation (P , .001, paired loss. Of the 210 fractured vertebrae, 113 were assessed
for type and grade of height loss and are listed in this
CI: 0.47, 0.71), which was indicative t test). On an individual basis, bone table.
of moderate or good agreement. FROC density of each age- and sex-matched
curves of system performance for com- control patient exceeded that of the
pression severity are shown in Figure 7. correlating case patient 67%–73% of
Additionally, system performance for the time, depending on the level from the determination of developmentally
severity grade of compres­sion was also T12 to L4. related scoliosis, and Scheuermann
determined relative to manual annota- disease.
tion for each of the Genant-enhanced This system performed with 95.7%
classification zones (A, C, P, R, or L). Discussion sensitivity in fracture detection and lo-
There was 59% overall agreement (66 We designed and validated an auto- calization to the correct vertebral level,
of 113, 95% CI: 0.49, 0.67) for cate- mated machine learning system to with a low false-positive rate. There
gorization of compression severity for detect and enumerate the level of ver- was a high level of overall agreement
zone A, 65% overall agreement (65 of tebral compression fractures and to (95%) for compression morphology
113, 95% CI: 0.49, 0.67) for categori- classify fracture morphology and se- and 68% overall agreement for sever-
zation of compression severity for zone verity according to Genant standards. ity categorization relative to radiologist
C, and 86% overall agreement (97 of We have updated the Genant schema classification.
113, 95% CI: 0.78, 0.91) for categoriza- to use the full power of 3D CT images. Quantitative data generated by our
tion of compression severity for zone P. Additionally, this system was designed system have the potential to decrease
There was 69% overall agreement (78 to delineate bone density, with the po- interobserver variability in compres-
of 113, 95% CI: 0.60, 0.77) for cate- tential to provide combined metrics for sion fracture detection and in morphol-
gorization of compression severity for risk stratification of progressive com- ogy and severity classification. Detailed
zone R and 69% overall agreement (78 pression fracturing of prevalent verte- information regarding morphology and
of 113, 95% CI: 0.60, 0.77) for cate- brae and incident vertebral fracturing severity, combined with the per-patient
gorization of compression severity for based on both shape and compositional speed of analysis (approximately 5 mi-
zone L. Figure 5 shows examples of feature metrics. nutes), can be used to predict future
compression fracture grading by a radi- The system holds potential to de- ability to provide repeatable, timely, and
ologist and with a computer. Additional crease interobserver variability of detailed evaluation of compression pat-
examples of compression fracture grad- Genant classification via systematic terns within the Genant schema and to
ing can be found in Figure E5 (online). and uniform quantitative compression allow extension to new 3D classification
Bone densitometry results showed assessment. The vertebral body is a schemas. Fracture risk stratification
average bone density in patients with 3D structure, and the system can be is of current interest in patients with
fractures was less than that in those used to evaluate asymmetric left- and osteopenia or osteoporosis (28). The
without fractures, as we intuitively right-sided compression, extending the stacked height compass model forms
expected (Table 4). The average bone two-dimensional radiographic Genant a global height loss view of the spinal
density for the T12-L4 interval was 146 system into three dimensions. As such, column and may be applicable to the
HU 6 29 in the case set and 173 HU the system also enables assessment semiquantitative spinal deformity index
6 42 in the control set. The difference of postcompression scoliosis and ky- risk matrix (29). Quantitative mor-
in bone density was significant between phosis. This system holds potential in phologic features of vertebrae, both

794 radiology.rsna.org  n Radiology: Volume 284: Number 3—September 2017


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

individually and statistically as a group, performed preliminary work in direct compression. Normal architecture of
in combination with bone density gen- determination of height loss in the the adjacent vertebrae can be variable
erated with the system, may further height compass model, without cre- due to degenerative change, variant
aid comparative effectiveness research, ation of a Genant model and without anatomy, and compression fracture,
thereby generating predictive features Genant categorization or assessment limiting its use as a reference point.
and risk profiles for further fracturing of bone density (35). A normal increase in vertebral body
and helping to guide the development The main limitation of this system size proceeding from cranial to cau-
of new treatment paradigms based on is thought to be the lack of concrete dal incorporates systematic error into
evidence-based medicine patient out- comparative manual reference stan- the measure of relative height loss.
come studies. dards with which to determine the The computer determined percentage
The bimodal distribution of percentage height loss. Posterior height loss by uniform and systematic
fractures mirrored that which was ex- body wall compression can limit use averaging of multiple adjacent ver-
pected and reported in other works as a reference standard for anterior tebrae; this was difficult to replicate
(15,21). Fracture severity was some-
what unusual. Grade 3 deformities were
the most common finding, whereas we Figure 7
had expected grade 1 deformities to be
the most common finding on the basis
of prior studies (30). Possible reasons
for this discrepancy include small sam-
ple size and sampling bias.
As intuitively expected, the system
calculated a significantly lower bone
density in the case population than in
the age- and sex-matched control sub-
jects, on average. On an individual case-
matched basis, bone density of patients
in the case group exceeded that of pa-
tients in the control group 27%–33% of
the time. This was thought to be most
likely related to selection bias and vari-
ation in the individual medical histories
of patients in this small sample.
Prior work to assess vertebral
compression fractures on lateral ra-
diographs has reached the stage of Figure 7:  FROC curves arranged by grade. FROC curves of system perfor-
clinical application (31,32). We have mance for Genant fracture severity grade 1 (,25%), grade 2 (26%–40%), and
designed a system to assess height grade 3 (.40%). The performance difference between grade 1 and grade 3
loss on CT images and to incorporate classification is significant (P = .05), while performance differences between
the 3D characteristics of the verte- grades 2 and 3 (P = .12) and between grades 1 and 2 (P = .64) are not.
bral body. When compared with the
method of Hsieh et al (33), vertebral Table 4
body height loss is measured directly
with 3D segmentation and partitioning Bone Density Comparison of Case and Control Studies for Various Thoracolumbar
of the vertebrae rather than indirectly Vertebral Levels
by measuring spinal curvature. We Control Attenuation Greater
have incorporated the left and right Vertebral Level Case Attenuation (HU) Control Attenuation (HU) than Case Attenuation (%) P Value
aspects for 3D lateralization, extend-
ing beyond anterior, middle, and pos- T12 146 6 29 169 6 37 67 ,.001
L1 141 6 26 173 6 43 73 ,.001
terior sagittal midline vertebral body
L2 147 6 28 171 6 40 68 ,.001
height loss as in Ghosh et al (34) and
L3 146 6 31 172 6 44 67 ,.001
Baum et al (25), and compared with
L4 146 6 29 179 6 45 73 ,.001
Baum et al (25), we include the en-
Average 146 6 29 173 6 42 71 ,.001
tire thoracic spine, not just the T5-T12
vertebrae. Additionally, we measured Note.—Bone attenuation is given as mean 6 standard deviation.
bone density in this study. Our group

Radiology: Volume 284: Number 3—September 2017  n  radiology.rsna.org 795


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

manually and was likely more accu- 2. American Society of Neurological Surgeons. 14. Delmas PD, van de Langerijt L, Watts NB,
rate than manual assessment. A sec- Vertebral compression fractures. In: Ameri- et al. Underdiagnosis of vertebral fractures
can Society of Neurological Surgeons/Pa- is a worldwide problem: the IMPACT study.
ond limitation was related to the finite
tient Information/Vertebral Compression J Bone Miner Res 2005;20(4):557–563.
size of each axial zone with (generally) Fractures. Rolling Meadows, Ill: American
continuous variation of vertebral body 15. Samelson EJ, Christiansen BA, Demissie S,
Society of Neurological Surgeons, 2016.
et al. Reliability of vertebral fracture assess-
height within each zone. The reference
3. Lenchik L, Rogers LF, Delmas PD, Genant ment using multidetector CT lateral scout
standard data were created by visual HK. Diagnosis of osteoporotic vertebral views: the Framingham Osteoporosis Study.
inspection of images for the approxi- fractures: importance of recognition and de- Osteoporos Int 2011;22(4):1123–1131.
mate midpoint, allowing for skewed scription by radiologists. AJR Am J Roent-
16. Schousboe JT, Debold CR. Reliability and

random variation in the choice of ref- genol 2004;183(4):949–958.
accuracy of vertebral fracture assessment
erence point. A third limitation was 4. Lentle BC, Brown JP, Khan A, et al. with densitometry compared to radiog-
that our estimation of sensitivity was Recognizing and reporting vertebral raphy in clinical practice. Osteoporos Int
based on use of a cutoff derived from fractures: reducing the risk of future os- 2006;17(2):281–289.
the data and the resulting FROC curve; teoporotic fractures. Can Assoc Radiol J
17. Alexandru D, So W. Evaluation and man-

thus, the calculated 95.7% sensitivity 2007;58(1):27–36.
agement of vertebral compression fractures.
may be an overestimate. Finally, the 5. Francis RM, Aspray TJ, Hide G, Sutcliffe Perm J 2012;16(4):46–51.
bone density determined was not re- AM, Wilkinson P. Back pain in osteopo- 18. Predey TA, Sewall LE, Smith SJ. Percuta-

normalized into bone mineral density rotic vertebral fractures. Osteoporos Int neous vertebroplasty: new treatment for
2008;19(7):895–903.
units, as an external reference was not vertebral compression fractures. Am Fam
available for all patients in this retro- 6. Fink HA, Milavetz DL, Palermo L, et al. Physician 2002;66(4):611–615.
spective study. However, Hounsfield What proportion of incident radiographic 19. Anselmetti GC, Muto M, Guglielmi G,

vertebral deformities is clinically diag-
units from clinical CT examinations Masala S. Percutaneous vertebroplasty
nosed and vice versa? J Bone Miner Res or kyphoplasty. Radiol Clin North Am
have been previously shown to corre- 2005;20(7):1216–1222. 2010;48(3):641–649.
late with dual energy x-ray absorpti-
ometry values (27). 7. Wong CC, McGirt MJ. Vertebral compres- 20. Ross PD, Genant HK, Davis JW, Miller PD,
sion fractures: a review of current manage- Wasnich RD. Predicting vertebral fracture
In conclusion, we have designed and
ment and multimodal therapy. J Multidiscip incidence from prevalent fractures and bone
validated a fully automated quantitative Healthc 2013;6:205–214. density among non-black, osteoporotic
system with which to detect and enu- women. Osteoporos Int 1993;3(3):120–126.
8. Siris ES, Genant HK, Laster AJ, Chen P,
merate levels of thoracic and lumbar
Misurski DA, Krege JH. Enhanced predic- 21. Jager PL, Jonkman S, Koolhaas W, Stieke-
vertebral body compression fractures tion of fracture risk combining vertebral ma A, Wolffenbuttel BH, Slart RH. Com-
on CT images, determine Genant fracture status and BMD. Osteoporos Int bined vertebral fracture assessment and
classification of compression severity 2007;18(6):761–770. bone mineral density measurement: a new
and morphology, generate extended 9. McCloskey EV, Vasireddy S, Threlkeld J, et standard in the diagnosis of osteoporosis
Genant scores for fracture lateraliza- al. Vertebral fracture assessment (VFA) with in academic populations. Osteoporos Int
tion, and generate bone densities. Our a densitometer predicts future fractures 2011;22(4):1059–1068.
work is proposed as a proof-of-concept in elderly women unselected for osteo- 22. Huopio J, Kröger H, Honkanen R, Saa-

system for fracture risk stratification porosis. J Bone Miner Res 2008;23(10): rikoski S, Alhava E. Risk factors for peri-
1561–1568. menopausal fractures: a prospective study.
and clinical decision support.
10. Delmas PD, Genant HK, Crans GG, et al. Osteoporos Int 2000;11(3):219–227.
Disclosures of Conflicts of Interest: J.E.B. dis-
Severity of prevalent vertebral fractures and 23. Genant HK, Li J, Wu CY, Shepherd JA.

closed no relevant relationships. J.Y. Activities
related to the present article: disclosed no rel- the risk of subsequent vertebral and non- Vertebral fractures in osteoporosis: a new
evant relationships. Activities not related to the vertebral fractures: results from the MORE method for clinical assessment. J Clin Den-
present article: receives patent royalties from trial. Bone 2003;33(4):522–532. sitom 2000;3(3):281–290.
iCAD, received graphics card donations from
11. Lunt M, O’Neill TW, Felsenberg D, et al. 24. Yao J, O’Connor S, Summers R. Automated
Nvidia. Other relationships: disclosed no rele-
vant relationships. R.M.S. Activities related to Characteristics of a prevalent vertebral spinal column extraction and partitioning.
the present article: disclosed no relevant rela- deformity predict subsequent vertebral In: International Symposium on Biomedical
tionships. Activities not related to the present fracture: results from the European Pro- Imaging (ISBI). Arlington, Va: IEEE, 2006;
article: receives patent royalties from iCAD, spective Osteoporosis Study (EPOS). Bone 390–393.
receives research support from iCAD and Ping 2003;33(4):505–513.
An, received graphics card donations from Nvid- 25. Baum T, Bauer JS, Klinder T, et al. Automatic
ia. Other relationships: disclosed no relevant 12.
Adams JE. Opportunistic identification detection of osteoporotic vertebral fractures
relationships. of vertebral fractures. J Clin Densitom in routine thoracic and abdominal MDCT.
2016;19(1):54–62. Eur Radiol 2014;24(4):872–880.
References 13. Roux C, Fechtenbaum J, Kolta S, Briot
26. Summers RM, Baecher N, Yao J, et al. Feasi-
1. Riggs BL, Melton LJ 3rd. The worldwide K, Girard M. Mild prevalent and incident bility of simultaneous computed tomographic
problem of osteoporosis: insights af- vertebral fractures are risk factors for new colonography and fully automated bone min-
forded by epidemiology. Bone 1995;17(5 fractures. Osteoporos Int 2007;18(12):1617– eral densitometry in a single examination. J
Suppl):505S–511S. 1624. Comput Assist Tomogr 2011;35(2):212–216.

796 radiology.rsna.org  n Radiology: Volume 284: Number 3—September 2017


MUSCULOSKELETAL IMAGING: Vertebral Body Compression Fractures Burns et al

27. Schreiber JJ, Anderson PA, Rosas HG, Bu- and recent advances. Eur Radiol 2008;18 van Ginneken B, eds. Proceedings of SPIE:
chholz AL, Au AG. Hounsfield units for as- (7):1484–1496. medical imaging 2011—computer-aided diag-
sessing bone mineral density and strength: 32. Guglielmi G, Palmieri F, Placentino MG,
nosis. Vol 7963. Bellingham, Wash: Interna-
a tool for osteoporosis management. J Bone D’Errico F, Stoppino LP. Assessment of oste- tional Society for Optics and Photonics, 2011;
Joint Surg Am 2011;93(11):1057–1063. oporotic vertebral fractures using specialized 796303.

28. Unnanuntana A, Gladnick BP, Donnelly E, workflow software for 6-point morphometry. 35. Yao J, Burns JE, Wiese T, Summers

Lane JM. The assessment of fracture risk. Eur J Radiol 2009;70(1):142–148. RM. Quantitative vertebral compression
J Bone Joint Surg Am 2010;92(3):743–753. 33. Hsieh MS, Tsai MD, Yeh YD, Jou SB.
fracture evaluation using a height compass.
Automatic spinal fracture diagnosis and In: van Ginneken B, Novak CL, eds. Pro-
29. Crans GG, Genant HK, Krege JH. Prognos-
surgical management based on 3D image ceedings of SPIE: medical imaging 2012—
tic utility of a semiquantitative spinal defor-
analysis and reconstruction of CT transverse computer-aided diagnosis. Vol 8315. Bell-
mity index. Bone 2005;37(2):175–179.
sections. Biomed Eng Appl Basis Commun ingham, Wash: International Society for
30. Diacinti D, Guglielmi G. Vertebral mor-
2002;14(5):204–214. Optics and Photonics, 2012; 83151X.
phometry. Radiol Clin North Am 2010;48(3):
34. Ghosh S, Alomari RS, Chaudhary V, Dhil-
561–575. lon G. Automatic lumbar vertebra segmen-
31. Guglielmi G, Diacinti D, van Kuijk C, et al. tation from clinical CT for wedge compres-
Vertebral morphometry: current methods sion fracture diagnosis. In: Summers RM,

Radiology: Volume 284: Number 3—September 2017  n  radiology.rsna.org 797