Trabecular Structure Quantified With the MRI-Based Virtual Bone Biopsy in Postmenopausal Women Contributes to Vertebral Deformity Burden Independent of Areal Vertebral BMD


  • Dr Wehrli is a Chairman of the Scientific Advisory Board of MicroMRI. All other authors state that they have no conflicts of interest.


In postmenopausal women with a wide range of vertebral deformities, MRI-based structural measures of topology and scale at the distal radius are shown to account for as much as 30% of vertebral deformity, independent of integral vertebral BMD.

Introduction: Trabecular bone architecture has been postulated to contribute to overall bone strength independent of vertebral BMD measured by DXA. However, there has thus far been only sparse in vivo evidence to support this hypothesis.

Materials and Methods: Postmenopausal women, 60-80 yr of age, were screened by DXA, and those with T-scores at either the hip or spine falling within the range of −2.5 ± 1.0 were studied with the MRI-based virtual bone biopsy, along with heel broadband ultrasound absorption and pQCT of the tibia. The data from 98 subjects meeting the enrollment criteria were subjected to νMRI at the distal tibia and radius, and measures of topology and scale of the trabecular bone network were computed. A spinal deformity index (SDI) was obtained from morphometric measurements in midline sagittal MR images of the thoracic and lumbar spine to evaluate associations between structure and deformity burden.

Results: A number of structural indices obtained at the distal radius were correlated with the SDI. Among these were the topological surface density (a measure of trabecular plates) and trabecular bone volume fraction, which were inversely correlated with SDI (p < 0.0001). Combinations of two structural parameters accounted for up to 30% of the variation in SDI (p < 0.0001) independent of spinal BMD, which was not significantly correlated. pQCT trabecular BMD was also weakly associated, whereas broadband ultrasound absorption was not. No significant association between SDI and structural indices were found at the tibia.

Conclusions: Structural measures at the distal radius obtained in vivo by νMRI explained a significant portion of the variation in total spinal deformity burden in postmenopausal women independent of areal BMD.


Although low BMD is a well-established risk factor for osteoporotic fracture, it accounts for only ∼60% of the contribution to bone strength.(1,2) In fact, the recently published results of the NORA study(3) indicated that more than one half of women sustaining osteoporotic fractures had BMD T-scores ≥ −2.0, and a significant portion had BMD in the range considered normal. During the past 10 yr, these and similar observations have spurred the search for other predictors of fracture susceptibility, often summarized under the notion of “bone quality.”(4) This term encompasses contributions from architecture at various length scales from macro to nano, as well as intrinsic material properties such as mineralization density and matrix chemistry (e.g., collagen cross-linking and mineral crystallinity).(5)

The concept of architecture as a risk factor for fracture independent of BMD is at least two decades old, but clinical evidence is still sparse. Most of the evidence for such a hypothesis is based on histomorphometry from sections of iliac crest bone biopsies.(6-9) Kleerekoper et al.(6) first showed lower mean trabecular plate density in patients with postmenopausal osteoporosis who had vertebral compression fractures than subjects without fractures matched for age, sex, race, menopausal status, and BMD. Similarly, Recker(7) matched patients with vertebral crush fractures by bone volume fraction to healthy subjects and found the former to differ in a number of structural indices. Most notably, the subjects with fractures had increased marrow star volume and decreased connectivity, both by 35-40%. A third study based on histomorphometry from iliac crest bone biopsies was in men with osteoporosis, some of whom had vertebral fractures.(9) The authors found that after adjusting for age, body mass index, and BMD, the two groups did not differ in BV/TV, Tb.Th, or star volume, unlike in the previous study, whereas parameters related to connectivity were significantly different in the two groups.

The emergence of noninvasive imaging methods—notably high-resolution MRI and CT—has provided an additional incentive to explore the role of microarchitecture as a determinant of fracture risk. During recent years, a number of studies, including work from this laboratory,(10,11) have appeared that examined the role of trabecular bone structure at a peripheral site as an indicator of vertebral fracture risk.(10-19) The potential for direct assessment of architecture in the axial skeleton by high-resolution CT as a means of differentiating subjects with vertebral fractures from their unfractured peers was first explored over a decade ago.(20,21) More recently, with the advent of multidetector CT, significant improvements in image quality have been achieved.(22) Concerns, however, remain about CT of the axial skeleton because of the significant radiation dose such studies entail.

Ongoing research in this laboratory during the past several years has led to a νMR-based procedure we refer to as “virtual bone biopsy” (VBB) that provides detailed quantitative information on trabecular bone architecture at peripheral sites such as the distal radius and tibia.(23) Prior work showed that women with vertebral deformities measured tomographically had a less connected trabecular network of the distal radial metaphysis than their unfractured peers.(11) The work further indicated that structural parameters were better discriminators of the two groups than DXA BMD measured either in the vertebrae or upper femur.

The principal aim of this study was to apply the VBB technique using substantially improved technology to quantify trabecular bone architecture in terms of scale and topology in a well-defined group of elderly postmenopausal women within a predefined narrow range of areal BMD and to correlate these VBB parameters with the degree of spinal deformity. The hypothesis was that trabecular bone structure measured at two distal sites—the distal radius and distal tibia—is a significant independent contributor to vertebral fracture susceptibility above and beyond that provided by BMD determined by DXA. For comparison, other measurements were performed as well, including broadband ultrasound absorption (BUA) in the heel and peripheral QCT of the distal tibia. This design also permits comparing VBB parameters at a load-bearing (distal tibia) and non-load-bearing sites (distal radius) as surrogate measures of vertebral deformity in osteoporosis.



Subjects were recruited from the Philadelphia area using flyers sent to their homes and advertisements in university publications and included patients who had participated in previous studies. The subjects were enrolled and examined in the Department of Radiology at the University of Pennsylvania and the Nutrition Core Laboratory of the General Clinical Research Center at the Children's Hospital of Philadelphia. A questionnaire was used to gather information on demographics, medication, and medical history. The Institutional Review Board approved the study protocol, and all study participants gave their written informed consent.

Inclusion criteria

Women were included who were >60 yr old, postmenopausal, and had T-scores of −2.5 ± 1.0 at either the spine or hip. The subjects were enrolled and examined from October 8, 2003 to February 9, 2005.

Exclusion criteria

Subjects were excluded who had ferromagnetic implants, were claustrophobic, or who were taking medications that affect mineral homeostasis (bisphosphonates within the previous 24 mo, calcitonin within the previous 6 mo, and glucocorticoids (≥20 mg/d) for >2 wk of the previous 6 mo). Other exclusion criteria were hyperparathyroidism, gastrointestinal disease resulting in malabsorption, current alcohol use (>3 drinks/d), or illicit drug use.

Of the women interviewed, 232 of 917 passed telephone screening criteria; those who qualified had their BMD measured by DXA. Of the 232 women, 118 were excluded because of BMD T-scores above −1.5, two had BMD T-scores below −3.5, one was excluded because of high alcohol consumption, one was taking a bisphosphonate, one had gastric bypass, one was recently treated with steroids, and one had high alkaline phosphatase suspicious of Paget's disease. Two patients were claustrophobic, four others refused for personal reasons, two subjects could not fit into the wrist coil, and one had metal bracelets that could not be removed. MRI scans of the distal tibia, wrist, and spine were obtained from the remaining 98 subjects.

Some scans could not be used for evaluation of trabecular topology because of poor image quality. To eliminate bias, three team members independently graded the image quality of the MRI scans with scores ranging from 1 to 4 (Fig. 1). The most common reasons for inferior image quality were artifacts resulting from involuntary subject movement that could not be corrected with the motion correction techniques described below. The threshold for exclusion was set so that exams with quality grade 4 by two of the three examiners were excluded. In this manner, 4 of 98 subjects were excluded because of poor image quality of the structural images. Finally, not all subjects yielded spine images that covered the vertebrae from T6 through L5, which reduced the number of analyzable subjects as explained later.

Figure Fig. 1..

Illustration of quality scoring for the MRI exams using tibia cross sections: grade 1 = near optimum quality; grade 2 = slight blurring; grade 3 = significant blurring that might affect fine features or cause some distortion of larger features; grade 4 = uncorrected motion artifacts too severe for trabecular features to be visualized and analyzed.

MRI-based VBB: image acquisition

All MRI scans were performed on a 1.5-T whole body clinical MRI scanner (Signa; General Electric, Milwaukee, WI, USA) equipped with 22-mT/m gradients. An elliptical, transmit-receive birdcage coil was used for the wrist and a receive-only, dual-element, phased-array surface coil for the ankle, both custom-built in the authors' laboratory. To prescribe the high-resolution imaging slab, whose inferior margin was 4 mm from the cortical endplate of the distal radius and 10 mm from the distal end plate of the tibia, axial and sagittal localizer spin-echo scans were acquired. Subsequently, a 3D FLASE (fast large-angle spin-echo) pulse sequence(24) was used for the high-resolution images of the wrist and tibia with a flip angle of 140°, TR/TE 80 ms/9.5 ms (fractional acquisition), 7.81 ± kHz bandwidth, field of view (FOV) 7 × 4 cm (wrist) or 7 × 5 cm (tibia), with one excitation. Thirty-two slices were obtained along the partition direction (inferior-superior), and 512 sample points were obtained along the frequency-encoding direction (anterior-posterior). Along the phase-encoding direction, 384 data points were obtained for the tibia and 288 for the wrist. Navigator data were taken every 0.2 s to monitor motion during the scan.(24) The voxel size of the raw data was 137 × 137 × 410 νm3, with the last dimension relating to the partition direction. The acquisition time was 16 min and 23 s for the tibia and 12 min and 18 s for the wrist.

MRI-based VBB: processing and analysis

The cascade of processing steps is shown in Fig. 2. After data acquisition, in-plane motion correction was applied to the spatial frequency (k-space) data using the navigator data obtained during the scan. After Fourier transformation of the motion-corrected k-space data yielding 32 contiguous image slices, a semiautomated masking routine was applied to segment the region of interest based on signal-to-noise ratio (SNR) criteria and gross anatomy. The result of this operation is a set of 28 images of the trabecular cross-section of either the distal radial or tibial metaphysis. From there, a bone volume fraction map was produced in which the voxel signal intensity represents the fractional occupancy by bone.(25) Apparent resolution was enhanced by subvoxel processing, a method in which voxels are subdivided and BV/TV reallocated according to proximity criteria,(26) yielding a final voxel size of 68.5 × 68.5 × 102.5 νm3.

Figure Fig. 2..

Cascade of image acquisition and processing steps. After data acquisition, the navigator data are used to correct for subject displacement during the scan. Subsequently, region of interest (ROI) for analysis is selected. After bone volume fraction mapping and subvoxel processing, the topological parameter densities are computed and a 3D rendering of a virtual core is computed for visual inspection.

Trabecular thickness (Tb.Th) was computed directly from the BV/TV maps by means of the fuzzy distance transform algorithm, which has been shown to return accurate values in the presence of partial volume blurring and noise.(27) Two additional parameters, which had previously been shown to be predictive of fracture,(10) were obtained by using spatial autocorrelation treating the trabecular network as a quasi-regular lattice. Both parameters are computed by normalizing the probability of finding bone in two adjacent voxels by the probability that two points from the same voxel are both in bone.(28) Transverse contiguity (TCON) was computed from voxels adjacent in the transverse plane (perpendicular to the bone's long axis), whereas the tubularity (TUB) is derived from voxels adjacent in the longitudinal direction (parallel to the bone's long axis). The latter thus measures the degree of longitudinal alignment.

To determine trabecular network topology, the preprocessed images were binarized to separate bone from marrow and skeletonized in a topology-preserving manner. Skeletonization results in a structure in which trabecular plates are converted to voxel surfaces (S) and rods to curves (C) as shown in Fig. 3. Digital topological analysis (DTA) produces a topological classification of each voxel of the skeletonized trabecular network.(29,30) In this manner, voxels in the interior of the surfaces and curves (SI and CI, respectively) can be distinguished from those at the edges (surface edge [SE] and curve edge [CE]). Intermediate structures called profiles, which are essentially two-voxel ribbons analogous to curves, are also classified (profile edge [PE] and profile interior [PI]). DTA further identifies voxels pertaining to junctions (J) between surfaces (SS), between curves (CC), and between surfaces and curves (SC). Topological densities are unitless quantities representing the fraction of voxels pertaining to a particular class relative to all voxels within the sampling volume. Finally, composite parameters were evaluated: the ratio S/C (the sum of S-type voxel densities divided by C-type voxel densities) and an erosion index (EI, the ratio of voxel densities that decrease, divided by those that increase with osteoclastic resorption).(11) These composite parameters were found to be particularly sensitive to structural alterations because perforative resorption, for example, has the dual effect of decreasing surface-type voxels and increasing curve-type voxels. Thus, for EI, such a process entails an increase in the numerator and a decrease in the denominator yielding an increase in the parameter.

Figure Fig. 3..

Digital representation of hypothetical skeletonized structure of trabecular network containing only surfaces and curves (the skeleton analogs of plates and rods, respectively) with their topological assignments.

Finally, the reproducibility in terms of the average CV was previously established to range from 4% to 7% and was found to be comparable for the two anatomic sites and for all parameters measured.(31)

Spine MRI protocol

In addition to the high-resolution trabecular bone scans, MRI was performed in the spine for the purpose of quantifying vertebral deformity similar to the protocol described in.(11) Toward this goal, sagittal images of the spine were obtained with the manufacturer's spine surface coil array using a fast spin-echo sequence (TR/TE of 4000/13.6 ms, echo train length = 8, bandwidth = 31.25 kHz, NEX = 2, field of view = 40 × 30 cm, 0.78 × 0.78 mm2 pixel size), with a variable number of slices of 5 mm thickness. MR overcomes some of the inherent limitations of projection imaging in that it allows for correction of errors caused by scoliosis and sagittal obliquity in that the imaging slice used for vertebral deformity can be individually selected so as to ensure that it transects the vertebra along its midline. Localizer scans were used to select the portion of the vertebral column to image the lumbar and thoracic vertebrae. With the longer side of the FOV selected in the superior-inferior direction, the number of vertebrae imaged for each subject varied with body habitus. The final spinal data provided cross-sectional images with a FOV of 40 × 30 cm (image matrix size of 512 × 256, TR/TE of 4000 ms/13.6 ms), which allowed for the measurement of vertebral dimensions in the sagittal plane. Depending on subject size, vertebrae were imaged from S1 to as high as C7, with scans capturing T6-L5 in most subjects.

Quantifying vertebral deformity

The method for quantifying vertebral deformity has been described earlier(11) and follows the approach by Eastell et al.(32) In brief, using custom-designed software written in IDL, the spine images were displayed, and each vertebra was manually marked at the four corners and the midpoints of the superior and inferior edges. On the basis of images as those shown in Fig. 4, the quantities Ha (anterior height), Hm (middle height), and Hp (posterior height) were computed on the interpolated sagittal images transecting the spine in the midline. The three types of deformity were obtained in this manner:

equation image

From the three types of deformities, a continuous measure of spinal deformity called spinal deformity index (SDI) was computed as a composite measure of wedge, biconcavity, and compression deformities. The formulae, first introduced elsewhere,(11) have been used for consistency with minor modifications. Furthermore, special precautions were taken in the measurement to avoid errors from osteophytes and depressions caused by endplate herniations (Schmorl's nodes).

Figure Fig. 4..

Midline sagittal images of the lumbar and portion of thoracic spine in three women showing the wide range of SDI values obtained by quantifying vertebral deformity: (A) SDI = 11.1, virtually all vertebrae visible show substantial deformities in this 88-yr-old woman, leading to an exceptionally high deformity burden; (B) SDI = 5.3, deformities are less pronounced in this 65-yr-old woman; (C) SDI = 2.2, no obvious deformities are discernable in this 71-yr-old woman.

Because variations in any one of the three types of vertebral deformities have different clinical relevance (i.e., a small change in compression deformity may have greater implications on overall deformity burden than a similar change in wedge deformity), it is appropriate to individually weight wedge, biconcavity, and compression deformities.(33) Accordingly, the weighted deformities were summed to create the SDI that expresses overall deformity burden:

equation image

Two different SDI values were computed. The first included all subjects for whom vertebral deformity data were available for T6 through L5 (SDIT6L5, N = 67). This range in vertebral coverage was not available for all subjects. To be able to include subjects who would otherwise be excluded because of inadequate vertebral coverage, the SDI was also computed by weighting it by the number of vertebrae for which deformity measures could be computed (SDInorm, N = 79).


DXA bone densitometry was performed at the lumbar spine and proximal femur using a Hologic Discovery bone densitometer (Version 12.3; Hologic, Bedford, MA, USA). All scans were reviewed by a single investigator (BSZ) for technical quality and to ensure consistent positioning and analysis. Scans were analyzed to generate measures of AP spine (vertebrae L1-L4) and proximal femur (femoral neck, trochanteric, and intertrochanteric regions) areal BMD (g/cm2). Following the recommendations of the International Society for Clinical Densitometry, T-scores were generated using reference data for white women based on the manufacturer's software. T-scores were used to establish whether the subject met entry criteria. The in vitro CV for spine BMD was <0.6% and the in vivo CV in adults was <1%.


pQCT (XCT 2000; Stratec, Pforzheim, Germany) measures of cortical bone were performed in the diaphysis of the left tibia 17 mm proximal to the line of growth plate fusion and 11 mm proximal to the line of growth plate fusion for the radius. A single tomographic slice of 2.3-mm slice thickness was taken at a pixel size of 0.4 × 0.4 mm2 at the respective distance proximal to the reference line. BMD values were computed with the manufacturer's software: BMDtotal (BMC divided by the volume enclosed within the periosteal envelope; mg/mm3) and trabecular BMD (BMC of the trabecular bone compartment divided by the area it encompasses, expressed in mg/mm3). The in vivo CV for trabecular BMD by pQCT is 1.3% and the in vitro CV is <0.5%.

Heel ultrasound

Ultrasound assessment of the calcaneus was performed using a contact ultrasound bone analyzer (CUBA; Cooper Surgical, Trumbull, CT, USA). BUA measurements of the left calcaneus, which reflect the frequency dependence of ultrasound attenuation in the range 200-600 kHz (expressed in dB/MHz), was measured three times, and the mean was used in the analysis. The in vivo CV for CUBA is 3.6%.

Statistical analysis

Normality of all parameters measured was examined from quantile plots. MRI-derived scale and topological parameters were analyzed against SDI by linear regression. Stepwise linear regression with forward selection was also performed involving up to nine parameters with retention of the three most significant ones. Similar analyses were also performed for DXA, pQCT, and BUA. Analyses were carried out with the statistical package JMPIN 4.0.4 (SAS Institute, Cary, NC, USA). Correlation coefficients reported were derived from the more conservative adjusted r2 values. In view of the multiple correlations performed involving the same basic structural images from which they were derived, the α level for statistical significance was set at 0.005. Correlations between trabecular parameters were also studied to assess the dependence of the DTA structural parameters on density measures and further to examine logical consistency.


Subject characteristics

Table 1 lists the characteristics of the study subjects included in the analysis. The number of subjects who had scans of sufficient quality to analyze differed between tibia and wrist, so the characteristics of each subset, as well as the total group, are given. No marked imbalances are noted between the subjects contributing to the wrist and tibia analyses. The subjects excluded from the analysis because of inadequate MR image quality did not differ in the vertebral BMD selection parameters, so that this exclusion could not have skewed the data in any manner.

Table Table 1.. Characteristics of Study Subjects
original image


By designing the experiment to include subjects with T-scores within the boundaries of −2.5 ± 1.0, a broad spectrum of spinal deformities could be expected. Figure 4 shows the wide range of SDI results from spine embodies a high deformity burden (Fig. 4A), whereas a low SDI results from a healthy spine (Fig. 4C). It is to be noted that a single severe deformity can produce very high SDI values by this analysis.

Images of the distal radius from three subjects are displayed in Fig. 5 to visually exemplify the range in the degree of structural integrity seen, in good qualitative agreement with the derived topological measures obtained by the analysis. The data highlight the dramatic differences among subjects with topological indices varying by as much as an order of magnitude among subjects.

Figure Fig. 5..

νMRIs of the distal radius (top row) with their respective virtual cores (bottom row) along with structural parameters from three subjects exemplifying a wide range in bone quality represented by the topological parameters that vary by over an order of magnitude between the extremes: (A) 68-yr-old woman having a well-connected bone structure; (B) 69-yr-old woman with less dense trabecular network; (C) 87-yr-old woman with sparse trabeculae and disconnected network.

Associations between architecture and vertebral deformity status

A number of structural parameters at the radius were found to be associated with SDI. Correlation coefficients for both metrics are given in Table 2. The strength of the correlations was somewhat greater for SDIT6L5 than for SDInorm. The strongest correlations between SDI and structural parameters were those at the distal radius involving the topological parameters but highly significant relationships were also found for trabecular thickness and bone volume fraction. Figure 6 shows correlations between surface density and SDI (r = −0.46, p < 0.0001) and the previously defined composite parameter denoted EI, a measure of the degree of structural degradation (r = 0.40, p = 0.0005). Other correlations of similar strength were obtained for surface components such as surface edges, SE (r = −0.47, p < 0.0001) or surface-surface junction voxels, SS (r = −0.42, p = 0.0002). Parameters representative of network scale were also significantly correlated (BV/TV: r = −0.46, p < 0.0001; TB.Th: r = −0.37, p = 0.001). The erosion index, as would be expected, is positively correlated with SDI (Fig. 6B). Several of the groups of parameters expressing the spatial distribution of the bone such as tubularity, shown previously to be strength and fracture likelihood related,(10,28) yielded significant associations with SDI.

Table Table 2.. Correlations (r) Between the SDI (both SDIT6L5 and SDInorm) and Architectural Parameters Measured in the Distal Radial Metaphysis
original image
Figure Fig. 6..

Associations between topological parameters in the radius and SDI showing negative and positive correlation, respectively, for voxel surface density (A) and erosion index (B) (r = −0.46, p < 0.0001 and r = 0.40, p = 0.0005).

Of relevance is the sign of the various correlations. Most structural parameters were negatively correlated (as one would expect because reduced BV/TV, tubularity, trabecular thickness, surface density, etc., are known to lower the network's mechanical competence, thus increasing the probability for occurrence of vertebral deformities), thereby corroborating earlier data.(10,11) In contrast, the erosion index, a quantitative measure of network erosion,(30) was positively associated with the SDI. The negative correlation between SDI and surface voxel density (Fig. 6A) indicates that increased surface voxel density (i.e., more platelike bone) is protective against fracture. Interestingly, areal BMD of the lumbar vertebrae was not correlated with the SDI (although approaching statistical significance, r = 0.16, p = 0.09), whereas hip areal BMD was more strongly associated (r = −0.37, p = 0.001).

Among the two-parameter correlations examined, the combination of PE with EI as independent variables suggested a stronger association with the SDI than any single parameter (r = 0.56, p < 0.0001), but the difference in correlation coefficients was not statistically significant. Nevertheless, this observation suggests that 30% (r2 = 0.31) of the variation in vertebral deformity in this group of patients can be explained in terms of variations in structural make-up at the measurement site. Inclusion of hip BMD into the model resulted in a three-parameter correlation, with r = 0.63 (p < 0.0001). Finally, subject age was weakly correlated with SDI (r = 0.28, p = 0.006) as were years after menopause, but there were no correlations of significance of the vertebral deformity with height, weight, or BMI.

MRI-derived structural indices were not significantly correlated with SDI at the distal tibia. This finding is surprising considering that tibia structural parameters have been shown to be sensitive to antiresorptive treatment(34) and also that associations were found for pQCT data obtained at this site in this study. Reasons for the failure of the tibia data to predict vertebral deformity status are likely to be technical and are addressed in the Discussion.

pQCT-derived densities and heel ultrasound BUA parameters and their correlation with SDI are given in Table 3. Trabecular density (TBD) and total density (TotD) at both locations yielded significant associations (r = −0.31 to −0.38, p ≤ 0.005), whereas no significant association was found for BUA (r = −0.21, p = 0.08).

Table Table 3.. Correlations (r) Between SDIT6L5 and pQCT Measures of Trabecular Bone Density (TBD), Cortical Density (CortD), and Total Bone Density (TotD)
original image

Interparameter correlations

Table 4 lists the cross-correlations among structural measures. The data provide compelling evidence in support of the main hypothesis, namely, that the topological parameters provide information different from BMD. Most notably, there was no significant correlation with areal BMD of the spine, and correlations between the topological parameters and hip BMD were weak or nonexistent. Because S/C and EI both depend on a ratio of sums of similar variables but interchanged numerator and denominator, a highly negative correlation is noted as expected (r2 = −0.80). Similarly, curve-type parameters are strongly correlated with profile-type parameters (PE and PI) because both derive from trabecular struts (r2 = 0.61-0.92). The strong association between junction and surface voxel densities (r2 = 0.93) implies that the main contribution to junction-type voxels arises from joining surfaces, i.e., trabecular plates. Surface edges are a subset of surfaces. Similarly, PI elements cannot exist without profile edges. Both parameter pairs are therefore strongly correlated with r2 > 0.8. Because joining of surfaces tends to entail more junction elements than the joining of curves, it is unsurprising to find the junction density to be strongly correlated with surface density but only weakly with curve density.

Table Table 4.. Correlations (r2) Between Various Structural Parameters Evaluated for the Radius
original image


The data in this study showed that detailed quantitative insight into the architectural implications of postmenopausal bone loss can be obtained from the MRI-based VBB. Structural variables relating to scale and topology of the trabecular network measured at a peripheral site are shown to be associated with global vertebral deformity burden, unlike areal BMD of the lumbar vertebrae. By treating vertebral deformity as a continuous variable, the total deformity burden can be captured in a manner that avoids the need for arbitrarily selecting a threshold. Accurate determination of a SDI is facilitated by evaluation of high-resolution midline sagittal MR images that can be acquired during the same imaging session. By correlating the SDI with parameters of scale and topology, the specific role of each architectural measure and its total contribution to deformity burden can be assessed. Individual parameters representative of scale (Tb.Th and BV/TV) and topology of the TB scaffold were found to correlate with SDI. Among these were surface and junction densities (r = −0.46, p < 0.0001 and r = −0.44, p = 0.0001, respectively), which were negatively correlated, the former indicating that a loss of plates weakens the bone's mechanical competence, as would be expected.(35)

The substantially higher correlations obtained by regressing parameter pairs suggest independent contributions from two different structural measures. A case in point is the combination of profile edges, PE (essentially representing free ends, i.e., disconnected rodlike trabeculae) with erosion index yielding r = 0.56 (p < 0.0001), suggesting that >30% of the variation in vertebral deformity burden (r2 = 0.31) can be attributed to variations in architecture. These observations further support the notion that loss of connectivity, along with a less platelike structure, weakens the trabecular network, thus increasing the risk for vertebral fracture. Last, it is interesting that for the chosen group of subjects (age > 60 yr and areal BMD T-scores within the band −2.5 ± 1.0), there was no significant association between SDI and areal vertebral BMD and only a relatively weak one with hip areal BMD (r = −0.37, p = 0.001). These findings are not unexpected given that BMD was intentionally constrained in our study but also when considering the limitations of spinal DXA in this population. The virtual absence of an association between the degree of vertebral deformity burden and DXA BMD is paralleled by recent data by Ito et al.(22) Even though in that work the range of BMD was not controlled and thus likely much larger than in our work, the authors found only a weak correlation between fracture status and areal BMD of the lumbar spine (p = 0.03), in contrast to structural measures and volumetric vertebral BMD, which were strongly associated (p = 0.0001).

In summary, our data suggest that structural alterations at the measurement site are paralleled by commensurate changes at the vertebral site. This comes as no surprise because osteoporosis is well known to be a diffuse disorder affecting all skeletal sites. The complementary nature of the information relative to areal BMD is further exemplified by the improvement in correlation seen when hip areal BMD is combined, for example, with surface edge density SE, thereby increasing the correlation coefficient from r = 0.37 from BMD alone to r = 0.57.

Topological indices of various kinds have previously been shown to be stronger discriminators than BMD or bone volume fraction of subjects with fractures from those without fractures. The trabecular bone pattern factor (TBPf),(36) for instance, a ratio of concave to convex surfaces, is a stereologic measure of the bone's connectedness. Similar to other structural measures of topology, such as the Euler number from which connectivity is derived,(37) these parameters, obtainable from digital images, have been shown to be independent discriminators of osteoporotic fracture.(15,38) More recently, Ito et al.(22) found such topological parameters as the structure model index that, similar to the surface-to-curve ratio, are a measure of the bone's “platelikeness” in the vertebrae to be particularly strongly associated with fracture status.

Analysis of the distal tibia data showed no significant correlations between vertebral deformity and the topological parameters. To study possible variables that could mask a physical effect, a posthoc analysis was performed focusing on the effects of image SNR and masking technique (choice of analysis region). Selecting a mask based on anatomic criteria followed by stratification of the resulting data into two groups with SNR >7.5 and <7.5 resulted in 45 subjects with SNR < 7.5 and 32 subjects with SNR > 7.5. Analysis of the subgroup with SNR < 7.5 yielded no significant correlations, whereas the subgroup with SNR > 7.5, despite lower power, suggested associations between SDI and several parameters, although none of these were significant (recalling that the α level was set to 0.005 in view of multiple correlations being tested). Among these were tubularity (r = −0.35, p = 0.03), Tb.Th (r = −0.36, p = 0.02), PI (r = −0.33, p = 0.04), and EI (r = 0.30, p = 0.05), paralleling the results found at the radial site. Whereas the variations in the derived structural parameters are greater at the lower SNR, they are, nevertheless, variable at higher SNR levels as well, albeit to a lesser extent, as shown elsewhere.(40) Once the functional dependence of the parameters is known, a retrospective correction for subject-to-subject variations in SNR is feasible. Furthermore, improvements in overall SNR will alleviate and possibly eliminate this problem.

The imaging technology underlying this work is undergoing further improvements, which should alleviate some of the error sources, key among which are subject and physiologic motion,(41) low or subject-dependent SNR, and image registration issues.(31,40,42) Correction for translational motion through navigator echoes, unless it is particularly severe, has been found to be effective. However, even very small rotations (on the order of a degree) can cause errors as large as 10%.(31) The distal tibia is generally less susceptible to motion artifacts.(40) Figure 1 shows an exam excluded for severe motion where the trabecular features are too blurred for accurate comparative analysis. Recent advances in the authors' laboratory in retrospective motion correction through autofocusing will, in the near future, allow for both rotational and translational correction.(41)

Another element of uncertainty concerns the quantification of the spinal deformity index, especially the delineation of the perimeter of each vertebra. In this population, degenerative joint disease can affect the measurement of vertebral dimensions. Interobserver variability may increase in light of extraosseus calcification in some of the elderly subjects. In this manner, osteophytes, which are not representative of compression deformities, can extend the anterior and posterior margins of the vertebrae in a vertical direction, thereby leading to an overestimation of the anterior and posterior height.

The approach presented for assessing the association between vertebral deformity and trabecular architecture has some limitations, one being that the structure analysis is conducted at a surrogate site rather than in the spine. Very few studies exist involving measurements of vertebral trabecular architecture as a means for fracture discrimination,(16,20-22) and only one study provides a comparison of the discriminating power of structure measures between vertebral and peripheral sites(16) (a study which involves measurements in intact cadavers only). Furthermore, the latter study is based on conventional structural parameters (rather than topological measures used in this work). The authors find the diagnostic performance, as expressed in terms of the area under the ROC curve to be comparable for direct CT assessment in the vertebrae and νMRI at the distal radius. There is thus, as yet, no conclusive evidence showing superiority of structural measurements at the fracture as opposed to a remote site.

In summary, the authors wish to emphasize the exploratory nature of this work, carried out with emerging technology that is likely to further evolve during the years ahead. Nevertheless, the data showed the potential of the MRI-based VBB to study the structural implications of osteoporosis.

The data obtained in postmenopausal women suggested that several structural measures, derived by νMRI at the distal radius—in particular those relating to the topology of the trabecular bone network—explain up to 30% of the variation in total spinal deformity burden independent of areal BMD of the lumbar vertebrae. Rather than treating vertebral fracture status as a dichotomous quantity, a spinal deformity index serving as a continuous variable to express total vertebral deformity burden was used. The results provide new insight into the structural manifestations of osteoporotic bone loss. Anticipated technological advances are likely to improve the method's robustness as an investigational and clinical tool.


This research was supported by NIH Grants RO1 AR49553, M01RR00240, T32 EB000814, T32DK07006, RR 024134, and UL1 RR024134 CTRC Univ. of PA SOM.