The incidence of osteoporotic hip fracture increases exponentially with age in both women and men and imposes an increasing societal, economic, and individual burden as the age of the population continues to increase throughout the world.1 Therefore, the prevention and reduction of osteoporotic hip fractures are clearly among the major current and future challenges.1 However, since the cost of preventive measures is high, the best approach consists of identifying patients at high risk of hip fractures and defining the dominant factors of such risk so as to target preventive interventions appropriately. Assessment of fracture risk is complicated additionally by the fact that there are major differences between patients with trochanteric-intertrochanteric fractures and cervical fractures.2–4 For example, patients with trochanteric-intertrochanteric fractures have a more severe and generalized BMD deficit, bone geometry appears to play a lesser role, and morbidity is higher.2–4
In clinical practice, the physician now can assess the risk of hip fracture for an individual patient from several clinical factors, which include age, weight, fracture after age 50 years, parental hip fracture, smoking status, alcohol intake, glucocorticoid therapy, and secondary causes of osteoporosis, with or without areal bone mineral density (aBMD).5, 6 These variables can be pooled into the recent World Health Organization (WHO) FRAX tool (www.shef.ac.uk/FRAX/) or similar algorithms to calculate a 10-year fracture risk.7, 8 Guidelines for treatment have been published in the United States and Europe,9–11 but targeting preventive interventions is still challenging.
A key issue is the difficulty in measuring the mechanical competence of the femur in vivo. This competence is determined by bone structural properties (ie, bone size, bone shape, and cortical and trabecular microstructure) and intrinsic material properties.12, 13 Dual-energy X-ray absorptiometry (DXA) is the established method to measure aBMD, but it does not provide other important features of a whole bone's mechanical competence, such as the 3D geometry of bone.12 aBMD predicts 50% to 80% of the breaking strength of bone in in vitro experiments14–16 and is the best validated in vivo parameter in predicting femoral fractures.17–19 However, in an individual subject, aBMD is still inadequate to predict whether or not the bone will fracture; other factors (eg, whether the patient falls, how she or he falls, and the response to the fall) play an additional role. Also, DXA measures integral aBMD and so does not separately measure cortical and trabecular bone; thus no information of the relative contributions of these two compartments to bone density and strength is provided.20
Separate assessment of cortical and trabecular bone is essential because the two types of bone differ in their age-related changes,21 mechanical roles, and response to preventive and curative treatments for osteoporosis. Trabecular bone is about eight times more metabolically active than cortical bone and is subjected to early and rapid changes with advancing age.21 Among available osteoporotic medications, some reduce trabecular bone loss [eg, bisphosphonates or selective estrogen receptor modulators (SERMs)] but have no effect on cortical bone, whereas parathyroid hormone seems to predominantly enhance cortical bone formation.22 The elastic modulus and ultimate strength of cortical and trabecular bone are different and decrease with advancing age in both men and women.23 However, there is still controversy about the respective contributions of trabecular and cortical bone compartments to hip fracture risk. Some authors have attributed the exponential increase in hip fracture risk with aging to thinning of the cortex of the neck,24, 25 whereas others suggest a synergistic action and a decline in both cortical and trabecular bone.26–28 Mechanical testing of excised femoral necks also has given variable results, ranging from cortical contributions of 40% to 60%14, 26 to almost 100%29 for overall femoral strength. Similarly, finite-element modeling (FEM) has suggested that in the femoral neck region, cortical bone supports 50% of the stresses associated with normal gait or during a fall on the side.27
Currently, only a few in vivo quantitative computed tomographic (QCT) studies investigating hip fractures exist.30–32 Owing to technical limitations of the QCT techniques applied, in particular, to assess cortical bone in the different compartments of the proximal femur, the ability of these studies to identify predictors of femoral fracture also is limited, despite the fact that QCT provides separate assessments of cortical and trabecular bone and measurement of 3D geometry. Therefore, in this 2-year case-control prospective in vivo study we aimed to identify QCT parameters that best discriminate hip fracture in postmenopausal women using an advanced image analysis software for 3D QCT images of the proximal femur that has already been used in in vivo33 and in vitro14 studies. BMD and geometric variables of the proximal femur were obtained separately for cortical and trabecular bone and from different volumes of interest (VOIs), including the femoral head and a proximal shaft VOI, which is predominantly composed of cortical bone. Trochanteric and cervical hip fractures were analyzed independently.
Materials and Methods
The study population consisted of 107 white women recruited in two centers, in Paris, France (center 1), and Manchester, England (center 2). Of the 107 subjects, 47 had acute low-trauma hip fracture (32 in center 1, 15 in center 2) confirmed on pelvic radiographs. Sixty subjects without hip fracture served as controls (26 center 1, 34 center 2). All subjects were 60 years of age or older. A previous hip fracture was an exclusion criterion. History of fracture other than a hip fracture and treatment with osteoporotic medications were not exclusion criteria.
In the patients with hip fracture, QCT of the contralateral hip had to be performed in a short interval between admission to hospital and hip surgery owing to good practice guidelines. DXA scans of the contralateral hip were obtained as soon as feasible following surgery. The non–hip fracture control subjects were drawn from patients referred to the centers who were invited and agreed to participate in the study. Exclusion criteria were hip fractures through preexisting bone lesions (pathologic fractures), hip arthroplasty/replacement on the contralateral side, and disorders that prevented the patient from giving informed consent. Approval for the study was given by the relevant local ethics committees in Paris and Manchester, and all subjects gave written informed consent.
Quantitative computed tomography (QCT)
Center 1 used a four-row CT scanner (Somatom Volume Zoom 4, Siemens, Erlangen, Germany) and a solid calibration phantom (Osteo phantom, Siemens). Center 2 had a 16-row CT scanner (LightSpeed 16, GE Medical Systems, Milwaukee, WI, USA) and a different solid calibration phantom (Mindways, Austin, TX, USA). The phantoms were placed below the patient for simultaneous calibration to convert Hounsfield units (HUs) into BMD. The European Spine Phantom (ESP) was scanned at both centers for cross-calibration.
The region scanned extended from just above the proximal surface of the femoral head to 5 cm below the lesser trochanter. CT acquisitions were performed at 120 kV, 170 mA, 1 second per rotation, pitch of 1, a slice thickness of 1 mm in center 1 and of 1.25 mm in center 2, and constant table height. For reconstructions, center 1 used the B40s kernel and center 2 the standard kernel. Data were reconstructed with a field of view (FOV) of 460 mm to analyze the calibration phantom and with a second FOV of 150 mm, in which the hip analyses were performed.
To minimize differences in QCT results protocols for acquisition and reconstruction, parameters were chosen to be as close as possible in the two centers, taking into account the different manufacturers of the scanners. Most of the BMD differences between the scanners were compensated for by use of the calibration phantom, the characteristics of which were taken into account in the image-analysis process. Remaining field inhomogeneities were corrected for based on the results from the ESP scans. The ESP scans also were used to compare the capabilities of the two scanners to assess cortical thickness. However, no attempt was made to cross-calibrate cortical thickness between the scanners because no validated method has yet been reported.
Using these protocols, the radiation dose [effective dose equivalent (EDE)] was about 2 to 3 mSv using dose-reduction techniques implemented on the scanners.34
The QCT data were analyzed using the dedicated software Medical Image Analysis Framework—Femur option (MIAF-Femur), which was developed at the Institute of Medical Physics (University of Erlangen, Erlanger, Germany). An important advantage of this software is the 3D segmentation that is not based on a global threshold but rather on local adaptive thresholding techniques combined with volume growing and morphologic operations. Another advantage is the automatic determination of a neck coordinate system (NCS), relative to which the volumes of interest (VOIs) analyzed are located.14, 35–37 The analysis is semiautomatic, with only occasional corrections by the operator being necessary.37 In this study, Version 4.0 was used; the processing time for each femur was 15 to 20 minutes (Pentium 4, 1 GB).
The analyzed VOIs comprised the neck, trochanter, intertrochanter, and the proximal shaft (Fig. 1). In each VOI, integral (intg), cortical (cort), and trabecular (trab) BMD and volume (Vol) and cortical thickness (thick) were analyzed. In addition, the 3D neck axis length, measured between the lateral edge of the greater trochanter and the medial edge of the femoral head (NALQCT), and polar and areal moments of inertia of the neck were obtained. Earlier measurements with the European Proximal Femur Phantom (EPFP)38 showed that accuracy errors for trabecular BMD were smaller than 5%. Owing to partial-volume artifacts, accuracy errors for cortical BMD were significantly higher and depended on cortical thickness. In vivo reanalysis precision errors were 1% to 3% for integral, trabecular, and cortical BMD at the various sub-VOIs.
Dual-energy X-ray absorptiometry (DXA)
Both centers used a Prodigy DXA scanner (GE Lunar, Madison, WI, USA). Image acquisition and analysis were performed as recommended by the manufacturer. Areal BMD (in g/cm2) was measured at the neck, trochanter, and intertrochanteric region and total femur. Ward's area, which is not established as a region of interest for study, was not used.
DXA hip axis length (HALDXA) was measured from the lateral edge of the greater trochanter to the medial edge of the femoral head.
The same ESP was scanned three times in both centers: The mean difference in aBMD in the area L1–L3 was 0.00567 g/cm2, which corresponded to a 0.52% difference. Additionally, both devices had regular quality-control procedures carried out thoughout the study period to ensure absence of drift.
Data were summarized as means and SDs. Pearson correlation coefficients (r) were computed to show age trends in densitometric and geometric values. Group means were compared using Student t tests or, if not applicable, the Mann-Whitney U test. Analysis of covariance was used to compare data after adjusting for age, weight, and height. Multivariable best subset selection logistic regression was performed from the set of QCT variables to identify independent correlates of the presence of hip fractures. Next, multivariable stepwise logistic regression was performed on DXA variables to identify independent correlates of the presence of hip fractures. The odds ratios (ORs) were standardized for a 1 SD decrease and reported with 95% confidence intervals (CIs). The areas under the receiver operating characteristic curve (AUCs) were used as discriminative performance criteria of the models. They are reported with their corresponding 95% CIs. To test the robustness of the models, internal validations were performed for each model using bootstrap procedures39 (1000 bootstrap samples) to estimate overoptimism associated with AUC, and the Brier score was calculated [model scores range from 0 (perfect) to 0.25 (worthless)]. The described procedures were applied to all hip fractures and also separately for cervical and trochanteric fractures. Finally, because some controls and some hip-fractured patients had previous non–hip fractures, a sensitivity analysis was performed, including the presence/absence of a previous fracture after age 50 years (ie, wrist, humerus, and vertebra) as a covariate in the QCT and DXA models. The aim was to examine whether the variables identified in the models are robust to that information (to check that a history of fracture does not change the significance of any of the variables identified by the models).
AUC values were compared between the QCT and DXA models.40 This comparison was performed for the subgroup of patients that had both CT and DXA.
Finally, we computed combined QCT-DXA models (using the previously statistically selected variables) to determine whether combining the two methods was more powerful than DXA alone. The discriminative performances again were compared using the AUC values.
All analyses were conducted using SAS 9.1 software (SAS Institute, Cary, NC, USA) and R 2.8 software (www.R-project.org); p values smaller than .05 were considered significant.
Study population and hip fracture characteristics
Patient demographics are given in Table 1. The hip fracture group was significantly older and had a significantly lower body mass index (BMI; p = .047). Height and weight were not significantly different. A history of fractures after age 50 years at sites other than the hip was present in 21 non-hip-fractured (control) subjects (ie, 13 wrist, 4 vertebral, and 2 humerus; two subjects each had a wrist and a vertebral fracture) and in 10 hip-fractured patients (ie, 3 wrist, 2 vertebral, and 2 humerus; three patients each had two fractures (ie, wrist and humerus 1, humerus and vertebra 1, wrist and vertebra 1). Fifteen non–hip fracture and two hip fracture women were receiving osteoporotic medication [ie, non–hip fracture subjects: bisphosphonates (n = 13) and raloxifene (n = 2); hip fracture patients: bisphosphonates (n = 1) and raloxifene (n = 1)]. Such information was not available in the 11 patients who did not attended for DXA.
Table 1. Characteristics of Subjects (Means, SDs in Parenthesis, Ranges in Italic)
n = number of subjects; BMI = body mass index = weight/height2; NS = nonsignificant.
p Values result from Student's t test.
73.4 (9.07) (n = 60)
81.59 (10.98) (n = 47)
157.06 (7.07) (n = 59)
157.17 (7.43) (n = 41)
64.73 (12.13) (n = 59)
60.17 (12.12) (n = 42)
26.27 (4.87) (n = 59)
24.37 (4.35) (n = 42)
The hip fracture type was determined from radiographs. There were 24 intracapsular (cervical) and 23 extracapsular (trochanteric) fractures. Mean age (SD) did not differ significantly between the two hip facture subgroups [intracapsular fracture 79.2 (12.1) years; extracapsular 84.1 (9.3) years], nor did height, weight, or BMI.
QCT was performed in all 107 women. DXA was performed in 59 non–hip fracture subjects, 19 patients with cervical fractures and 18 with trochanteric fractures. In control non–hip fracture subjects, QCT and DXA were performed on the same day. In hip fracture patients, QCT had to be performed before surgery and was obtained within the first 24 hours in 43 patients, on day 2 in 2, on day 3 in 1, and on at day 6 in 1. The mean separation between hip fracture patients' QCT and DXA was 11.9 days (<7 days, n = 17; between 8 to 14 days, n = 9; >15 days, n = 10; 1 month, n = 3; 2 months, n = 2). DXA could not be performed in 10 patients following hip fractures because of death or because their clinical conditions precluded attendance. In one patient, the DXA had been performed 3 months before the hip fracture and was not repeated.
QCT and DXA results
The cross-calibration of the two CT scanners used in Paris and Manchester was performed by applying different field inhomogeneity corrections (FICs), originating principally from differences in beam hardening across the CT scanner field of view. In Manchester, the CT scanner specific field inhomogeneity resulted in a 10% BMD overestimation of the 50 mg/cm3 insert of the L1 vertebra of the ESP and a 3% BMD underestimation of the 200 mg/cm3 insert (L3 vertebra). In Paris, L1 BMD was underestimated by 10% and L3 BMD by 3%. After FIC, differences between scanners were reduced to 1.8 mg/cm3 for the L1 insert and to 0.8 mg/cm3 for the L3 insert.
The impact on cortical thickness was evaluated using the L2 and L3 vertebrae of the ESP, with a nominal thickness of 1 and 2 mm, respectively. Scans in Manchester resulted in thickness measurements of 1.7 and 2.6 mm and in Paris of 1.4 and 2.2 mm. This is mainly due to the larger slice thickness (1.25 versus 1 mm) on the GE scanner in Manchester and the different reconstruction kernels.
Integral BMD in the head, neck, trochanter, intertrochanter, and shaft decreased significantly with advancing age (r2 = 0.58 to 0.62, p < .0001; for the shaft, r2 = 0.28, p = .005), as did trabecular BMD (r2 = 0.48 to 0.55, p < .0001). The same significant trend was observed with cortical BMD in the trochanter (r2 = 0.44, p < .0001). DXA aBMD of the neck, trochanter, intertrochanter, and total femur decreased with advancing age (r2 = 0.49 to 0.53, p < .0001). Cortical thickness in the neck, trochanter, intertrochanter, and shaft decreased significantly with advancing age (r2 = 0.32 to 0.38, p < .0008). Integral volume of the shaft increased with advancing age (r2 = 0.32, p = .001). No significant age-related change was found with other bone volume or hip axis length (measured by either QCT or DXA).
QCT results are shown in Table 2. Integral BMD values in the head, neck, trochanter, and intertrochanter were significantly lower in the hip fracture group by 23.1%, 16.8%, 25.0%, and 19.5%, respectively. The major contributor was the trabecular BMD component, with percentage differences of 41.4%, 50.9%, and 45.1% for neck, trochanteric, and intertrochanteric VOIs, respectively. Apart from the trochanteric VOI, cortical BMD was not significantly different. Integral BMD values in the head, neck, trochanter, intertrochanter, and shaft VOIs were lower in the trochanteric than in the cervical fracture group, but not significantly. Integral BMD mean values (and SDs) ranged between 180.88 (50.56) and 572.37 (173.77) in patients with cervical fractures (in the trochanteric and shaft VOIs, respectively). These values ranged between 161.12 (47.55) and 495.29 (136.64) in patients with trochanteric fractures (in the trochanteric and shaft VOIs, respectively).
Table 2. Quantitative Computed Tomographic (QCT) Densitometric and Geometric Variables (Means and SD in Parenthesis), All Hip Fractures Included
Controls (n = 60), means (SD)
Hip fracture subjects (n = 47), means (SD)
Mean difference (% difference between means)
p Value (adjusted p value)
Note: p Values result from Student's t tests or, if not applicable, from Mann-Whitney U tests. In six subjects (five hip-fractured subjects, one control subject), the proximal shaft VOI could not be analyzed because the scan ended just below the lesser trochanter. Head, Neck, Troch, Shaft = volumes of interest analyzed in QCT, head, neck, trochanter, and proximal femoral shaft, respectively; Cort = cortical bone compartment; Trab = trabecular bone compartment; Intg = integral bone compartment; NALQCT = 3D neck axis length measured using QCT; Thick = cortical thickness; PMIM and AMIM = polar and areal moments of inertia of the neck, respectively.
Cortical thickness was significantly thinner in the neck, trochanter, intertrochanter, and proximal shaft VOIs in the hip fracture group. The percentage differences between mean values for the neck, trochanteric, intertrochanteric, and proximal shaft thicknesses were, respectively, 4.31%, 4.44%, 5.73%, and 13.22%. Volume of integral bone in the shaft was significantly larger in the hip fracture group than in controls but was not significantly different in the neck, trochanteric, and intertrochanteric regions. Areal and polar moments of inertia were significantly lower in hip fracture patients than in the non–hip fracture controls (p = .02 and p = .009, respectively).
Among all geometric variables obtained in the integral, cortical, and trabecular bone, only one, the volume of cortical bone in the trochanteric region, was significantly different between the two types of hip fractures (ie, lower in the trochanteric fracture group, p = .04). However, all cortical thicknesses and most volumes had lower values in trochanteric fractures than in cervical fractures.
DXA results are shown in Table 3. DXA aBMD values of the neck, trochanter, intertrochanter, and total femur were significantly lower in hip fracture patients by between 15% and 18% (p < .0001). HAL did not differ significantly between the two groups. Mean aBMD (in g/cm2 and SDs) for cervical and trochanteric fracture patients were in the neck, trochanter, intertrochanter, and total femur 0.67 (0.09) to 0.65 (0.12) g/cm2, 0.55 (0.09) to 0.54 (0.09) g/cm2, 0.84 (0.17) to 0.82 (0.15) g/cm2, 0.70 (0.11) and 0.68 (0.11) g/cm2, respectively. There was no significant difference in aBMD values of the neck, trochanter, intertrochanteric regtion, or total femur between the two subgroups of hip fractures.
Table 3. Dual-Energy X-ray Absorptiometry (DXA) Variables (Means and SDs in Parentheses), All Hip Fractures Included
Using logistic procedures to discriminate the hip fracture event, three different QCT models combining two variables were obtained. The three models are provided in Table 4, without and with age, height, and weight adjustment. All three models were found to combine one BMD value, either integral BMD of the head (which is a measure of trabecular BMD) or trabecular BMD of the trochanter, with a cortical thickness value, either the mean cortical thickness of the neck (CortNeckThick) or the cortical thickness of the shaft (CortShaftThick). AUCs, given in Table 4, ranged from 0.80 to 0.84 and were not significantly different (p = 0.70). Table 4 also provides the optimism of the models' performances and the Brier scores.
Table 4. Logistic Procedures (Before Adjustment and in Italic after Age, Weight, and Height Adjustment) Using Quantitative Computed Tomography (QCT) Variables
Notes: Odds ratios (OR) are provided for 1 SD decrease. BMD = bone mineral density; AUC = area under the receiver operating characteristic curve; Head, Neck, Troch, Shaft = volumes of interest analyzed in QCT ofr head, neck, trochanter, and proximal femoral shaft, respectively; Cort = cortical bone compartment; Trab = trabecular bone compartment; Intg = integral bone compartment; NALQCT = 3D neck axis length measured using QCT; Thick = cortical thickness.
Cervical hip fractures
2.648 [1.511, 4.640]
0.770 [0.666, 0.874]
3.032 [1.396, 6.587]
0.810 [0.710, 0.909]
Trochanteric hip fractures
3.412 [1.585, 7.345]
0.866 [0.784, 0.948]
2.750 [1.294, 5.846]
3.275 [1.237, 8.670]
0.875 [0.793, 0.958]
2.429 [1.007, 5.857]
When only trochanteric fractures were considered, AUCs were 0.87 (0.78, 0.95) and 0.88 (0.79, 0.96) for nonadjusted and adjusted models, respectively, with the combination of the mean thickness and BMD of the trochanteric VOI. For cervical fractures, AUCs were lower: 0.77 (0.67, 0.87) and 0.81 (0.71, 0.91) for nonadjusted and adjusted models, respectively (Table 4B).
For DXA, aBMD of the total hip (aBMDtot) was the best discriminant variable of hip fracture risk when all hip fractures were included [AUCs were 0.78 (0.69, 0.87) and 0.80 (0.72, 0.89) for nonadjusted and adjusted models]. For trochanteric fractures, AUCs were 0.79 (0.68, 0.90) and 0.85 (0.76, 0.94) for nonadjusted and adjusted models (Table 5). aBMD of the neck (aBMDneck) was the best discriminant variable for cervical fractures [AUCs were 0.76 (0.64, 0.88) and 0.77 (0.65, 0.90) for nonadjusted and adjusted models] (Table 5).
Table 5. Logistic Procedures (Before Adjustment, and in Italics after Age, Weight, and Height Adjustment) Using Dual Energy X-ray Absorptiometry (DXA) Variables
aBMD = areal bone mineral density; AUC = area under the receiver operating characteristic curve.
Total femur aBMD
3.070 [1.523, 6.190]
0.790 [0.680, 0.900]
Total femur aBMD
3.090 [1.260, 7.578]
0.849 [0.760, 0.939]
No significant differences between DXA and QCT hip fracture discrimination was observed when considering all hip fractures (p = .49 and p = .71 when comparing QCT model 1 and DXA model, without and with adjustment, respectively) (Fig. 2) or hip fracture subgroups separately (p > .5).
The combination of the variables statistically selected in the previous QCT and DXA models was found to improve hip fracture discrimination significantly for QCT model 1 (IntgHeadBMD and CortShaftThick) (p = .029) (Table 6 and Fig. 2). The significance for QCT model 3 was borderline (p = .0502).
Table 6. Logistic Procedures (Age, Weight, and Height Adjustment) Using Quantitative Computed Tomographic (QCT) Variables (Models 1, 2, and 3) or Dual-Energy X-ray Absorptiometry (DXA) Variables or a Combination of QCT and DXA
6A: QCT Model 1 (IntgHeadBMD and CortShaftThick)
QCT model 1
QCT + DXA
6B: QCT Model 2 (IntgHeadBMD and CortNeckThick)
QCT model 2
QCT + DXA
6C: QCT Model 3 (TrabTrochBMD and CortTrochThick)
Notes: The AUC of the QCT-DXA model was compared with the AUC of DXA alone. BMD = bone mineral density; Thick = cortical thickness; Head, Neck, Troch, Shaft = volumes of interest analyzed in QCT of head, neck, trochanter, proximal femoral shaft; Cort = cortical bone compartment; Trab = trabecular bone compartment; Intg = integral bone compartment; AUC = area under the receiver operating characteristic curve.
QCT model 3
QCT + DXA
Because 21 non-hip-fractured (control) subjects and 10 hip-fractured patients had previous non–hip fractures, the presence/absence of previous fracture was included as a covariate in the QCT and DXA models. This additional information did not changed the significance of any of the variables identified by our models.
We investigated the in vivo discrimination of hip fracture using QCT. The analysis was performed using MIAF-Femur software, a powerful dedicated image-analysis tool. We further provided insights into the mechanisms underlying hip fracture in elderly women.
We found that a combination of one trabecular BMD parameter with one cortical thickness parameter was particularly powerful in discriminating hip fracture. This result stimulates two threads of thought. First, the best discrimination was obtained by a combination of one density parameter (ie, a bone quantity) and one geometric parameter. This is in perfect agreement with several in vitro14, 15, 41, 42 and in vivo studies.41, 43 Second, the best discrimination was obtained by combining one parameter representative of the trabecular compartment and one representative of the cortical compartment.
As covered in the Introduction, the respective contribution of the trabecular network and the cortical shell to hip fracture risk is still debated. Our results show that cortical and trabecular bone act in synergy to provide mechanical competence to the proximal femur, which is in agreement with Lotz and colleagues, who showed by using finite-element analyses (FEA) that 50% of the applied load during gait or a sideways fall was supported by the trabecular bone.27 In this study, the differences in trabecular BMD between hip fracture and non–hip fracture subjects were much higher than the differences in cortical BMD. This highlights the importance of increasing trabecular BMD to prevent fractures in particular because age-related BMD decrements of trabecular bone are also higher than those in cortical BMD.21 However, from our in vivo study, the importance of cortical bone for hip fracture discrimination is difficult to assess. On a percent basis, cortical BMD differences between the hip fracture and non–hip fracture groups were small compared with the trabecular differences. However, cortical bone quality (encompassing, for example, tissue mineralization or abundance of microcracks), an important determinant of cortical bone strength,44 and some technical limitations of in vivo QCT to measure cortical BMD accurately (see discussion below) may explain why cortical BMD was not a contributing variable in any of the three QCT models.
With QCT, the determination of cortical thickness and of total and cortical bone volume may be more promising than the determination of cortical BMD. With increasing age, at least in the femoral neck, cortical BMD decreases, and cortical thinning, owing to endosteal resorption, is accompanied by compensatory periosteal apposition.21 In our study, we could not detect differences in bone volume between the hip fracture and non–hip fracture groups in either VOI (with the exception of the shaft VOI), although cortices of the neck, trochanteric and intertrochanteric regions, and shaft were significantly thinner in the hip fracture group. Since cortical thinning without periosteal expansion causes a larger decrease in bone strength compared with a similar percent homogeneous decrease in cortical BMD, it is not surprising that cortical thickness and not cortical BMD was included in our QCT models. Moments of inertia calculated in the neck did not contribute to the QCT models, although cortical and trabecular BMD and bone geometry contribute to their calculation. However, since neck BMD played no significant importance in discriminating between hip fracture patients and non–hip fracture controls, the variation in QCT model 2 was captured by cortical neck thickness alone.
A very important finding in our study was that BMD of the femoral head was the most powerful hip fracture discriminator, particularly of those in the cervical region. There is growing evidence that the femoral head, although not a site of osteoporotic fracture, is a powerful discriminator of hip fracture risk. In another in vivo QCT study, hip QCT images of 37 patients with hip fracture and 38 age-matched controls were transformed into a voxel-based statistical atlas.45T-tests were performed between the two groups to identify which regions were most different. The authors found that BMD differences were localized mainly in three small regions, one of them being the femoral head and the two others being more classically the neck and trochanter.45 In another study from the same group,46 bone loss associated with long-duration space flight also was found to occur preferentially in the femoral head.
In an experimental study using radiographs and a combined analysis of trabecular bone structure and bone geometry to discriminate hip fracture load, the femoral head appeared to be the most sensitive for trabecular resorption in cervical fracture and the trabeculae of the greater trochanter for trochanteric fracture,47 which reflects our results exactly. These original and consistent findings clearly emphasize the need for assessing the femoral head to predict hip fracture. In this, QCT has a great advantage over 2D projectional methods such as DXA or radiographs, in which the femoral head and acetabulum are superimposed. However, even though QCT inherently provides images with a clear visual separation of the femoral head from the acetabulum, image analysis remains challenging,36 probably explaining why previous clinical QCT studies did not investigate the femoral head VOI.31
In our study, the ratio of trochanteric to cervical fractures was close to 1, with a mean age of 81.6 years for the hip fracture patients. This is in agreement with data from a meta-analysis of 36,000 hip fracture cases by Baudoin and colleagues,48 who found in women over 60 years of age a progressive increase in the number of trochanteric fractures with advancing age, reaching a ratio of 1 in the very elderly. Records of incident hip fractures in Sweden between 1985 to 2000 also showed an increase in the number of trochanteric compared with cervical fractures.49 Trochanteric fractures also have been shown to be associated with increased mortality when compared with cervical fractures, and this could not be explained by differences in age or comorbidity.4
Since there is evidence in both in vitro14, 15, 42, 47, 50 and in vivo3, 30 studies that cervical and trochanteric fractures have different risk factors, we compared densitometric and geometric variables between the two subgroups and analyzed hip fracture discrimination separately for each subgroup. Compared with cervical fracture cases, we found lower DXA aBMD and QCT BMD values in trochanteric fracture cases, but differences were not statistically significant. The only significant difference was a lower cortical bone volume in the trochanteric region in trochanteric fracture. In our study, a model combining trochanteric BMD and mean trochanteric cortical thickness provided the best discrimination of trochanteric fractures, whereas the best discriminant of cervical fractures was BMD of the femoral head. These results are in agreement with current knowledge that women with trochanteric fractures have a more severe and generalized low bone density involving particularly the trabecular component, with lower aBMD in the trochanteric region and lumbar spine.51 Cervical fractures seem to depend more on proximal femur or pelvic geometry (such as a larger femoral neck shaft angle or a longer hip axis length in cervical fractures2, 3, 30, 50) combined with focal bone loss.51
In the EPIDOS study, aBMD values of novel upper and lower femoral neck subregions were lower in trochanteric fractures than in controls, but prediction of cervical fractures was improved by measuring aBMD only in the upper subregion of the femoral neck.3 It is well established that bone density measurements made at the specific fracture site provide stronger fracture risk prediction.2, 17, 18, 30, 51 Using statistical parametric mapping, Li and colleagues45 demonstrated recently the existence of a small region between the femoral neck and head with significantly lower BMD values in cervical compared with trochanteric fracture cases. Also, a small region of the trochanter was shown to contain voxels with lower BMD values in trochanteric fractures.45
Although in our study QCT better discriminated hip fracture than did DXA, which is the standard method for clinical evaluation of fracture risk, the difference was not statistically significant, in agreement with previous clinical52 and experimental studies.14, 42 QCT, which measures cortical and trabecular compartments separately and densitometric and geometric variables and cortical thickness in various sub-VOIs of the proximal femur, logically may capture more information relevant to bone strength than does DXA. However, although QCT better discriminates hip fracture than DXA, unless significant superiority can be demonstrated, QCT will not substitute for DXA because of limited accessibility and higher ionizing radiation dose. For this in vivo protocol and using the dose-reduction techniques provided by the scanner manufacturers, the effective-dose values calculated in two human cadavers using thermoluminescent dosimeters varied between 2 and 3 mSv,34 which is high compared with the 1 to 5 µSv of a DXA hip scan. To increase the clinical acceptability of 3D QCT of the hip, further optimization of acquisition protocols with lower radiation doses is required, particularly for application in younger subjects than those in our study.54
QCT should be considered as a powerful tool for studying the etiology of hip fractures and targeting the impact of therapeutic and nonpharmacologic interventions in osteoporosis. In addition to bone-strength assessment, QCT can be used to measure reduction in subcutaneous fat and muscle cross-sectional area and increased muscle adiposity. These parameters relate to declining physical function in the elderly, with a resulting increased risk of falling and a decreased capacity to protect bone from impact.55 Combining these variables with BMD and geometric assessment of the proximal femur has been shown to improve hip fracture prediction compared with bone measures alone.56
One of the strengths of our study is MIAF-Femur because there is a paucity of analysis tools available for QCT. The tool has been developed over 10 years of collaborative research at the Institute of Medical Physics in Erlangen and the Experimental Radiology Laboratory of the Lariboisière–St-Louis Research and Training Unit in Paris. The tool is based on advanced 3D local adaptive segmentation techniques to outline the periosteal and endosteal bone surfaces. It does not use a global threshold for which cortical measurements are highly dependent on the BMD and thickness of the patient being studied. The tool provides a 3D evaluation of both BMD and bone geometry, with separate evaluations of cortical and trabecular compartments. Variables are measured within a large variety of VOIs that are highly reproducibly located using a 3D neck coordinate system. Some of the VOIs have been chosen to reproduce DXA regions of interest (ROIs); others, such as the head or proximal shaft VOI, have not been implemented previously in QCT analysis software. Additionally, the performance of the software has been tested extensively.57
Our study has several limitations. First, the two groups were not matched for age and BMI. The initial plan was to age-match patients and controls, but despite the long recruitment period, we were not able to enroll elderly controls and young hip fracture patients. The potential effects on the results have been reduced by the age, weight, and height adjustments of our discriminative models. Also, the non–hip fracture controls had a history of more non–hip fractures (n = 21) than the patients with hip fractures (n = 10). However, this would cause an underestimation in our study of the discriminatory power of QCT.
Second, the study population was small relative to the number of variables, particularly when comparing the two hip fracture types, resulting in a lack of significance of several interesting comparisons. However, studying hip fracture patients is always challenging because these patients are elderly and frail, and CT cannot delay surgical treatment, which good practice guidelines advise should be performed as soon as clinically feasible.58 Also, validating our models by using random subsets (bootstrapping) provided support for the variables selected in the logistic regression models. Brier scores (measure the accuracy of a set of probability assessments) ranged between 0.14 and 0.20. They indicate a weak discriminative power of QCT for hip fracture. This is probably explained by the limited power of the study because the Brier scores for QCT are lower (more discriminative), so better, than those for hip DXA aBMD, which is established to be a relatively good predictor of hip fracture.
Third, as a consequence of practical clinical issues, the mean delays between hip fracture and QCT or DXA were different. DXA was performed several days after the hip fracture, and aBMD results could have been affected by disuse bone loss, which is known to occur after immobilization. However, this would artefactually increase the capability of DXA to discriminate between hip-fractured and non-hip-fractured women, and if circumstances could have been avoided for this, then the discriminatory capabilities of QCT would have been enhanced.
Fourth, the comparison of cortical thickness measurements between the two scanners, which were based on the ESP, showed some differences. Since the tomographic reconstruction kernel and the spatial resolution differed between the two scanners, in theory, a cross-calibration for geometric measurements should be performed. However, a cross-calibration of cortical thickness is very difficult because phantom results cannot be extrapolated easily to patients owing to the facts that (1) the correction depends on thickness (in particular, for nominal thicknesses of below 1 mm, which are more affected by partial-volume artifacts) and (2) because of the anisotropic spatial resolution, the corrections depend on the position of the bone structure relative to the scan direction (ie, will be different in the neck and the shaft). To our knowledge, an appropriate cross-calibration of cortical thickness (that certainly requires more than a simple multiplicative factor) has not yet to be reported.
Fifth, we did not consider factors related to fall biomechanics, although they are known to play an important role in the etiology of hip fracture.59, 60 Sixth, we measured trabecular and cortical bone quantity and its distribution at the organ level but could not address the other important aspect of fracture risk assessment of bone quality that relates to bone tissue properties.13
Finally, there is an inherent accuracy limitation of cortical BMD and thickness measurements using QCT owing to partial-volume artifacts.61 Even at the relative small voxel size of 0.300 × 0.300 × 1.0 mm3 (0.300 × 0.300 × 1.25 mm3 in center 2) used in this study, a thin cortex may result in underestimation of cortical BMD and overestimation of cortical thickness. The femoral neck is particularly affected by this issue because the neck cortex is relatively thin and in an oblique plane relative to the CT acquisition plane. We used an optimized segmentation method that is superior to global thresholding. However, the CT technique used enables an accurate assessment of cortical thickness down to 1 mm. Prevhral and colleagues showed that with spiral CT scanners, cortical thickness can be determined accurately to a true thickness of 1 mm but that for an accurate cortical BMD determination, a minimum cortical thickness of 2 mm is required.62 However, even if absolute cortical thickness and BMD cannot be measured accurately because the cortex is too thin, the relative differences between two consecutive measurements are accurate when the difference exceeds a certain threshold.62 Since we observed significant differences in some cortical thicknesses in hip fracture patients and control subjects, the in vivo diagnostic value of cortical measurements in hip fracture risk evaluation using QCT clearly exists. The contribution of cortical, compared with trabecular, bone in hip fracture discrimination using QCT is limited by the lack of accuracy in cortical bone evaluation.
In conclusion, in this in vivo case-control study of hip fracture, our results support the view that trabecular and cortical bone compartments act in synergy to provide mechanical competence to the hip in elderly women. QCT hip fracture discrimination was not significantly higher than DXA discrimination, but further studies are required with larger study populations. There is also a need for high-quality image-analysis software to be available for 3D QCT to have a clinical role in identifying subjects with the high risk of hip fracture and to be used as a tool to study the effects of disease and interventions in osteoporosis.
All the authors state that they have no conflicts of interest.
This work was supported by a EU grant (Contract Number QLK6-CT-2002-02440-3DQCT). The data were presented in part at the ASBMR 2009 and the IOF 2010 conferences.
We thank the Assistance Publique-Hôpitaux de Paris, the Parisian technologists involved in DXA and QCT scan acquisition, especially Patricia Pierre, Max Moutai, Samira Feroudj, and Patricia Bouaziz, and the staff of the emergency room and orthopedic department in Lariboisière Teaching Hospital. In Manchester, the QCT scans were performed by Mrs Jo Anderson, the database was prepared by Mr Mike Machin, and the study was executed in collaboration with Mr Nick Kenny, orthopedic surgeon, Manchester Royal Infirmary.