A sub-cohort of postmenopausal women with osteoporosis who were treated with the monoclonal antibody denosumab within the framework of a large randomized controlled trial termed FREEDOM and its open-label extension phase, was investigated [16,17,18]. In order to support the clinical findings of this study in regard to bone structural indices like BSV, a supplementary ex vivo study in human vertebrae was performed using high-resolution quantitative computed tomography (hr-QCT).
The FREEDOM pivotal trial was an international, randomized, placebo-controlled trial in postmenopausal women with osteoporosis. Women aged 60–90 years with a lumbar spine or total hip BMD T-score less than − 2.5 at either site but not less than − 4.0 at both sites were eligible to enroll in this study. Subjects received placebo or denosumab 60 mg subcutaneously every 6 months for 36 months, with daily supplements of ≥ 1000 mg of calcium, and ≥ 400 IU of vitamin D. Overall, data from 7808 women were available in the FREEDOM trial, including 3902 in the denosumab group and 3906 in the placebo group. Denosumab is a fully human monoclonal antibody to the receptor activator of nuclear factor-κB ligand (RANKL), which blocks its binding to RANK, inhibiting the development and activity of osteoclasts, followed by suppression of bone resorption. Details of the study and the main results have been reported previously [16,17,18].
All women who completed the FREEDOM study (i.e., completed their 3-year visit) in either the denosumab or placebo arm were eligible to enter the 7-year open-label Extension Phase of the trial, provided all inclusion criteria were met [17, 18]. In the Extension Phase, all participants were scheduled to receive 60 mg denosumab subcutaneously every 6 months (± 1 month) with daily calcium and vitamin D supplementation. Results from 5, 6, and 8 years of denosumab exposure in women included in the extension trial have been published previously [17,18,19].
For the purpose of the study presented here, postmenopausal women who had been recruited at the Medical University of Graz (one out of 178 participating study centers) and who completed the 3-year FREEDOM study and the 5-year open-label extension were included in the analyses. The study was registered in the European Union Drug Regulating Authorities Clinical Trials (EudraCT) database (2007–001041-17), and approved by the ethics committee of the participating institution (EK-07-146-0807).
BMD at the lumbar spine was assessed by DXA measurement (HOLOGIC QDR 4500; HOLOGIC Inc., Bedford, MA, USA) according to the FREEDOM protocol . Results were expressed in grams per square centimeter. Measurements were performed at the pivotal trial screening visit (year 0) and at year 3 (which was corresponding to the screening visit of the extension phase), and at years 1, 2, 3, and 5 (which correspond to years 4, 5, 6, and 8 of the entire study) of the extension phase (Fig. 2a and b) [17,18,19]. As the Medical University of Graz did not participate in the DXA substudy of the FREEDOM pivotal trial, lumbar spine DXA measurements were not available for years 1 and 2 of the pivotal trial. Furthermore, data from year 7 of the extension phase were not included in the analyses, as official results from this study phase had not yet been published as a peer-reviewed article.
Furthermore, for analyses as conducted in the present study, local read data have been used instead of central data.
According to the FREEDOM pivotal trial study protocol, radiographs of the thoracic and lumbar spine were taken in anterior-posterior (a.p.) and lateral projection at baseline, years 2 and 3 . During the Extension Phase, radiographs of the thoracic and lumbar spine were taken at baseline (which was corresponding to the end of the FREEDOM pivotal trial) and at years 2, 3, and 5 [17, 18]. Vertebrae were locally assessed, visually inspected, and excluded from analysis if fracture was present and height-loss was equal to or exceeded 20 % according to a semi-quantitative classification .
For BSV analysis, digitally stored radiographs were available in 16-bit DICOM format from which date of acquisition, the modality, and the pixel spacing were extracted. However, the current version of the analyzer used (IB LAB TX™ Analyzer, IBL, Vienna, Austria) only allows image processing in 12-bit depth, yielding a gray level range of 0 to 4096. Furthermore, due to the retrospective design of this pilot study and due to the fact that the underlying randomized controlled trial was carried out over a period of 8 years, standardized pixel size was not available. In fact, depending on the detector plates used, the following pixel sizes were available: 100, 111, 114, and 150 μm (Table 1).
As evident from the table, pixel sizes of 150 μm were used only after year 3 of the pivotal trial, raising the possibility that the steep increase in BSV thereafter could be due to the larger pixel sizes used. However, correlation analysis did not show a significant association between BSV and pixel size (R ~ 0.170; p > 0.09).
Defining the region of interest
Digital radiograph images of the lumbar spine were analyzed using a semi-automatic software application (IB Lab TX Analyzer, IBL, Vienna, Austria). Thus, positioning of the ROI involved a two-step procedure as outlined in the following:
In step one, anatomical landmarks were placed manually on the anterior and posterior edges of each of the vertebrae L1 – L4, depending on their respective eligibility (Fig. 1). If only three out of four vertebrae were eligible, then three vertebrae were analyzed, and if only two out of four were eligible, then only two were analyzed. If less than two of the vertebrae L1-L4 were eligible, the corresponding spine radiograph was excluded from further analysis.
Given that the landmarks on each vertebra constitute the technical pillars on which the consecutive, software-driven automated process is based, we sought to quantify potential intra- and interobserver bias. Thus, six different radiographs were studied by three observers (experienced radiologists), and each of these six radiographs was analyzed five times by each observer. The reading was done in a randomly selected order, with a minimum of 24 h between consecutive analyses. Observers were blinded to the respective results they had obtained previously. Both the anterior and the posterior height (mm) were measured for each vertebra, since (in a non-deformed vertebra) the crossing point of the diagonals between, e.g., the upper-posterior and the lower-anterior landmark is mainly determined by these parameters. For each series of five analyses of the same radiograph, a coefficient of variation (CV) was calculated and the results were then averaged, providing the final CV. The mean CV (%) for the intraobserver reproducibility was 0.60 ± 0.16 (SD).
The interobserver reproducibility was assessed by analyses performed by three observers (experienced radiologists) who were not identical with the observers of the intraobserver substudy. A total of six radiographs were analyzed by each radiologist, and the same parameters measured as for the intraobserver substudy. Observers were blinded to the results obtained by the other observers. A coefficient of variation (CV) was calculated and the results were then averaged, providing the final CV. The mean CV (%) for the interobserver reproducibility was 0.68 ± 0.4 (SD).
In step two, based on the manually placed landmarks, the final ROIs were then set automatically so that the center of each vertebra was congruent with the center of the according ROI (Fig. 1). The shape of the ROI was chosen as rectangular and their size was arbitrarily set at 28 × 14 mm, yielding a total of 2613 to 3920 pixels per ROI, depending on the detector plate used. Given that digitally stored radiographs were processed in 12-bit pixel depth, 4096 Gy-levels were available for analysis per ROI. For each ROI, the BSV was then calculated automatically, and the average BSV of all measurable vertebrae (i.e., at least two) was a patient’s final BSV result. However, in order to evaluate if ROI size would affect the BSV results, together with the chosen ROI of 28 × 14 mm, two additional ROI sizes were tested in each of the vertebrae L1-L4, in three randomly selected patients. The largest ROI was set manually so that the largest possible area of trabecular bone in each vertebra was captured, without including the end plates. The smallest ROI was set at 14 × 7 mm using the same semi-automatic software application as introduced above. The three different ROI showed no significant difference in BSV, but a highly significant correlation was observed (pairwise Mann–Whitney–Wilcoxon rank sum test: p > 0.4; one-way ANOVA: p > 0.85; multiple correlation coefficient: 0.99, p < 0.0001).
Derivation and calculation of the BSV
The BSV, as performed in the prevailing study, is based on the principle of a fractal model, which has been described in more detail previously . In short, fractals are a class of mathematical functions which can be used to characterize the geometrical properties of sets . Accordingly, fractals can be applied to relate a metric property such as the length of a line (two-dimensional) or the area of a surface (three-dimensional) to the elemental length or area used as a basis for the calculation . Two-dimensional patterns as present on plain radiographs of trabecular bone are thus well suited to fractal-based analyses. In fact, it has been shown that this two-dimensional information can be well translated into three-dimensional textural information by the use of a fractal model which involves the so-called fractional Gaussian noise, with the latter being the increment of the fractional Brownian motion (FBM) . The only parameter of interest in the FBM is the so-called H parameter, also referred to as the Hurst exponent . In order to assess this parameter, a maximum likelihood estimator must be applied . This very method has been shown to yield high intra- and inter-observer as well as long-term reproducibility, with coefficients of variation being 0.61, 0.67, and 2.07%, respectively . Based on the principles described above, oriented textural analysis software was developed, using grey-level variations together with a specific machine-learning-based algorithm. All vertical and horizontal pixel lines within each rectangular ROI (i.e., depending on the detector plates used: 187–280 vertical lines, and 93–140 horizontal lines, respectively) were analyzed in such way that each pixel (or gray-value) of a line was compared to each pixel (or gray-value) within the same line, yielding a vertical BSV (BSVv) as well as a horizontal BSV (BSVh). The final result for each ROI, and hence its textural characterization, was the BSV value obtained by averaging the BSVv and BSVh results of all lines within the same ROI, and the final result for the respective lumbar spine (or patient) was obtained by averaging the BSV value of all vertebrae included in the analysis. It should be noted that BSV is a unitless value.
Supplementary ex vivo study
In order to test the hypothesis that BSV, as utilized for the clinical pilot study presented herein, provides information on structural indices of (trabecular) bone, we also carried out a supplementary ex vivo study. For this purpose, ten human vertebrae (five taken from a 67-year-old male patient, and five taken from an 87-year-old female patient; provided by the Macroscopic and Clinical Anatomy section of the Gottfried Schatz Research Center for Cell Signaling, Metabolism and Aging of the Medical University of Graz, Austria) was measured using two different technologies: high-resolution computed tomography (HR-pQCT; XTreme CT II, SCANCO Medical, Switzerland), and conventional X-ray (Multix Fusion Max, Siemens Healthcare, Germany).