Osteoporosis is globally highly prevalent and characterized by low bone mass and micro-architectural deterioration, predisposing mostly older individuals to fragility fractures [1]. Fractures can affect multiple regions of the body, but hip fractures are the second most common ones following vertebral fractures and entail high morbidity and mortality, thus making up a high socio-economic burden [1]. To diagnose osteoporosis and assess the individual fracture risk, imaging methods including dual-energy X-ray absorptiometry (DXA) and quantitative computed tomography (CT) are used in clinical routine [2,3,4,5]. While DXA represents the widely used reference standard technique for bone densitometry by measuring areal bone mineral density (aBMD), quantitative CT is an alternative that generates three-dimensional (3D) datasets enabling volumetric BMD (vBMD) assessments [2, 3, 5].

The 3D nature of a CT dataset provides the opportunity to perform more advanced analyses beyond extraction of vBMD [2, 4]. This is of high clinical importance as a patient’s susceptibility to fragility fractures cannot solely be explained by decreased vBMD, because bone strength and resistance to fractures are also determined by other parameters such as bone geometry and microstructure with different contributions from the trabecular and cortical bone compartments [6,7,8]. To date, the multi-parametric capability of CT at the proximal femur has been extended beyond vBMD calculations to estimates of bone strength with finite element modeling (FEM) [9], dedicated measurements of cortical bone thickness (Ct.Th) [10], as well as to statistical models of shape and imaging features [11], and to localized parametric comparisons using computational anatomy approaches such as statistical parametric mapping (SPM) [10, 12,13,14]. While those approaches including underlying segmentations of the proximal femur are oftentimes performed in isolation and are labor-intensive, an automated framework to perform multi-parametric assessments of the proximal femur with extraction of major determinants of bone strength has been introduced [10].

However, quantitative CT for diagnosing osteoporosis and assessing the individual fracture risk comes at the cost of higher radiation exposure to the scanned patient when compared to DXA, which becomes even more relevant in the common clinical scenario when multiple measurements are needed due to reevaluation purposes or therapy monitoring [2,3,4, 15]. Consequently, strategies to reduce radiation exposure from imaging and increase patient safety are highly desirable but need to be carefully balanced against sufficiently high image quality and accuracy of CT-derived measures [16, 17]. Reductions of radiation dose can be established, among other options, with systematically lowered tube current or by acquisitions of fewer projections during scanning, which is referred to as sparse sampling [18,19,20,21,22,23]. When brought together with advanced image reconstruction algorithms such as statistical iterative reconstruction (SIR), considerable radiation dose reductions may become feasible while assuring sufficient image quality [18,19,20, 24]. In this context, it has been shown that the application of sparse sampling with SIR to multi-detector CT (MDCT) images with a reduction down to 10% of original projections may still produce robust integral (cortical + trabecular) vBMD values with small and clinically acceptable bias (e.g., + 6 mg/cm3) compared to full-dose imaging for measurements at the femoral neck [25]. On a similar note, a radiation dose reduction of 50% through sparse sampling with SIR may be established for FEM-based estimation of femoral failure load [20]. However, it has not been investigated to date whether and to what level strategies for radiation dose reduction may be applied in keeping with the requirement of accurately delivering major determinants of bone strength (i.e., vBMD and Ct.Th) when considering the entire proximal femur, its compartments (trabecular and cortical), and its subregions at multiple scales.

Against this background, the hypothesis of this study was that sparse sampling with SIR applied to MDCT of the proximal femur allows for reductions of radiation exposure of 50% or more, while still producing accurate determinants of bone strength across the whole proximal femur. Therefore, we aimed to assess the independent effects of both sparse sampling and tube current reductions with SIR on MDCT-based calculations of Ct.Th and cortical and trabecular vBMD in the total proximal femur, in three of its subregions (femoral head, femoral neck, and trochanteric region), and at a local level using the SPM framework.

MethodsStudy design and participants

This retrospective single-site study was approved by the local institutional review board (registration number 5022/11-A1) and conducted in accordance with the Declaration of Helsinki. The requirement for written informed consent was waived due to the retrospective study design.

Patients above the age of 50 years who had undergone thoraco-abdominal MDCT during clinical routine and following clinical indications were included and identified by searching our institutional Picture Archiving and Communication System (PACS). The exclusion criteria were (1) any known history of malignant bone lesions (such as bone metastases or primary bone cancer), (2) hematological disorders, (3) any history of proximal femur fractures, (4) presence of hip endoprostheses, and (5) any known metabolic bone disorders aside from osteoporosis. In total, 40 subjects were included (26 males and 14 females, mean age ± standard deviation [SD] 69.1 ± 10.0 years). This cohort (except for one patient, who was excluded due to image processing issues) has been previously investigated to determine the effects of methods for radiation dose reduction on integral vBMD measurements at the femoral neck [25]. In brief, this previous study used simulated low-dose scans by lowering tube currents and applying sparse sampling, and integral vBMD values were then compared between different dose levels, numbers of projections, and image reconstruction approaches [25]. Application of sparse sampling with a reduction down to 10% of projections of original scans delivered robust values, with clinically acceptable relative changes [25]. Hence, this previous study did not investigate other parameters beyond integral vBMD, did not investigate the entire proximal femur together with its subregions, and did not provide detailed spatial assessments [25].

Imaging by multi-detector computed tomography

All MDCT scans were performed during clinical routine and due to clinical indications (e.g., oncological staging) in one tertiary care center (academic medical center). None of the patients provided more than one MDCT dataset to the present study (i.e., 40 scans from 40 different patients retrospectively included). Scanning by MDCT covered the proximal femur of both sides at least down to the lesser trochanter in all subjects [25]. Imaging was performed with a 256-row MDCT scanner (iCT; Philips Healthcare, Best, The Netherlands) and with a standard reference phantom (Mindways Osteoporosis Phantom; Austin, TX, USA) that was placed beneath the patients for calibration purposes. Tube voltage and rotation time were 120 kVp and 0.4 s, the pitch ranged from 0.59 to 0.91, and the overall maximum tube current during the thoraco-abdominal scans ranged from 200 to 400 mA (exact tube current implicitly modulated by the scanner, tube current at the proximal femur ranging between approximately 100 and 200 mA) [25]. The imaging examinations were performed after the administration of an intravenous contrast agent (Imeron 400; Bracco, Konstanz, Germany) using a high-pressure injector (Fresenius Pilot C; Fresenius Kabi, Bad Homburg, Germany) and considering a delay of 70 s, a flow rate of 3 mL/s, and an individual body weight-dependent dose (i.e., 80 mL for body weight < 80 kg, 90 mL for body weight between 80 and 100 kg, 100 mL for body weight > 100 kg). In addition, all patients drank approximately 1000 mL of an oral contrast agent (Barilux Scan; Sanochemia Diagnostics, Neuss, Germany) over a period of about 1 h prior to scanning. The effective dose of MDCT scans ranged between approximately 3.6 and 9.1 mSv [25].

Image data analysis Low-dose simulations and statistical iterative reconstruction

Low-dose scans were simulated by using (1) virtual tube current reduction and (2) sparse sampling in all patients using in-house developed scripts. The simulation algorithm to generate lower tube currents for MDCT scans was based on raw projection data, and MDCT system parameters were known to take the electronic noise into account [18, 19, 21, 22, 25, 26]. Furthermore, the raw imaging data of the examinations were exported directly from the MDCT system and used for simulations of sparse-sampled imaging. For sparse sampling, the number of projections per full rotation was systematically reduced, while other parameters (e.g., projection geometry and subject location) were kept the same [18, 19, 21, 22, 25, 27].

For both virtually lowered tube current and sparse sampling, a stepwise approach was followed, generating low-dose simulations at 50% and 10% of the original tube current (D50, D10) and at 50% and 10% of the original projection data for sparse sampling (P50, P10) by reading only every second or tenth projection angle and deleting the remaining projections in the sinogram. The original MDCT imaging data with 100% tube current and 100% projections (i.e., original full-dose images) were defined as D100 P100.

The full-dose and dose-reduced simulated images were reconstructed with an in-house developed SIR algorithm [18,19,20]. We applied ordered-subset separable paraboloidal surrogation and a momentum-based accelerating approach for SIR [28]. A proper regularization was used for SIR to optimize the image quality for virtual low-dose scans. To enhance convergence and to further depress image noise while achieving adequate bone/soft tissue contrast, a regularization term based on a Huber penalty was applied, with the distinct strength of the regularization term being selected in consensus with three board-certified radiologists (aiming at visually high diagnostic image quality). The voxel intensities (mass attenuation coefficients, m2/kg) were calibrated to Hounsfield units (HU) by using air/water information from the MDCT calibration data.

Extraction of determinants of bone strength

To derive cortical and trabecular image-based determinants of bone strength from the full-dose and dose-reduced simulated images, an automated image analysis pipeline was applied (to left and right periosteal proximal femoral segmentations) [10]. Briefly, cortical bone was segmented using a non-local fuzzy c-means algorithm where the vBMD and the distances of each bone voxel to the closest soft tissue voxel were used as clustering features. Surface-based Ct.Th maps were then generated using the streamline integral thickness (SIT) technique, which takes into account partial volume effects. To perform subregional analyses at corresponding anatomical subregions across all scans in the study (full-dose and dose-reduced scans), volumes and surfaces of interest (femoral head, femoral neck, and trochanteric region) were predefined in a femoral template (Fig. 1) and transferred to all scans using affine and non-linear image registrations based on shape. Similarly, maps of vBMD and Ct.Th were spatially normalized to a femoral template using affine and non-linear registrations of the femoral shapes to perform localized volume-based and surface-based analyses using the SPM framework [10, 29]. We calculated integral vBMD (in mg/cm3) for the total hip; cortical vBMD (in mg/cm3) for the femoral neck and trochanteric region; trabecular vBMD (in mg/cm3) for the femoral head, femoral neck, and trochanteric region; and Ct.Th (in mm) for the femoral neck and trochanteric region. Cortical bone at the femoral head was not analyzed due to its thinness.

Fig. 1

Coronal cross-section of the proximal femoral template showing the femoral head (green), femoral neck (blue), and trochanteric (white) regions of interest used in the analyses of this study

All parameters were computed for both the left and right hips for the full-dose data (D100 P100) as well as for the simulated low-dose data of all levels (D50 P100, D10 P100 and D100 P50, D100 P10). The reproducibility values of compartmental vBMD and Ct.Th calculations for this automated pipeline in test–retest analyses have been reported to range under a root mean square coefficient of variation (CVRMS) of 1.7% (absolute precision error = 1.53 mg/cm3) and 2.7% (absolute precision error = 0.042 mm), respectively. Moreover, the reproducibility for local Ct.Th calculations has been reported to vary spatially across the proximal femur (CVRMS better than 7% with absolute precision errors smaller than 0.07 mm, excluding the femoral head) [10]. Analyses were performed independently for the left and right femur.

Statistical analysis

Statistical data analyses were performed with GraphPad Prism (version 10.1.1.; GraphPad Software, Boston, MA, USA) and MATLAB (version R2022b; The Mathworks Inc., Natick, MA, USA).

Descriptive statistics were calculated for all measures, including mean, SD, median, minimum, and maximum values, considering full-dose data (D100 P100) as well as the simulated low-dose data of all levels (D50 P100, D10 P100 and D100 P50, D100 P10). Shapiro–Wilk tests indicated a normal distribution for the quantitative data. Trabecular vBMD, cortical vBMD, and Ct.Th values for the femoral neck of both sides were used to create boxplots. To assess the accuracy of the low-dose bone parameters, we performed linear regression and Bland–Altman analyses with respect to the full-dose parameters. We also calculated relative differences for the low-dose data in relation to the full-dose MDCT data (in %) and used paired t-tests to compare relative differences between D50 P100 and D100 P50 and between D10 P100 and D100 P10 data. Differences were considered statistically significant at a p-value < 0.05.

We used the SPM framework to assess the spatial distribution of the accuracy of low-dose vBMD and Ct.Th values on a voxel-by-voxel and vertex-by-vertex basis, respectively, yielding maps of parameters derived from linear regression analyses.