Voxel size dependency, reproducibility and sensitivity of an in vivo bone loading estimation algorithm

Abstract

ResearchCite this article: Christen P et al. 2016 Voxel size dependency, reproducibility and sensitivity of an in vivo bone loading estimation algorithm. J. R. Soc. A bone loading estimation algorithm was previously developed that provides in vivo loading conditions required for in vivo bone remodelling simulations. The algorithm derives a bone's loading history from its microstructure as assessed by high-resolution (HR) computed tomography (CT). This reverse engineering approach showed accurate and realistic results based on micro-CT and HR-peripheral quantitative CT images. However, its voxel size dependency, reproducibility and sensitivity still need to be investigated, which is the purpose of this study. Voxel size dependency was tested on cadaveric distal radii with micro-CT images scanned at 25 mm and downscaled to 50, 61, 75, 82, 100, 125 and 150 mm. Reproducibility was calculated with repeated in vitro as well as in vivo HR-pQCT measurements at 82 mm. Sensitivity was defined using HR-pQCT images from women with fracture versus non-fracture, and low versus high bone volume fraction, expecting similar and different loading histories, respectively. Our results indicate that the algorithm is voxel size independent within an average (maximum) error of 8.2% (32.9%) at 61 mm, but that the dependency increases considerably at voxel sizes bigger than 82 mm. In vitro and in vivo reproducibility are up to 4.5% and 10.2%, respectively, which is comparable to other in vitro studies and slightly higher than in other in vivo studies. Subjects with different bone volume fraction were clearly distinguished but not subjects with and without fracture. This is in agreement with bone adapting to customary loading but not to fall loads. We conclude that the in vivo bone loading estimation algorithm provides reproducible, sensitive and fairly voxel size independent results at up to 82 mm, but that smaller voxel sizes would be advantageous.

Upscaling ofIn VivoHR-pQCT Images Enables Accurate Simulations of Human Microstructural Bone Adaptation

“…Furthermore, HR-pQCT images tend to have more noise 54[18] and other potential imaging artefacts, such as those due to movement [19]. 55The reduction in resolution is a known obstacles for the translation of computational techniques 56 from the lab into the clinical setting [20][21][22]. Thus, the use of clinical images in microstructural bone 57 adaptation simulations requires us to first understand the convergence of existing algorithms with 58 respect to image resolution [22,23] and second, evaluate whether supersampling of HR-pQCT data to 59 the resolution of desktop micro-CT images on which the algorithms have been validated produces 60 accurate results [18,24].…”

“…55The reduction in resolution is a known obstacles for the translation of computational techniques 56 from the lab into the clinical setting [20][21][22]. Thus, the use of clinical images in microstructural bone 57 adaptation simulations requires us to first understand the convergence of existing algorithms with 58 respect to image resolution [22,23] and second, evaluate whether supersampling of HR-pQCT data to 59 the resolution of desktop micro-CT images on which the algorithms have been validated produces 60 accurate results [18,24]. Supersampling of magnetic resonance imaging (MRI) data has been shown 61 to produce micro-FE results in good agreement with those from micro-CT images of a higher 62 4 resolution [18].…”

“…This algorithm tries to linearly combine three different load cases to achieve 149 the most homogeneous SED distribution possible across the given bone structure. The target mean 150 SED value was 0.02 MPa, as has been used previously [22]. Furthermore, soft pads were added to 151 the distal and proximal ends of the images with a pad-thickness of 246 µm and a Young's modulus of 152 15 MPa, which has previously been found to improve the load estimation [36].…”

“…The strain-adaptive in silico microstructural bone adaptation simulation by Adachi et al [28] was re-164 implemented in Python using NumPy [37] and pybind11 [38]. In short, this algorithm is iterative; for 165 The high-resolution images were segmented using a threshold of 450 mg HA/cm 3 [22] and the bone 172 volume over total volume (BV/TV) for the trabecular regions was computed for reference. Finally, 173 voxels identified as bone were set to 750 mg HA / cm 3 and background voxels were set to 0 mg HA / 174 cm 3 .…”

Towards in silico prognosis using big data

“…The analysis is performed with the dedicated micro-FE solver ParOSol [8] on the CPU. Boundary conditions are defined according to a bone loading estimation algorithm [9], providing physiological in vivo loading for this particular patient. They include compression (zz-direction) and shear strains (zx-and zy-direction).…”