Introduction

Major trauma due to vehicle accidents or conflict trauma often results in severe bone damage. Orthopaedic reconstruction of these cases is especially challenging with large defects [46] or amputation. If amputation is not avoidable, the general aim is to preserve as much tissue as possible. An example of tissue salvage is in knee disarticulations for which there is a higher functional outcome than for trans-femoral amputations [32, 38]. This is particularly important for children and adolescents for whom a fusion or loss of epiphyseal plate results in a disturbance of growth causing problems in their future development and the necessity for many future interventions [20], and for patients from develo** countries, where aftercare often challenges the patients and their families [3, 30].

Orthopaedic reconstruction requires knowledge of the premorbid bone shape to achieve accurate joint alignment. The shape can be used in pre-surgical planning to create cutting guides for implant placements, design patient-specific implants or be used in computer-aided systems during surgery (e.g., surgical robots, patient-specific instrumentations, augmented reality etc.). In clinical practice, the first choice to predict the intact three-dimensional (3D) geometry is to use the geometry of the contralateral side extracted from X-ray computed tomography (CT) or magnetic resonance images (MRI) [22, 40, 51]. This method is limited to cases that show no obvious asymmetries or bilateral defects and in situations, where there is ready access to costly 3D medical imaging. In develo** countries and areas of armed conflict such access to MRI or CT facilities is limited.

Reconstructions of 3D bone geometry have been studied previously. The methods include reconstructions using planar X-ray images [17, 18, 28, 48], and bi-planar X-ray systems [10, 19, 41, 52] that morph a bone template to match the contour lines of the X-ray image(s). In an intraoperative navigation application, Hurvitz and Joskowicz [24] used an active appearance model to reconstruct bone surfaces from calibrated X-ray images using a C-arm system. Reconstructing 3D volumes from 2D X-ray image(s) is a challenge, and Henzler et al. [21] demonstrated the use of deep learning-based convolutional neural networks for bone reconstruction. However, the reconstruction of a bone from a partially available bone has not previously been studied, neither SSM-based nor AI-based.

For orthopaedic reconstruction, the prediction of missing parts allows for the accurate restoration of joint kinematics in the musculoskeletal system [35]. A method widely used in the literature is the prediction of missing bone parts using statistical shape models (SSM) [1, 27, 31, 34, 2]. Therefore, predictions from image intensities might be more significant in paediatric bone shape predictions if an age-specific SSAM is used.

The projection of isolated bone geometries allowed easy segmentation of the contour lines. In clinical practice, this might require more, potentially manual, work and is, therefore, a potential source of error which could affect the reconstruction accuracy. Automated methods to segment bone geometries have been described in the literature [29] which would help minimise segmentation errors. Due to the way bones overlap, such as at the hip joint and due to the patella, local differences between clinical radiographs and the simulated radiographs in this study were not taken into account. As this would only affect local regions it is assumed that it will not have large effects on the reconstruction using the matching of intensity values and only affects the reconstruction using the contour measure through the segmentation, which is not evaluated in this study. In this study, the same projection method to simulate radiographs of the target shape and the reconstructions were used. This cannot be assumed for applications in clinical practice, where images from different sources might be used. This study tried to minimise this effect using an image intensity measure which has been shown to be robust for the comparison of images from different modalities [8]. Nevertheless, the evaluation of the robustness was not part of this study and needs to be investigated separately. The RMSE were used to evaluate the surface reconstruction error in this study. This is a commonly used parameter in such studies as it assesses average errors, whereas the Hausdorff distance gives information on the maximal error. The RMSE and Hausdorff distance were closely related in our study, likely due to the rather smooth surface coming from the segmented surfaces and automated surface smoothing algorithm. Reconstruction with bone defects from 0% and 50% of the whole bone were simulated in this study; therefore, average error was prioritised to take account of the different sizes of reconstruction. Finally, the bone geometries reconstructed in this study and the shape models were accurately aligned with regard to the anatomical directions, so that the projection directions did not need to be adjusted. In clinical practice, this might not be the case and an algorithm to maximise similarities between radiograph and projection would be necessary to optimise reconstruction results. This is a research question on its own and is not addressed in this study. The X-ray image used in this study is produced from CT scans using digitally reconstructed radiographs (DRR). We used in this study linear projection with a camera distance of 1500 mm (the standard clinical distance) without applying magnification of the images, so the projected X-ray size might appear different from the true size. However, as the same DRR-based X-ray projection setup is used for the sample preparation and shape reconstruction, we believe this is unlikely to amend our study conclusions. For clinical use, the DRR parameters need to be calibrated and so further study on this point is needed.

Conclusions

In summary, this study showed that reconstructions from 2D planar images using statistical shape and appearance models had an accuracy which would support their potential use in orthopaedic reconstruction after trauma, in examples such as a template for the reconstruction using AR systems, creating personalised instrumented guides for joint reconstruction and construction of patient-specific implants.

Materials and methods

Subject data

Full-body CT scans of 40 female cadaveric specimens obtained from the Digital Korean data set (http://dk.kisti.re.kr) [36] were used to create statistical shape and appearance models (SSAMs) following the method described in Nolte and Bull [34] using a leave-one-out strategy. In short, the models were created by aligning shapes using rigid body transformations calculated using an iterative closest point algorithm [6]; morphing a reference shape into all other shapes using free-form deformations [42]; creating a tetrahedral mesh of the reference shape and morphing it to all other shapes by solving a Laplace boundary condition problem as per Shontz and Vavasis [45],map** the Hounsfield units (HU) from CT scans to the volume meshes by map** the grey value of the closest voxel; and using a principal component analysis [13]. To allow the Hounsfield units to be read and assigned to the created shape model, we use the tetrahedral mesh to divide the shape model (i.e., surface only) into multiple “cells” (i.e., as a solid model). Since the extracted surface model shares the same coordinates with its DICOM CT images. For every single cell, with the known coordinates of the cell centroid point, we can acquire the greyscale or Hounsfield unit from the DICOM image, by repeating this for all the cells, a model can then be built with greyscale information (appearance model). The statistical shape and appearance models have previously been published [34], demonstrating compactness, generalization ability and specificity.

For each specimen, digitally reconstructed radiographs (DRRs) of the segmented right femur bones were created using a volume rendering method implemented in VTK (VTK 6.3.0, www.vtk.org) to simulate X-ray images. For the volume rendering, a linear transfer function describing the relationship between HU and opacity (op) values with coefficients op = 0.25 HU/1700, which was determined empirically, was used. DRRs were created for projections in the sagittal and coronal planes in anterior–posterior (AP) and mediolateral (ML) directions, respectively (Fig. 4). The projections were made with a consistent camera distance of 1500 mm to replicate a standard clinical distance.

Fig. 4
figure 4

Projection directions used for reconstructions from digitally reconstructed radiographs in a anterior–posterior and b anterior–posterior and mediolateral directions

Full size image

Shape reconstruction

The femur shapes of the 40 specimens were reconstructed from the SSAMs by calculating parameters for the modes of variation to minimise an error metric. For the reconstructions two error metrics were used: (1) quantification of the similarity of the contour of the DRR of the source and the projection of the shape model, and (2) quantification of the similarities in grey value intensities of the target DRRs and the DRRs of projections of the shape model instances. For the first metric, the contour line was extracted from the DRR of the source shape and compared to the aligned contour line of the projection of the shape model instantiation by calculating the average distance between contours. For the second metric, the grey values of the aligned images were compared by calculating a Pearson correlation coefficient of the grey values, which has been shown to be a robust measure in case the grey value ranges are not calibrated to the same scale [8]. For both measures, the images were aligned by extracting the contour lines of both images, pre-aligning to match the most proximal point of the contour lines and afterwards minimising the distance between them using an iterative closest point search algorithm [6] implemented in VTK. To determine the parameters for the modes of variation, the error measures were minimised using a Broyden–Fletcher–Goldfarb–Shanno (BFGS) optimisation algorithm for bound optimisation implemented in Python (L-BFGS-B, www.scipy.org).

Evaluation

To simulate bone defects, the distal part of the bones on the DRRs were cropped at levels of 0%, 10%, 20% and 50% of the bone length. Shapes were reconstructed using the two methods mentioned above using DRRs in AP only or AP and ML directions. Shapes were reconstructed using 14 modes of variation. The initial estimate for the optimisation problem was determined iteratively by estimating the parameters for one mode of variation using a guess of 0.0, and iteratively using the solution as an initial guess to calculate the solution for the optimisation problem with one mode of variation more.

The accuracy of the reconstructed shapes was evaluated by aligning the reconstructed shapes to the segmented intact bone shapes and calculating the root-mean-square error (RMSE) between them. In addition, anatomical measures used in the literature were evaluated [15, 34]. These are used in orthopaedic interventions and are the 3D angle between two tangent lines connecting the proximal and distal end of the femoral anatomic axis (bow angle), the angle between the mechanical and anatomical axis (FAA–FMA angle), the version angle using the trans-epicondylar axis (Version TEA) and the posterior condylar axis (Version PCA), the diaphyseal–condylar angle and the femoral neck angle. Furthermore, the radius of the femoral head was estimated by fitting a sphere to points on the bone surface (Fig. 5).

Fig. 5
figure 5

Definition of anatomical measures. Three-dimensional angles were defined as the femoral neck angle, the angle between the femoral anatomical and mechanical axis (FAA–FMA angle), the bow angle and the diaphyseal–condylar angle between the distal condylar axis and femoral anatomical axis. Version (PCA) and (TEA) were calculated from projections of the neck axis and posterior condyle and trans-epicondylar axis, respectively, onto the transverse plane

Full size image

Statistical analysis

The measures for evaluating reconstructions were compared for differences between error metrics, defect level and the number of projection planes. Results were analysed using non-parametric Kruskal–Wallis tests with paired Wilcox signed rank tests in the post-hoc analysis. All tests were performed with a significance level of α = 0.05 using R (v3.5.1, www.r-project.org).