This study investigated if a statistical shape and appearance model of the femur can be used to reconstruct patient-specific bone shapes from planar X-ray images with an accuracy that enables the model to be used for orthopaedic reconstruction. Bone shapes were reconstructed from simulated radiographs representing different levels of bone defect using two different methods. The reconstructions were evaluated by calculating RMS errors of the reconstructed surfaces and the surfaces segmented from CT images, and calculating the error of commonly used anatomical measures calculated from the surfaces.
To evaluate the methods for the use in clinical applications, the precision of anatomical measures was compared to deviations reported in the literature. Knee replacement surgeries for the treatment of osteoarthritis or sports injuries typically aim for preserving the joint alignment and therefore the kinematics. Whereas the accuracy of the preservation is influenced by the implant type (Bull et al., 2008), the positioning of the implant has the largest effect on the kinematics. A typical threshold for studies evaluating the lower limb alignment in varus-valgus is 3° (Cip et al., 2014); Yau et al. (2007) used the same threshold for the rotational alignment to calculate the success rate of joint replacement surgeries and found that over 50% of cases had angles above this threshold with values ranging from 17° internal to 15° external rotation.
In this study, the angle between anatomical and mechanical axes and the diaphyseal-condylar angle was evaluated to estimate the accuracy of the varus-valgus alignment. Whereas median errors between anatomical and mechanical axes are well below the threshold of 3°, median errors for diaphyseal-condylar angle are slightly above this threshold but still within the ranges reported in clinical studies. To estimate the accuracy of the rotational alignment, the version angles were evaluated. The median errors in these angles are slightly above the threshold used in clinical practice, but within the range observed in clinical practice with computer navigation (Yau et al., 2007).
For unicompartmental knee arthroplasties, Ng et al. (2017) reported errors in implant placement between 2° and 7° in femoral rotation for a lateral implant, Jaffry et al. (2014) reported errors between 3° and 7°. These measures were evaluated using the reconstruction of the transepicondylar axis using pre- and post-op CT scans. The errors in diaphyseal-condylar angles and version angles reported in this study were close to these values. The benefit of the method described in this study lies in the reduced exposure to ionising radiation compared to evaluating CT images and therefore can be an alternative to the estimation of the 3D shapes in surgical planning for unicondylar knee replacements.
Applications using augmented reality (AR) for orthopaedic reconstructions of the glenoid reported deviations for models of the glenoid of about 2.3 mm mean error (Berhouet et al., 2018, Berhouet et al., 2019). Typical registration errors for head mounted AR systems were reported between 0.8 mm (Chen et al., 2015) and 1.3 mm (Birkfellner et al., 2002). The average errors reported in this study had a comparable magnitude.
The shapes reconstructed using the image intensity metric were more accurate when only using projections in AP direction compared to using AP and ML projections. Previous studies comparing reconstructions with more than one perspective for planar radiographs using an intensity error metric reported only small differences in shape accuracy (Humbert et al., 2012). In our study, the CT images were only empirically calibrated, which might result in inaccuracies in the calibration and might result in variations in the image intensities. Creating SSAM using CT image calibration using a phantom might reduce the variation which could change this behaviour.
From the results of this study, reconstructions from intensity values seem less important than contour only for the prediction of the bone shape. Intensity values represent the internal structure and bone density and therefore to a large degree the bone strength. The images for the construction of the shape model were obtained from cadavers covering a large age range and bone density. As bone density reduces with age (Demontiero et al., 2012, Khosla, 2013), an SSAM for a narrow age range might produce more accurate reconstructions from matching intensity values. Also, bone strength and shape is is influenced by loading in growing bones (Ambrose et al., 2018). Therefore, predictions from image intensities might be more significant in paediatric bone shape predictions if an age-specific SSAM is used.
The study has a number of limitations which could have effects on the clinical application of the described methods. In this study, radiographs were simulated by projections of isolated bones without surrounding soft tissue, which is different from radiographs from patients, although this is a commonly-used and validated method (Galibarov et al., 2010, Zhang et al., 2017, Shi et al., 2022). The projection of isolated bone geometries allowed an easy segmentation of the contour lines. In clinical practice, this will require more, potentially manual, work and is therefore a source of errors which could affect the reconstruction accuracy. Automated methods to segment bone geometries have been described in the literature (Lindner et al., 2013) which would help minimise segmentation errors.
Due to the overlap of bones, for the femur especially at the hip joint and for 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 of the reconstructions was 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 by using an image intensity measure which has been shown to be robust for the comparison of images from different modalities (Birkfellner et al., 2009). Nevertheless, evaluation of the robustness was not part of this study and needs to be investigated separately.
Lastly, 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.