Introduction
Musculoskeletal modeling is required to estimate muscle and joint function, often with a long-term goal of understanding the form-function relationship of the musculoskeletal system (Erdemir et al., 2007). The clinical impact of musculoskeletal models is currently limited, due to the difficulty in generating patient-specific parameters, such as bone and joint geometry. The predicted muscle and joint contact forces are dependent on the accurate estimation of bone geometry and subsequent paths and lines-of-action of muscle-tendon units (Gerus et al., 2013). Estimating bone shape and pose from motion-capture landmarks is an essential part of patient-specific biomechanical simulation.
Simple length scaling of template models to landmarks cannot account for variations in bone shape (Blemker et al., 2007). Musculoskeletal software, such as OpenSim (Delp et al., 2007), typically scale a generic model (e.g. Delp et al., 1990) linearly and often isotropically according to experimental markers or anthropometric measurements. In the Anybody software (Aalborg, Denmark) the anatomy from a single cadaveric specimen (Horsman et al., 2007) is scaled nonlinearly using radial basis functions (Lund et al., 2015, Marra et al., 2015). Other non-linear scaling methods, such as host-mesh fitting (Fernandez et al., 2004), and the elastic registration method of Redert et al. (1999) applied to muscle morphing (Pellikaan et al., 2014) deform a template model to match experimental data. While providing additional degrees of freedom for shape morphing, these methods do not guarantee an anatomically realistic shape and often require extra smoothing constraints, which are chosen arbitrarily.
Statistical shape models are efficient and accurate in capturing realistic variations in anatomy (Allen et al., 2003, Bryan et al., 2009, Styner et al., 2003). In a typical shape model based on principal component analysis, realistic shapes can be generated as linear combinations of principal components. In musculoskeletal model generation, non-rigid registration using statistical shape models so far have been restricted to one or two bones. Kainmueller et al. (2009) presented a shape model of the hip joint for estimating full pelvis and femur geometry from limited field-of-view images. Yang et al. (2008) presented a shape model of the scapula and humerus, showing the correlation in shape between the two bones. The shape of the knee joint, in terms of the distal femur, patella, and proximal tibia, have been modelled by Fripp et al. (2007) and Rao et al. (2013), demonstrating the ability to model the shape variations of multiple articulated bodies. However, to the best of the authors’ knowledge, shape modeling has not been used to customize a lower limb musculoskeletal model from motion-capture landmarks.
We present an articulated shape model of the left lower limb, including the pelvis, femur, patella, tibia and fibula, with embedded muscle attachment regions. We use this articulated shape model to estimate bone geometry, pose, and muscle attachment locations from a sparse set of 7 common motion-capture landmarks. The output is a unified set of geometries that can be used for both rigid-body and continuum musculoskeletal analysis modelling. The method is validated using clinical computed tomography (CT) data to show improved accuracy compared to the standard isotropic linear scaling method.
Methods
The lower limb anatomy model is composed of a statistical shape model, model articulation, and embedded muscle attachment sites. A shape and pose optimization process fits the model to patient specific bony landmarks. We validate the accuracy of the optimization in a leave-one-out experiment and compare the results to conventional isotropic scaling.
Results
Fig. 4 illustrates the lower limb shape model. The first principal component captured mainly variations in bone size, while the second component captures variations in pelvis width coupled with an anterior-posterior shift of the lower limb. Subsequent components captured more subtle non-rigid variations within each bone. A typical principal component power spectrum was exhibited, with 75% of variation captured in the first five components.
For the shape model-based estimation, mean landmark.
Results
Fig. 4 illustrates the lower limb shape model. The first principal component captured mainly variations in bone size, while the second component captures variations in pelvis width coupled with an anterior-posterior shift of the lower limb. Subsequent components captured more subtle non-rigid variations within each bone. A typical principal component power spectrum was exhibited, with 75% of variation captured in the first five components.
For the shape model-based estimation, mean landmark.