Automated Bone Landmarks Prediction on the Femur using Anatomical Deformation Technique

Seung-Yeob Baek a Joon-Ho Wang b Insub Song c Kunwoo Lee a Seungbum Koo d,
b Department a School of Mechanical and Aerospace Engineering, Seoul National University, Seoul, South Korea of Orthopaedic Surgery, School of Medicine, Sungkyunkwan University, Samsung Medical Center, Seoul, South Korea c Department of Radiology, School of Medicine, Chung-Ang University, Seoul, South Korea d School of Mechanical Engineering, Chung-Ang University, Seoul, South Korea

Abstract Anatomical landmarks on bones play important roles in musculoskeletal simulations and surgical planning. This study develops an anatomically deformable model of the femur to predict bone landmarks automatically and quanties its prediction accuracy. Fortythree angiographic computed tomography (CT) images of femurs were collected and 14 bone landmarks were manually marked on these images by experts. Surface mesh models of the femur were extracted from the CT images and combined with the bone landmark information to create an anatomical deformable model. The anatomical deformation technique developed in this study predicted bone landmarks automatically as the surface of a deformable model was matched to the surface of a given femur model. The prediction accuracy was quantied using the leave-one-out cross-validation method. The average prediction error for the 14 landmarks ranged from 2.80 to 5.93 mm. While the prediction accuracies of anterior and posterior cruciate ligaments and lateral epicondyle sites were high with averages (standard deviation) of 3.00 (1.55), 2.80 (1.76) and 2.97 (1.87) mm, respectively, those of gluteus minimus, ligament of head of femur and piriformis sites were low with averages of 5.93 (3.77), 4.89 (3.49) and 4.87 (2.70) mm, respectively. Accuracy can be expected to increase with the use of more population data as is the nature of a population-based statistical deformable model.
Key words: bone landmarks, anatomical deformation technique, femur, statistical shape analysis, joint biomechanics

1. Introduction Anatomical landmarks on the bone surface play very important roles in determining the surgical reference points for total joint replacements and biomechanical applications. In navigation surgery techniques bone landmarks are marked using a probe with optical ball markers to obtain 3-dimensional (3D) positions. Based on the bone landmark positions surgical guides and cutting tools are aligned with the bone and joint [1]. Biomechanical human musculoskeletal simulation requires a skeletal model with muscles attached. The muscle attachment locations that can be obtained from bone landmarks determine the length of muscle moment arm and signicantly aect the accuracy of simulation results [2]. Previous studies reported that the visual marking of the bone landmarks on denuded cadaveric bones during a simulated and navigation based total knee surgery had rela Corresponding Author

tively large errors in both intra-observer and inter-observer repeatability [35]. The mean errors for 6 anatomical points around the knee ranged from 4.9 to 11.1 mm and statistically signicant dierences existed between observes in identifying most of the anatomical points [5]. On the other hand, the identication of bone landmarks on the femur and tibia in 3D anatomical images from computed tomography (CT) had relatively high intra-observer and interobserver reproducibility. When tested with 3 observers on 17 bone landmarks on the femur and tibia, the mean intraobserver error ranged from 0.4 to 1.4 mm and the mean inter-observer error ranged from 0.3 to 3.5 mm [6]. Automated bone landmark prediction has been introduced in previous studies. Beil et al. could detect bone landmarks based on the local curvature of isocontours in 3D images with information on the initial manual localization [7]. Automatic prediction of bone landmarks based on pure geometric analysis achieved high reproducibility but it could not be guaranteed that the anatomical landmarks were always on the geometrically distinctive points [7,8]. A

3D template matching technique to localize the bone landmarks on the knee was introduced [9] but the technique used the bone landmarks pre-determined from a single template model, which did not reect the population variation of the bone landmark positions. The statistical shape modeling technique to predict target model shape from the shapes of database models has achieved a great advancement for last decades [10,11]. Previously, we have developed a human body shape modeling method that creates a statistical shape model with variations from a large number of human body shape population data [12] utilizing the principal component analysis (PCA) [13,14]. Related to the bone shapes, similar insights are applied recently to the human femur in a purpose of obtaining the statistically-based nite element (FE) models of the bone [1517]. This type of statistical approach utilizes the population shape database and predicts a target shape within the variations that exist in the population. In this study, we developed a novel statistical bone landmarks prediction method by utilizing the statistical shape modeling and shape prediction methods. The base data set contained not only the coordinates of points that compose model shapes but also the coordinates of points for bone landmarks. In the pre-process of creating a consistent mesh for each of the base data, the bone landmarks were assigned the consistent index throughout all base data. Assuming that the 3D bone surface shape of a new subject could be relatively easily obtained from 3D imaging data such as CT [18,19] or from a reconstruction of two-dimensional images [20], our statistical shape model calculated from the consistent mesh models could predict the bone landmarks on this surface model statistically based on the population data. The suggested algorithm was tested with 43 human femur models with bone landmarks marked by clinical experts. The performance of the bone landmarks prediction was measured using the leave-one-out cross-validation method. The accuracy for the prediction of bone landmarks against the landmarks marked manually by clinical experts was quantied. 2. Methods 2.1. Data acquisition An internal review board (IRB) approval was obtained from the Chung-Ang University Hospital before the study. Forty-three angiographic CT images without patients identication were obtained from the database of the Department of Radiology. Fourteen anatomical landmarks on the femur were selected for this study, which are frequently used in navigation surgery of the knee [6] and in biomechanical musculoskeletal simulation [21,22]. The target bone landmarks are listed in Table 1. Two landmarks in the list were the bony eminences, 3 landmarks were the ligament attachment sites, and the rest were the insertion and origin sites of muscles. Some of the landmarks were frequently used 2

Fig. 1. Dissection of cadaveric knees to examine the attachment sites for ligaments and muscles on the femur

Fig. 2. Determination of ACL attachment site on the femur in Osirix Table 1 List of selected anatomical landmarks on the femur No. Name of landmark Type 1 Obturator internus Muscle insertion 2 Piriformis Muscle insertion 3 Gluteus minimus Muscle insertion 4 Psoas Muscle insertion 5 Adductor magnus distal Muscle insertion 6 Obturator externus superior Muscle insertion 7 Gastrocnemius medialis Muscle origin 8 Popliteus Muscle origin 9 Gastrocnemius lateralis Muscle origin 10 Ligament of head of femur Ligament attachment 11 Anterior cruciate ligament (ACL) Ligament attachment 12 Posterior cruciate ligament (PCL) Ligament attachment 13 Medial epicondyle Bony eminence 14 Lateral epicondyle Bony eminence

to determine a joints coordinate system [21,23]. The muscles in the list are the major muscles used during various activities; thus, their attachment locations are important for determining muscle activation level and internal muscle forces in musculoskeletal simulations [22]. An experienced musculoskeletal radiologist and an orthopedic surgeon identied 14 bone landmarks on the right femur in 43 angiographic CT images. The orthopedic surgeon had previously performed intensive dissections on the distal femur of a number of cadaveric knees to understand the attachment sites for ligaments and muscles as shown in Figure 1. The observers used the Osirix medical image viewer [24] to locate bone landmarks in the CT images as this shows very high accuracy and reproducibility [6]. The multi-planar reconstruction tool in Osirix visualizes CT data in any oblique plane aiding identication of the bone landmarks as in Figure 2. 2.2. Consistent mesh with bone landmarks 3D geometric models of the femur in the right leg were obtained from 43 CT scans using custom medical image segmentation software [19]. The accuracy of 3D models reconstructed using the software was previously reported

[19]. The surface model of the femur consisted of triangular meshes and vertices. In mesh-based statistical shape processing, mesh models should have identical topology or connectivity between triangular meshes so that the compatibility between the models is guaranteed. However, the surface mesh topologies and number of vertices of bone models generated from CT data using surface reconstruction techniques dier from each other. Therefore, before performing the statistical shape analysis, the meshes of a bone model should be re-parameterized so that it has consistent and identical topology. For this reason, we employed a template-based mesh re-parametrization technique [12]. In this re-parametrization technique, template mesh was deformed to minimize shape disparity so that the resultant deformation had the same shape as the target model while the original mesh topology of the template remained intact. This technique was applied to every model in the database, to give a set of models with the shape of the original individual and the unique topology inherited from the template. Shape disparity was dened by an error functional, which includes 4 energy terms: disparity error, smoothness error, landmark error, and distortion error. Most importantly, the role of the landmark error was to guarantee that the bone landmarks in the template model were exactly matched to the corresponding bone landmarks in the target. By doing so, we could guarantee one-to-one correspondences between the anatomical landmarks, and therefore, the vertex indices of the landmarks were identical despite variations in bone shape. 2.3. Anatomical deformation of the bone shape with landmarks With conrmed consistent topology between femur models, we dened a meaningful mathematical quantity representing the shape of a bone; the shape vector was dened as an array of coordinate values of a consistent mesh model and represented the shape of the corresponding bone model: x = [x1 , y1 , z1 , x2 , y2 , z2 , ..., xn , yn , zn ]T The shape vectors of each model were collected to build a data matrix X = [x1 , ..., xm ] where m was the number of models in the database. By using a matrix decomposition technique, one can can decompose the column space of the matrix X into 2 separate subspaces, one of which is possibly a meaningful subspace and the other is noise. In our study, we employed the PCA as the matrix decomposition technique to obtain m principal component vectors that span the column space of the matrix X in order of importance. In our study, around 16 principal component vectors in the rst part could explain 90% of shape variation in the population. In other words, the sum of the standard deviations in directions of rst 16 components was greater than 90% of the total sum of them. Thus, other remaining components share only less than 10% of the shape variations all together and thus are negligible. Hence, we used the rst 16 principal component vectors from the PCA to compose 3

Fig. 3. Eects of changing the weight values for the rst 4 signicant variation modes

the meaningful subspace of matrix X. Therefore, excluding the noise factors, we could approximate an arbitrary shape as the weighted sum of the 16 principal component vectors:
16

y=x +
i=1

wi pi

(1)

where y is the new data to be approximated and x is the mean shape vector. pi s are the principal component vectors and wi s are their corresponding weight values. In a specic case when the sources are the shape vectors, the dominant principal component vectors imply the major shape variation modes. Figure 3 shows the variation of the average femoral model that was manipulated by varying the weight values of the rst 4 signicant variation modes. The rst and the most signicant variation mode was related to the angles between the shaft and the both ends, and the second signicant variation mode was related to the torsion while the third and fourth aected the bending of the shaft and the scaling the distal and proximal ends, respectively. This anatomical deformation technique could vary the mean bone model with bone landmarks along natural bone shapes since the deformation modes were extracted from population data. Most importantly, since we conrmed that the anatomical landmarks have identical vertex indices, the resultant models derived from this technique will always have anatomical landmarks with the same indices. 2.4. Landmark localization The 16 signicant variation modes could be used as shape tuning parameters of the femoral bone and bone landmarks. One can generate an arbitrary natural bone shape by deforming the mean bone model according to Equation 1. We call this the deformable model. Therefore, newly observed bone models can be automatically estimated (called the target model) by conforming the deformable model whose mesh connectivity and vertex indices of landmark points are already known. To achieve this, the initial position and orientation of the deformable

bone model was roughly aligned to that of the target bone model. The position was aligned by coinciding the centroids of the 2 models, and the orientation was aligned by coinciding the principal axes of the bone models [25,26]. After the initial alignment, corresponding pairs of each vertex in the deformable bone model was selected from among the target models vertices. Various kinds of local differential geometry properties (e.g., curvatures) were taken into account for the selection of corresponding pairs. However, as bone models had very similar morphology between individuals, and we had already aligned the initial position and orientation of the 2 models, it was sucient to select the closest vertex in the target model from a vertex in the template model as a corresponding pair. It should be noted that the models include an average of more than several hundreds of thousands of vertices, which is therefore the number of candidates for a corresponding pair of one single vertex.). After the corresponding pairs were determined for every vertex in the deformable bone model, a rigid body rotation R and translation t that minimizes the sum of distances between the corresponding pairs was computed by minimizing the following equation: min
R,t

Fig. 4. Surface distance deviation between the tted model and the target model for a representative sample (left and middle), and box plots of average and maximum distance deviations for the 43 models in the leave-one-out cross-validation

2.5. Validation of automatic localization of the landmarks on subjects data The prediction accuracy for 14 landmarks in an arbitrary femur model was validated using the leave-one-out crossvalidation method [28]. Among the 43 femur models with 14 landmarks, 1 model was used as the validation model and the remaining 42 models were used as training models. A mean model and the rst 16 signicant variation modes of the training models were calculated using the anatomical deformation technique. Variation modes with dierent weight values were combined with the mean model to nd a model that best t the validation model. The resultant model incorporated the predictions for the 14 bone landmarks. The predicted landmarks positions were compared with the original landmarks positions in the validation model. This process was also repeated for other 42 models. 3. Result The accuracy of the bone surface shape tting was quantied for 43 femur models using the leave-one-out crossvalidation method. The distance deviation between the tted model and the validation model was quantied along the normal direction on the surface of the validation model as shown in Figure 4. The average surface distance deviation ranged from 0.56 2.18 mm with the average (standard deviation) being 1.08 ( 0.38) mm for the 43 models. The 25th, 50th and 75th percentiles of the average distance deviation were 0.75 mm, 1.03 mm and 1.26 mm, respectively as shown in the boxplot in Figure 4. The maximum distance deviation on the surface ranged from 2.74 8.00 mm with the average of 5.24 ( 1.33) mm for the models. The 14 anatomical landmarks on the femur were predicted in the process of tuning the weight variables of the 16 shape variation modes to t the surface shape of the target bone. In the leave-one-out cross-validation, the prediction accuracy for the 14 anatomical landmarks was quantied by calculating the distances between the predicted landmark locations and the locations manually determined by experts. The average prediction error for the 14 landmarks ranged from 2.80 5.93 mm, with highest accuracy for the posterior cruciate ligament (PCL) site and lowest accuracy for 4

Rai + t bi

(2)

where ai and bi are the vertices of deformable model and the target model separately. Then, the errors between the corresponding pair were minimized again by deforming the deformable model with anatomical deformation, which could essentially be done by tuning the wi s in Equation 1:
16

min
R,t

ai +
j =1

wj pi,j bi

(3)

where ai is the vertex of deformable model after the rigid body registration, (i.e., ai = Rai + t) and pi,j is the i-th element of j -th principal component vector. For these two minimization problems, there are many alternative ways for obtaining the solutions. In this research, we employed the quaternion method for 3D-3D rigid body registration [27] for solving Equation 2, and the least-squares method for Equation 3. Since application of the quaternion method for rigid body registration is direct and obvious, we do not discuss the details in this paper. Also we do not discuss the details of application of least squares, since the vector expression in Equation 3 can be easily converted to a matrix form, and hence the least squares method could be directly applied. By repeating this procedure (matching corresponding pairrigid body registrationanatomical deformation) until the average error between the corresponding pairs was lower than a threshold 1 and the improvement of the average error was lower than a threshold 2 , one could nally conform the deformable model to the target data. Then, as the vertex indices of the landmarks in the deformable model were already known, we could eventually estimate the locations of the landmarks in the target model.

Fig. 5. Average (standard deviation) distances between the predicted locations and the manually determined locations for the 14 anatomical landmarks

Fig. 6. Predicted anatomical landmarks on the distal and proximal femur for a representative sample

the insertion site of the gluteus minimus muscle. While the sites for anterior cruciate ligament (ACL), PCL and lateral epicondyle had higher prediction accuracies with averages of 3.00 (1.55), 2.80 (1.76) and 2.97 (1.87) mm, respectively, the sites for gluteus minimus, ligament of head of femur and piriformis had lower accuracies with averages of 5.93 (3.77), 4.89 (3.49) and 4.87 (2.70) mm, respectively. The results for the 14 landmarks are summarized in Figure 5. The illustration of the predicted anatomical landmarks in the distal and proximal femur is shown for a representative sample in Figure 6. Qualitatively speaking, the bony landmarks, i.e., medial and lateral epicondyles, were very close to the bony eminence. The sites for ACL and PCL were observed to be almost identical to the dissection results. On the proximal femur, the sites for muscles e.g., the psoas, were well predicted on the bony eminences. 4. Discussion Anatomical landmarks on the bone surface from population data were incorporated into a bone shape matching process through the creation of consistent meshes. The PCA based template matching and the landmark localization method could predict the bone landmarks with accuracies comparable to the direct identication of the landmarks in CT images [6] and was superior to identication on denuded knee bones [35]. This method used the bone shape of the population [9] and the bone landmarks positions of the population. The target data was given as a 3D bone surface model in this study. It was assumed that a 3D bone model can be relatively easily obtained from a 3D volume image such as CT data [18,19] or a bone model could be reconstructed using a computer vision technique from a few radiograph images with minimum radiation exposure [20]. Even the 3D bone surface model could easily obtainable accurate 5

prediction of the bone landmarks is a challenging work [3 5]. The deformable femur model can be tted to a new femur model with an average (standard deviation) surface deviation of 1.08 (0.38) mm in the leave-one-out crossvalidation. Because of the nature of a population based statistical deformable models, accuracy is expected to increase with the additions of further population models. In the course of tting the template model to the new femur model, the 14 bone landmarks on the template model matched the anatomical locations in the new femur model with average (standard deviation) distance errors ranging from 2.80 (1.76) to 5.93 (3.77) mm. It is very promising that 10 out of 14 landmarks had an average distance error of less than 4 mm. Previous studies on the identication accuracy of bony landmarks in the femur using palpation on the skin reported errors ranging from 14.1 19.2 mm [29,30]. An experiment on cadaveric knees to visually locate bone landmarks in the distal femur reported that the mean error between two observers was 7.1 mm and 11.1 mm for the medial and lateral femoral epicondyles, respectively [5]. In CT images, the inter-observer errors for 17 bone surface landmarks ranged 0.3 to 3.5 mm [6]. It is highly probable that the deformable femur model expanded by the 42 models could not provide enough degrees of freedom to t to an arbitrary femur model with surface deviation of less than 1 mm and average landmark prediction deviation of less than 2.8 mm. This database will be expanded with more femur models in the future to increase the accuracy. Ideally, bone landmarks data should be obtained from the 3D positions of the landmarks on dissected cadaveric bones as shown in Figure 1, but this eort may be an enormous endeavor. One of our observers (orthopedic surgeon) previously dissected 10 cadaveric femurs and identied the anatomical landmarks employed in this study (unpublished). Ligament insertion sites were identied using Hounseld values in CT in a previous study [31]. We took a similar approach using the Osirix [24]. The predicted bone landmarks can be used not only to align surgical tools during navigation surgery but also to build a subject-specic musculoskeletal model. Intraarticular joint force can be predicted from a musculoskeletal human simulation, which model a skeletons ligaments and muscles [22,32]. The accuracy of the simulation can vary depending on the muscle and ligament attachment locations that aect the moment arms in the simulation. The accurate prediction of the bone landmarks can decrease the discrepancies between patient measurements and musculoskeletal simulation results. In conclusion, a population data based anatomical deformation technique was developed. The technique accurately and robustly predicted 14 bone landmarks on the femur, outperforming any previously published manual methods. It is expected that the accuracy may increase by adding more collected data to the base model set. The predicted bone landmarks can be used to make subject-specic musculoskeletal models with subject-dependent anatomical co-

ordinate systems and muscle attachment locations. Furthermore, the vertices on the template model not included in the landmarks were also positioned at similar anatomical locations in the target model. Acknowledgements This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (2012-0000787 and 2012-0004102). References
[1] H. Bathis, L. Perlick, M. Tingart, Alignment in total knee arthroplasty. a comparison of computer-assisted surgery with the conventional technique, Journal of Bone and Joint Surgery, British Volume 86 (2004) 682687. [2] M. G. Hoy, E. Felix, F. E. Zajac, M. E. Gordon, A musculoskeletal model of the human lower extremity: The eect of muscle, tendon, and moment arm on the moment-angle relationship of musculotendon actuators at the hip, knee, and ankle, Journal of Biomechanics 23 (1990) 157169. [3] J. Y. Jenny, C. Boeri, Low reproducibility of the intraoperative measurement of the transepicondylar axis during total knee replacement, Acta Orthopaedica Scandinavica 75 (2004) 7477. [4] W. P. Yau, A. Leung, K. Y. Chiu, W. M. Tang, T. P. Ng, Intraobserver errors in obtaining visually selected anatomic landmarks during registration process in nonimagebased navigation-assisted total knee arthroplasty: a cadaveric experiment, Journal of Arthroplasty 2005 591601. [5] W. P. Yau, A. Leung, K. G. Liu, Interobserver and intra-observer errors in obtaining visually selected anatomical landmarks during registration process in non-image-based navigationassisted total knee arthroplasty, The Journal of Arthroplasty 22 (2007) 11501161. [6] J. Victor, D. Van Doninck, L. Labey, B. Innocenti, P. M. Parizel, J. Bellemans, How precise can bony landmarks be determined on a CT scan of the knee?, The Knee 16 (2009) 358365. [7] W. Beil, K. Rohr, H. S. Stiehl, Investigation of approaches for the localization of anatomical landmarks in 3D medical images, in: Proceedings of Computer Assisted Radiology and Surgery, Elsevier Science, 1997, pp. 265270. [8] K. Subburaj, B. Ravi, M. Agarwal, Automated identication of anatomical landmarks on 3d bone models reconstructed from CT scan images, Computerized Medical Imaging and Graphics 33 (2009) 359368. [9] E. Stindel, J. L. Briard, P. Merloz, S. Plaweski, F. Dubrana, C. Lefevre, J. Troccaz, Bone morphing: 3D morphological data for total knee arthroplasty, Computer Aided Surgery 7 (2002) 156168. [10] B. Allen, B. Curless, Z. Popovic, The space of human body shapes: reconstruction and parameterization from range scans, ACM Transactions on Graphics (Proc. SIGGRAPH 2003) (2003) 587594. [11] D. Anguelov, P. Srinivasan, D. Koller, S. Thrun, J. Rodgers, J. Davis, Scape: shape completion and animation of people, ACM Transactions on Graphics (Proc. SIGGRAPH 2005) 24 (2005) 408416. [12] S. Y. Baek, K. Lee, Parametric human body shape modeling framework for human-centered product design, Computer-Aided Design 44 (2012) 5667. [13] R. O. Duda, P. E. Hart, D. G. Stork, Pattern classication, 2nd Edition, Wiley-Interscience, 2000.

[14] J. Shlens, A tutorial on principal component analysis, Institute for Nonlinear Science, USCD (2005). [15] R. Bryan, P. S. Mohan, A. Hopkins, F. Galloway, M. Taylor, P. B. Nair, Statistical modelling of the whole human femur incorporating geometric and material properties, Medical Engineering and Physics 32 (2010) 5765. [16] L. Grassi, N. Hraiech, E. Schileo, M. Ansaloni, M. Rochette, M. Vicecontia, Evaluation of the generality and accuracy of a new mesh morphing procedure for the human femur, Medical Engineering and Physics 33 (2011) 112120. [17] B. Schuler, K. Fritscher, V. Kuhn, F. Eckstein, R. Schubert, Using a statistical appearance model to predict the fracture load of the proximal femur, Medical Imaging 2009: Visualization, Image-Guided Procedure, and Modeling 7261 (2009) 72610W 10. [18] M. Fiebich, C. M. Straus, V. Sehgal, B. C. Renger, K. Doi, K. R. Homann, Automatic bone segmentation technique for CT angiographic studies, Journal of Computer Assisted Tomography 23 (1999) 155161. [19] S. Koo, G. E. Gold, T. P. Andriacchi, Considerations in measuring cartilage thickness using MRI: Factors inuencing reproducibility and accuracy, Osteoarthritis and Cartilage 13 (2005) 782789. [20] N. Baka, B. L. Kaptein, M. de Bruijne, T. van Walsum, J. E. Giphart, W. J. Niessen, B. P. Lelieveldt, 2D-3D shape reconstruction of the distal femur from stereo X-ray imaging using statistical shape models, Medical image analysis 15 (2011) 840850. [21] T. P. Andriacchi, E. J. Alexander, Studies of human locomotion: past, present and future, Journal of Biomechanics 33 (2000) 12171224. [22] S. Kim, Three-dimensional dynamic model of the knee, Journal of Mechanical Science and Technology 12 (1998) 10411063. [23] J. H. Mun, A method for the reduction of skin marker artifacts during walking: Application to the knee, Journal of Mechanical Science and Technology 17 (2003) 825835. [24] A. Rosset, L. Spadola, O. Ratib, OsiriX: An open-source software for navigating in multidimensional DICOM images, Journal Digital Imaging 17 (2004) 205216. [25] E. Paquet, M. Rioux, A. Murching, T. Naveen, A. Tabatabai, Description of shape information for 2-D and 3-D objects, Signal Processing : Image Communication 16 (2000) 103122. [26] D. Vranic, D. Saupe, J. Richter, Tools for 3d-object retrieval: Karhunen-Loeve transform and spherical harmonics, in: Proceedings of 2001 Workshop on Multimedia Signal Processing, Cannes, France, 2001. [27] B. K. P. Horn, Closed form solution of absolute orientation using unit quaternions, Journal of the Optical Society of America A 4 (1987) 629642. [28] T. Hastie, R. Tibshirani, J. Friedman, The elements of statistical learning: data mining, inference, and prediction, 2nd Edition, Springer, 2009. [29] U. D. Croce, A. Cappozzo, D. C. Kerrigan, Pelvis and lower limb anatomical landmark calibration precision and its propagation to bone geometry and joint angles, Medical and Biological Engineering and Computing 37 (1999) 155161. [30] M. Rabuetti, G. Baroni, M. Ferrarin, Self-marking of anatomical landmarks for on-orbit experimental motion analysis compared to expert direct-marking, Human Movement Science 21 (2002) 439455. [31] M. L. Purnell, A. I. Larson, W. Clancy, Anterior cruciate ligament insertions on the tibia and femur and their relationships to critical bony landmarks using high-resolution volumerendering computed tomography, American Journal of Sports Medicine 36 (2008) 20832090. [32] M. Damsgaard, J. Rasmussen, S. T. Christensen, Analysis of musculoskeletal systems in the anybody modeling system, Simulation Modelling Practice and Theory 14 (2006) 11001111.