Validation and comparison of three different methods for automated identification of distal femoral landmarks in 3D
-
Luisa Berger
,
Peter Brößner
Abstract
Objectives
Identification of bony landmarks in medical images is of high importance for 3D planning in orthopaedic surgery. Automated landmark identification has the potential to optimize clinical routines and allows for the scientific analysis of large databases. To the authors’ knowledge, no direct comparison of different methods for automated landmark detection on the same dataset has been published to date.
Methods
We compared 3 methods for automated femoral landmark identification: an artificial neural network, a statistical shape model and a geometric approach. All methods were compared against manual measurements of two raters on the task of identifying 6 femoral landmarks on CT data or derived surface models of 202 femora.
Results
The accuracy of the methods was in the range of the manual measurements and comparable to those reported in previous studies. The geometric approach showed a significantly higher average deviation compared to the manually selected reference landmarks, while there was no statistically significant difference for the neural network and the SSM.
Conclusions
All fully automated methods show potential for use, depending on the use case. Characteristics of the different methods, such as the input data required (raw CT/segmented bone surface models, amount of training data required) and/or the methods robustness, can be used for method selection in the individual application.
Introduction
The identification of bony landmarks in medical image data is of high relevance for various clinical applications as well as for basic research on large case databases. Preoperative planning [1], 2], computer-assisted surgical navigation [3], 4], evaluation of off-the-shelf as well as patient-specific implants [5], [6], [7], [8], [9] or research on morphotypes and their functional consequences [8], 10], 11] are just a few examples that rely on accurate landmark detection.
While in the past morphological analysis has mostly been performed in 2D [12], 13], 3D approaches have gained importance in recent years [14], [15], [16]. The higher accuracy of 3D approaches is still debated in the literature [17], [18], [19]. One advantage of 3D approaches is the ability to account for rotation between the scanner reference frame and the anatomical reference due to different positions in the scanner or deformities [16]. Reference planes and axis determined within 2D images are highly dependent on the exact position of the patient and accuracy is limited [18], 20].
In the field of landmark identification, manual selection by the surgeon is the gold standard. However, manual selection suffers from intra- and interobserver variability [21], 22] and is time-consuming and therefore expensive [22]. In a recent study by Okamoto et al., the identification of knee landmarks and subsequent measurements of 5 rotational parameters took on average of 30 min [23]. To assist the surgeon and minimise time and resources required, an automated identification of landmarks and subsequent morphological analysis and/or planning would be beneficial. Automated processing appears to be essential for the analysis of large case databases, [9], 24].
Several methods for automated identification of landmark detection on the knee joint have been presented in the literature [25], [26], [27], [28], [29], [30], [31], but, to the authors’ knowledge, there has been no comparison of different methods on the same database. In this study we compared 3 different approaches for automated identification of 6 distal femoral landmarks (neural network (NN), statistical shape model (SSM) and geometric approach (GA)) on the same database regarding the aspects: input data (CT/surface model), setup (training, adaption), robustness and accuracy with respect to manually selected reference landmarks.
Materials and methods
This study was performed on a database containing CT images of 101 Japanese THA patients who underwent unilateral primary total hip arthroplasty. Average patient age was 62.6 (±11.0, range 34–91) years. For each patient, left and right femora were evaluated, resulting in a total number of 202 femora. The femora were segmented according to the approach of Fischer et al. [32]. Next, the meshes underwent manual landmark identification by 2 raters (one single measurement per dataset each). Subsequently, patients were classified as osteophyte and non-osteophyte cases by visual inspection of the osteophyte volume in the distal femoral region. 12 patients (24 femora) were categorised as osteophyte cases, 89 patients (178 femora) were classified as non-osteophyte cases. 80 % of these non-osteophyte cases (n=142 femora) were used to train the approaches, if training was required. The remaining 20 % (n=36 femora) were used to test of all 3 approaches.
A potential application for automatic landmark identification is preoperative planning for joint surgery, where pathological deformities of the bones are to be expected. Therefore, the automatic landmark detection must be robust to shape differences. Therefore, we additionally analysed the bones previously classified as osteophyte cases with the three different approaches (Figure 1).
Workflow for automated landmark identification.
The evaluation of the 3 methods was performed against reference landmarks that were defined as average of the two manually identified landmarks. The mean absolute deviation from the reference landmarks were determined for the purpose of evaluating all methods. The selected landmarks included the medial and lateral epicondyle (MEC & LEC), the most distal point on the medial and lateral condyle (MDC & LDC) and the most posterior point on the medial and lateral condyle (MPC & LPC) (Figure 2). For manual landmark identification, the meshes were oriented with the shaft pointing in vertical direction. From this position, the most distal and most posterior points on each condyle were selected in frontal and sagittal planes, respectively. To identify the epicondyles, the meshes were additionally manually aligned in the frontal plane. The most prominent points in medial and lateral direction in the epicondylar region were selected. An additional check in the sagittal plane ensured that the chosen points were indeed part of the epicondylar area.
Landmarks for automated landmark identification.
Neural network
For the NN approach, the self-configuring convolutional neural network nnU-Net was used, as presented by Isensee et al. [33]. The network configuration comprised a 3D full resolution architecture, with 5 downsampling operations starting with 32 convolutional kernels up to a maximum of 320. The training configuration comprised a fixed data split, an initial learning rate of 0.01, a batch size of 2 and patch size of (128, 128, 128), and with mirroring disabled for data augmentation. The landmark identification was addressed as a semantic segmentation task with 13 classes (6 landmarks on the left side, 6 landmarks on the right side, background), as described by Hess et al. [34], with landmarks annotated as spheres of 5 pixels diameter. The neural network was trained on annotated DICOM data; hence no bone segmentation was necessary. As the approach uses CT data, the bones are not analysed individually, rather, left and right femora are analysed simultaneously.
Statistical shape model
In the context of the SSM analysis, the bone surface models of the femora were aligned in a bone-specific coordinate system, in accordance with the methodology presented by Asseln et al. [24]. The only additional information required for alignment of the mesh in the bone specific coordinate system is the femoral head centre (FHC), which was automatically determined according to the first steps of the A&A method [31]. The fully automatic positioning calculates a unified sagittal plane (USP) based on the work from Li et al. [35]. Based on this plane, prominent medial and lateral landmarks are calculated. Together with the FHC, these epicondyle landmarks are used to define the coordinate system, with the frontal plane being spanned between the FHC and the epicondyles. The definition of the coordinate system is as defined by Kai et al. [36]. The x-axis of the coordinate system is oriented in mediolateral direction between the condyles, the y-axis is oriented in anteroposterior direction, perpendicular to the x-axis and the frontal plane and the z-axis is oriented in proximo-distal direction, perpendicular to the x- and y-axes. Subsequent to this positioning, the training data were utilised to generate the annotated mean shape. An arbitrary shape from all training data sets is used as the initial reference shape, and point correspondences are determined by morphing this initial reference shape to the training data using the N-ICP-A algorithm from Amberg et al. [37]. Using the so found correspondences, the shapes were aligned with the initial reference shape. Finally, the determined shape vectors of the training dataset were averaged to compute the final mean shape. The process is iterated three times to reduce the influence of the initial reference shape. The mean shape thus created was then transformed to each individual femur of the testing dataset.
Geometric analysis
The GA is embedded in a software for automated morphological analysis of the distal femur [24]. In a first step, the bone surface model of the femur was oriented in the bone-specific coordinate system, which was also used for the SSM approach. The subsequent automated morphological analysis extracts a total of 44 morphological parameters such as dimensions, angles and radii that describe the shape of the distal femur. These parameters are subjected to a plausibility check to allow automatic detection of failed analyses. The identification of the landmarks analysed in this study based on geometric criteria was added to the workflow. The MEC and LEC were defined as points with maximum distance in the direction of the USP. MDC and LDC were defined as points with minimum z-value for positive and negative x-values (medial and lateral from centre of the coordinate system). Similarly, MPC and LPC were defined as points with minimum y-value. The software is fully automatic and requires no training data.
Results
With the NN, all test femora in the non-osteophyte group were analysed without error. The approach failed to determine 1 of the 6 landmarks for 2 of the osteophyte cases. The SSM was able to analyses all cases, both non-osteophyte and osteophyte. However, the prepositioning failed for 1 non-osteophyte case of the test dataset and for 2 osteophyte cases. These cases were excluded from the evaluation as the accuracy of the method may be different for these cases due to different baseline conditions. The GA failed for 3 non-osteophyte cases and for 7 osteophyte cases (Table 1). In 3 osteophyte cases, the failure occurred in the same bone with at least 2 methods (1 SSM+GA, 2 NN+GA). In 5 osteophyte cases and in 3 non-osteophyte cases, the analysis failed with only one method (OC: 1 SSM, 4 GA; NOC: 1 SSM, 2 GA). The NN only failed in cases that also failed in one of the other methods.
Robustness of the approaches for cases with and without osteophytes.
| Successful analysis of non-osteophyte cases (n=36) | Successful analysis of osteophyte cases (n=24) | |
|---|---|---|
| NN | 36 (100 %) | 22 (92 %) |
| SSM | 35 (97 %) | 22 (92 %) |
| GA | 34 (94 %) | 17 (71 %) |
The accuracy of the respective methods is reported as mean ± standard deviation for each landmark in Table 2. To ensure comparability, evaluation was performed for all cases that were successfully analysed by all approaches (33 non-osteophyte and 16 osteophyte cases). To test the statistical significance, the Welch test was performed (α=0.05) to compare each method with manual landmark selection. The NN showed a significant difference in accuracy for determining LPC in the non-osteophyte group, and the GA showed a significant difference in accuracy for MEC in the non-osteophyte and the osteophyte groups, for MPC in the osteophyte group, and for LPC in the non-osteophyte and osteophyte groups. In addition, the GA was the only method with a statistically significant difference in the mean deviation for all landmarks for the non-osteophyte and the osteophyte groups (Table 2).
Mean deviations between the reference landmarks and the different methods’ predictions for non-osteophyte cases (NOC) and osteophyte cases (OC); statistically significant differences are marked with a star (*).
| Landmark | Manual | NN | SSM | GA | ||||
|---|---|---|---|---|---|---|---|---|
| NOC | OC | NOC | OC | NOC | OC | NOC | OC | |
| MEC | 3.6 ± 1.2 mm | 3.2 ± 1.7 mm | 3.6 ± 2.1 mm | 3.2 ± 2.2 mm | 3.2 ± 2.0 mm | 3.8 ± 1.6 mm | 5.6 ± 2.8 mm* | 6.3 ± 2.5 mm* |
| LEC | 1.6 ± 0.9 mm | 2.3 ± 1.7 mm | 1.7 ± 1.0 mm | 2.3 ± 1.2 mm | 1.8 ± 1.0 mm | 2.4 ± 1.0 mm | 2.5 ± 1.8 mm | 2.4 ± 1.4 mm |
| MDC | 2.6 ± 2.1 mm | 2.3 ± 1.9 mm | 2.6 ± 1.6 mm | 3.1 ± 2.5 mm | 2.3 ± 1.2 mm | 3.3 ± 2.8 mm | 3.9 ± 2.5 mm | 3.1 ± 2.0 mm |
| LDC | 3.3 ± 2.3 mm | 3.5 ± 2.6 mm | 2.9 ± 1.7 mm | 3.4 ± 2.3 mm | 3.4 ± 1.7 mm | 3.8 ± 3.1 mm | 4.3 ± 2.8 mm | 4.8 ± 2.5 mm |
| MPC | 1.7 ± 1.3 mm | 2.4 ± 1.2 mm | 1.4 ± 0.8 mm | 2.2 ± 1.0 mm | 1.6 ± 0.9 mm | 2.0 ± 1.1 mm | 3.5 ± 1.9 mm* | 3.8 ± 1.9 mm |
| LPC | 2.4 ± 1.3 mm | 2.2 ± 1.1 mm | 1.5 ± 0.7 mm* | 2.0 ± 0.8 mm | 2.1 ± 1.0 mm | 2.8 ± 1.5 mm | 4.0 ± 2.5 mm* | 3.6 ± 1.3 mm* |
| Average | 2.6 ± 1.7 mm | 2.6 ± 1.8 mm | 2.3 ± 1.6 mm | 2.7 ± 1.9 mm | 2.4 ± 1.5 mm | 3.0 ± 2.1 mm | 4.0 ± 2.5 mm * | 4.0 ± 2.3 mm * |
The average results for all landmarks are also visually presented as violin plots in Figure 3.
Violin plots for average deviations between reference landmarks and landmarks detected by the different methods: the left side of each plot shows the distribution for the non-osteophyte cases, the right side for the osteophyte cases. The dotted lines represent the interquartile range and the dash line gives the median value.
Discussion
In our study, we compared 3 different methods for automatic landmark identification on CT images or surface models against manually identified reference landmarks. The interrater accuracy for the manual measurements was comparable to the accuracy reported in previous studies [21], 31].
All methods show potential for robust automated analysis in the field of basic research and clinical applications such as preoperative planning or intraoperative navigation [38], 39]. The choice of the best method depends on the application and the type of data available. The previously mentioned evaluation criteria, listed below, may be helpful in selecting the appropriate method for the specific application.
Input data
Regarding the required input data, SSM and GA require segmented surface meshes. This requires additional time and effort for segmentation. NN works directly on CT images and is therefore advantageous when considering time requirements. However, the NN is sensitive to varying image sections of the knee. Therefore, for image sections not represented in the training data, CT images must first be cropped to a common region of interest. The segmented surface meshes that are used for SSM and GA can be easily moved to a desired position, in order to achieve compatibility with the respective workflow.
Setup
NN and SSM require training data, whereas the GA is based on geometrical considerations and can therefore be used without training. When establishing new landmarks in the workflow, the GA can be adapted by adding new geometrical definitions. For SSM, new landmarks can be added by creating a new mean shape with training data or by annotating new landmarks directly on the already existing mean shape. For the latter approach, no further training is required. Changing landmarks for NN require adjustment of the training data and retraining. The number of training data sets required for the NN is generally larger than for SSM.
Robustness
All methods showed a high robustness of more than 90 % except for the GA for osteophyte cases. Robustness should not be the only criterion as methodically robustness does not directly correlate with robust landmark identification. While the NN and SSM are able to handle almost all cases, the usefulness of the result is not questioned during the analysis. Inappropriate input data can be processed successfully, while the results may be useless. Therefore, for these methods, either the input data or the output data must be carefully controlled.
In the case of GA, the method’s design provides a set of parameters that are needed for geometric definition of the bone-specific coordinate system and landmarks. These parameters can be used for an intrinsic plausibility check. Incorrect input data will result in a processing error and incorrect output data will be detected within the plausibility check. Both effects are reflected in a lower robustness of the method.
NN and SSM were able to identify landmarks for cases with osteophytes with a similar level of robustness as for non-osteophyte cases. The lower robustness of GA for osteophyte cases can be explained by the structure of the approach. As all processing steps are defined by geometric descriptions, shape deviations can lead to a change in the calculation basis. In the case of severely deformed bones, which can be expected in some pathologies, the implementation of the GA should be adapted, or another method should be used to ensure high robustness. Automated osteophyte removal may also be considered to increase the robustness [40].
As most of the cases that failed during the analysis failed for only one method, a combination of the methods might be useful for plausibility control and to increase robustness.
Accuracy
The accuracy of all three investigated methods for automated landmark detection was comparable to the manual method and also comparable to the accuracy reported in previous studies on automated landmark identification [18], 20], [29], [30], [31]. However, GA showed statistically significant differences compared to manual landmark selection not only for individual landmarks but also on average for both osteophyte and non-osteophyte cases.
When the osteophyte cases are compared with the non-osteophyte cases, there is a slight deterioration for all methods, but the differences are not statistically significant. Thus, the accuracy of all methods is in the same range for the presence of deformities and for non-deformed bones.
Limitations
Our study is a general comparison of methods, but its validity is limited by several restrictions. Each of the methods analysed depends on implementation and adaptation, which limits the generalisability of our findings.
Our data set was based on Japanese patients undergoing hip arthroplasty. Therefore, the transferability of our results to other ethnicities and different patient groups remains to be demonstrated. Our analysis was restricted to 101 patients. The number of training data, especially for the neural network, is therefore comparatively small compared to other studies. In addition, the manually selected reference landmarks were based on the manual selection of only two raters. A higher number of raters would be desirable in order to increase the accuracy of the manually selected reference landmarks.
We categorised the bones into osteophyte and non-osteophyte cases for deformity analysis. In the context of surgical treatment, many different types of deformities can occur. Further deformity analysis would be desirable to investigate the accuracy and robustness of the methods with respect to deformed bones, as expected in many surgical procedures.
Acknowledgments
The authors declare that they have no competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. No author received specific funding for this work.
-
Research ethics: Not applicable.
-
Informed consent: Not applicable.
-
Author contributions: All authors have accepted responsibility for the entire content of this manuscript and approved its submission.
-
Use of Large Language Models, AI and Machine Learning Tools: None declared.
-
Conflict of interest: The authors state no conflict of interest.
-
Research funding: None declared.
-
Data availability: Data is not publicly available.
References
1. Rajamani, KT, Styner, MA, Talib, H, Zheng, G, Nolte, LP, Gonzalez Ballester, MA. Statistical deformable bone models for robust 3D surface extrapolation from sparse data. Med Image Anal 2007;11:99–109. https://doi.org/10.1016/j.media.2006.05.001.Search in Google Scholar PubMed
2. Habor, J, Fischer, MCM, Tokunaga, K, Okamoto, M, Radermacher, K. The patient-specific combined target zone for morpho-functional planning of total hip arthroplasty. J Personalized Med 2021;11. https://doi.org/10.3390/jpm11080817.Search in Google Scholar PubMed PubMed Central
3. Hohmann, E, Bryant, A, Tetsworth, K. Accuracy of acetabular cup positioning using imageless navigation. J Orthop Surg Res 2011;6:40. https://doi.org/10.1186/1749-799x-6-40.Search in Google Scholar
4. Robinson, M, Eckhoff, DG, Reinig, KD, Bagur, MM, Bach, JM. Variability of landmark identification in total knee arthroplasty. Clin Orthop Relat Res 2006;442:57–62. https://doi.org/10.1097/01.blo.0000197081.72341.4b.Search in Google Scholar PubMed
5. Wong, KC, Kumta, SM, Chiu, KH, Cheung, KW, Leung, KS, Unwin, P, et al.. Computer assisted pelvic tumor resection and reconstruction with a custom-made prosthesis using an innovative adaptation and its validation. Comput Aided Surg 2007;12:225–32. https://doi.org/10.3109/10929080701536046.Search in Google Scholar PubMed
6. Grothues, S, Becker, AK, Hohlmann, B, Radermacher, K. Parameter-based patient-specific restoration of physiological knee morphology for optimized implant design and matching. Biomed Tech 2023;68:537–44. https://doi.org/10.1515/bmt-2023-0017.Search in Google Scholar PubMed
7. Hohlmann, B, Asseln, M, Xu, J, Radermacher, K. Investigation of morphotypes of the knee using cluster analysis. Knee 2022;35:157–63. https://doi.org/10.1016/j.knee.2022.03.006.Search in Google Scholar PubMed
8. Asseln, M, Grothues, S, Radermacher, K. Relationship between the form and function of implant design in total knee replacement. J Biomech 2021;119:110296. https://doi.org/10.1016/j.jbiomech.2021.110296.Search in Google Scholar PubMed
9. Grothues, S, Hohlmann, B, Zingde, SM, Radermacher, K. Potential for femoral size optimization for off-the-shelf implants: a CT derived implant database analysis. J Orthop Res 2023;41:1198–205. https://doi.org/10.1002/jor.25464.Search in Google Scholar PubMed
10. Grothues, S, Radermacher, K. Variation of the three-dimensional femoral J-curve in the native knee. J Personalized Med 2021;11. https://doi.org/10.3390/jpm11070592.Search in Google Scholar PubMed PubMed Central
11. Fantozzi, S, Catani, F, Ensini, A, Leardini, A, Giannini, S. Femoral rollback of cruciate-retaining and posterior-stabilized total knee replacements: in vivo fluoroscopic analysis during activities of daily living. J Orthop Res 2006;24:2222–9. https://doi.org/10.1002/jor.20306.Search in Google Scholar PubMed
12. Jeon, JH, Lee, SW, Song, KS, Bae, KC. Computed tomography in evaluation of femoral anteversion: transverse section versus axial oblique section. Keimyung Med J 2020;39:72–8. https://doi.org/10.46308/kmj.2020.00045.Search in Google Scholar
13. Scorcelletti, M, Reeves, ND, Rittweger, J, Ireland, A. Femoral anteversion: significance and measurement. J Anat 2020;237:811–26. https://doi.org/10.1111/joa.13249.Search in Google Scholar PubMed PubMed Central
14. Kim, JS, Park, TS, Park, SB, Kim, JS, Kim, IY, Kim, SI. Measurement of femoral neck anteversion in 3D. Part 2: 3D modelling method. Med Biol Eng Comput 2000;38:610–6. https://doi.org/10.1007/bf02344865.Search in Google Scholar
15. Grothues, S, Berger, L, Radermacher, K. Automated analysis of morpho-functional interbone parameters of the knee based on three dimensional (3D) surface data. In: Proceedings of the 20th annual meeting of the international society for computer assisted orthopaedic surgery; 2022, vol 5:81–8 pp.10.29007/8nb5Search in Google Scholar
16. Hess, S, Hirschmann, MT. 3D planning of total knee arthroplasty: why and how? In: Becker, R, Hirschmann, MT, Kort, NP, editors. Basics in primary knee arthroplasty. Cham: Springer International Publishing; 2022:233–48 pp.10.1007/978-3-030-58178-7_19Search in Google Scholar
17. Tiefenboeck, S, Sesselmann, S, Taylor, D, Forst, R, Seehaus, F. Preoperative planning of total knee arthroplasty: reliability of axial alignment using a three-dimensional planning approach. Acta Radiol 2022;63:1051–61. https://doi.org/10.1177/02841851211029076.Search in Google Scholar PubMed
18. Radtke, K, Becher, C, Noll, Y, Ostermeier, S. Effect of limb rotation on radiographic alignment in total knee arthroplasties. Arch Orthop Trauma Surg 2010;130:451–7. https://doi.org/10.1007/s00402-009-0999-1.Search in Google Scholar PubMed
19. Schwendicke, F, Chaurasia, A, Arsiwala, L, Lee, JH, Elhennawy, K, Jost-Brinkmann, PG, et al.. Deep learning for cephalometric landmark detection: systematic review and meta-analysis. Clin Oral Invest 2021;25:4299–309. https://doi.org/10.1007/s00784-021-03990-w.Search in Google Scholar PubMed PubMed Central
20. Kawakami, H, Sugano, N, Yonenobu, K, Yoshikawa, H, Ochi, T, Hattori, A, et al.. Effects of rotation on measurement of lower limb alignment for knee osteotomy. J Orthop Res 2004;22:1248–53. https://doi.org/10.1016/j.orthres.2004.03.016.Search in Google Scholar PubMed
21. Victor, J, Van Doninck, D, Labey, L, Innocenti, B, Parizel, PM, Bellemans, J. How precise can bony landmarks be determined on a CT scan of the knee? Knee 2009;16:358–65. https://doi.org/10.1016/j.knee.2009.01.001.Search in Google Scholar PubMed
22. Subburaj, K, Ravi, B, Agarwal, M. Automated identification of anatomical landmarks on 3D bone models reconstructed from CT scan images. Comput Med Imag Graph 2009;33:359–68. https://doi.org/10.1016/j.compmedimag.2009.03.001.Search in Google Scholar PubMed
23. Okamoto, S, Mizu-uchi, H, Okazaki, K, Hamai, S, Tashiro, Y, Nakahara, H, et al.. Two-dimensional planning can result in internal rotation of the femoral component in total knee arthroplasty. Knee Surg Sports Traumatol Arthrosc 2016;24:229–35. https://doi.org/10.1007/s00167-014-3370-1.Search in Google Scholar PubMed
24. Asseln, M, Hanisch, C, Schick, F, Radermacher, K. Gender differences in knee morphology and the prospects for implant design in total knee replacement. Knee 2018;25:545–58. https://doi.org/10.1016/j.knee.2018.04.005.Search in Google Scholar PubMed
25. Brehler, M, Thawait, G, Kaplan, J, Ramsay, J, Tanaka, MJ, Demehri, S, et al.. Atlas-based algorithm for automatic anatomical measurements in the knee. J Med Imaging 2019;6:026002. https://doi.org/10.1117/1.jmi.6.2.026002.Search in Google Scholar
26. Barbosa, RM, Serrador, L, da Silva, MV, Macedo, CS, Santos, CP. Knee landmarks detection via deep learning for automatic imaging evaluation of trochlear dysplasia and patellar height. Eur Radiol 2024;34:5736–47. https://doi.org/10.1007/s00330-024-10596-9.Search in Google Scholar PubMed PubMed Central
27. Chen, HC, Lin, CJ, Wu, CH, Wang, CK, Sun, YN. Automatic Insall-Salvati ratio measurement on lateral knee x-ray images using model-guided landmark localization. Phys Med Biol 2010;55:6785–800. https://doi.org/10.1088/0031-9155/55/22/012.Search in Google Scholar PubMed
28. Van Cauter, S, De Beule, M, Van Haver, A, Verdonk, P, Verhegghe, B. Automated extraction of the femoral anatomical axis for determining the intramedullary rod parameters in total knee arthroplasty. Int J Numer Meth Biomed Eng 2012;28:158–69. https://doi.org/10.1002/cnm.1478.Search in Google Scholar PubMed
29. Yang, D, Zhang, S, Yan, Z, Tan, C, Li, K, Metaxas, D. Automated anatomical landmark detection on distal femur surface using convolutional neural network. In: 2015 IEEE 12th international symposium on biomedical imaging (ISBI); 2015:17–21 pp.10.1109/ISBI.2015.7163806Search in Google Scholar
30. Baek, S, Wang, J-H, Song, I, Lee, K, Lee, J, Koo, S. Automated bone landmarks prediction on the femur using anatomical deformation technique. Comput Aided Des 2013;45:505–10. https://doi.org/10.1016/j.cad.2012.10.033.Search in Google Scholar
31. Fischer, MCM, Grothues, SAGA, Habor, J, La Fuente, M, Radermacher, K. A robust method for automatic identification of femoral landmarks, axes, planes and bone coordinate systems using surface models. Sci Rep 2020;10:20859. https://doi.org/10.1038/s41598-020-77479-z.Search in Google Scholar PubMed PubMed Central
32. Fischer, MCM, Krooß, F, Habor, J, Radermacher, K. A robust method for automatic identification of landmarks on surface models of the pelvis. Sci Rep 2019;9:13322. https://doi.org/10.1038/s41598-019-49573-4.Search in Google Scholar PubMed PubMed Central
33. Isensee, F, Jaeger, PF, Kohl, SAA, Petersen, J, Maier-Hein, KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021;18:203–11. https://doi.org/10.1038/s41592-020-01008-z.Search in Google Scholar PubMed
34. Hess, H, Herren, M, Gerber, N, Scheidegger, O, Schär, M, Daneshvar, K, et al.. Automatic quantification of fatty infiltration of the supraspinatus from MRI. In: CAOS 2022 the 20th annual meeting of the international society for computer assisted orthopaedic surgery; 2022:107–10 pp.10.29007/xq8mSearch in Google Scholar
35. Li, K, Tashman, S, Fu, F, Harner, C, Zhang, X. Automating analyses of the distal femur articular geometry based on three-dimensional surface data. Ann Biomed Eng 2010;38. https://doi.org/10.1007/s10439-010-0064-9.Search in Google Scholar PubMed
36. Kai, S, Sato, T, Koga, Y, Omori, G, Kobayashi, K, Sakamoto, M, et al.. Automatic construction of an anatomical coordinate system for three-dimensional bone models of the lower extremities-pelvis, femur, and tibia. J Biomech 2014;47:1229–33. https://doi.org/10.1016/j.jbiomech.2013.12.013.Search in Google Scholar PubMed
37. Amberg, B, Romdhani, S, Vetter, T. Optimal step nonrigid ICP algorithms for surface registration. Proc CVPR IEEE 2007:1491.10.1109/CVPR.2007.383165Search in Google Scholar
38. Alagha, MA, Logishetty, K, O’Hanlon, C, Liddle, AD, Cobb, J. Three-dimensional preoperative planning software for hip resurfacing arthroplasty. Bioengineering 2023;10. https://doi.org/10.3390/bioengineering10080939.Search in Google Scholar PubMed PubMed Central
39. Jones, CW, Jerabek, SA. Current role of computer navigation in total knee arthroplasty. J Arthroplast 2018;33:1989–93. https://doi.org/10.1016/j.arth.2018.01.027.Search in Google Scholar PubMed
40. Hänisch, C, Radermacher, K. Automatic osteophyte removal from distal femur surfaces obtained from CT scans. In: The 16th annual meeting of the international society for computer assisted orthopaedic surgery; 2016.Search in Google Scholar
© 2025 the author(s), published by De Gruyter, Berlin/Boston
This work is licensed under the Creative Commons Attribution 4.0 International License.
