Abstract
INTRODUCTION: A combination of clinical hip MRI and a variety of 2D radiographs are used for the diagnosis and management of developmental dysplasia
of the hip (DDH), as well as for pre-surgical planning (1). However, these require manual identification of anatomical landmarks. Statistical shape models
(SSM) have been regarded as a robust tool for generating point-to-point correspondence and allowing for automatic identification of relevant landmarks (2).
In this study, we propose to automatically calculate a series of anatomical measures of the femur relevant to DDH using the correspondences generated by a
statistical shape model (3), and to validate the results by comparing them to manually measured results from 2D radiographs.
METHODS: A SSM of a femur bone (3) built from healthy lower limb CT scans (4) was used to reconstruct 23 clinical femur samples extracted from MRIs
of young adults with DDH with a field of view limited to the proximal 15 cm of the femur (REFgroup). For fitting the SSM on REFgroup, 5,000 points were
sampled from the REFgroup femurs to use as inputs for a Bayesian linear regression algorithm. The output of the fitting task were fully reconstructed 3D femurs
(SSMGroup) in point-to-point correspondence, allowing for automatic landmark identification for the SSMGroup femurs. Landmarks had to be manually selected
on the REFgroup femurs, as they were not in correspondence. Five anatomical measures (Table 1) were then automatically calculated from the landmarks using
a custom written Matlab script. The parameters were also manually measured using the 2D radiographs taken during the clinical care of the subjects. The
manual measures were done by an expert clinician, and was taken as the mean of 2 randomized and anonymized measures (2Dgroup). A dependent T-test was
used to compare the error of the manual 2D measurements and shape model generated predictions versus the REFgroup as a ground truth.
RESULTS SECTION: No significant differences were found between the error of the measures of the SSMGroup and 2Dgroup relative to the REFgroup (p>0.05)
(Figure 1, Table 1), however, SSMGroup tended to overestimate the values compared to 2Dgroup. The largest relative error occurred in a large overestimation and
underestimation of the height of the greater trochanter (GT) in the SSMGroup , and 2Dgroup respectively (71.76%, -68.28%). While the relative error in GT was
much larger than that of lesser trochanter in 2D, the absolute error between the measures was only 3 mm. Femoral neck-shaft angle had the lowest error of all
measures, and was predicted to less than 1° of error for both SSMGroup and 2Dgroup.
DISCUSSION: The preliminary results of our model show our error scores and both our reference and predicted values to be generally in good agreement
with the available literature for anatomical measures (1). Similar error scores seen between our model prediction and manually measured 2D values suggest
our tool is within levels of clinical acceptance for the margin of error. However, errors in the predictions of the height of the greater trochanter, femoral head
diameter and femoral offset highlight the effect of both model error due to the lack of DDH femurs in the training population, and the limitations of 2D imaging
methods, which have been shown to have errors due to positional and rotational effects of the leg during image acquisition (5). We aim to improve our results
by incorporating DDH femurs into the model building process in order to reduce the error between SSMGroup and REFgroup.
SIGNIFICANCE/CLINICAL RELEVANCE: This study highlights the limitations of 2D imaging methods in describing morphological parameters of bones
due to positioning errors. The use of 3D models opens possibilities for precise automatic calculation of several parameters used for diagnosis and surgical
planning in several orthopedic pathologies.
REFERENCES: 1) Wells et al. Clin Orthop Relat Res (475) 2017: 1045-54. 2) Ebert et al. Forensic Sci Int. (332) 2022. 3) Kwasny et al. World Congr.
Biomech, 2022:1938-1939. 4) Kistler et al. J Med Internet Res (15) 2013:e245. 5) Guidetti et al. Comp Meth in Biomech & Biomed Eng: Imag & Vis 2022.
ACKNOWLEDGEMENTS: Funding by German Sport University Cologne.
of the hip (DDH), as well as for pre-surgical planning (1). However, these require manual identification of anatomical landmarks. Statistical shape models
(SSM) have been regarded as a robust tool for generating point-to-point correspondence and allowing for automatic identification of relevant landmarks (2).
In this study, we propose to automatically calculate a series of anatomical measures of the femur relevant to DDH using the correspondences generated by a
statistical shape model (3), and to validate the results by comparing them to manually measured results from 2D radiographs.
METHODS: A SSM of a femur bone (3) built from healthy lower limb CT scans (4) was used to reconstruct 23 clinical femur samples extracted from MRIs
of young adults with DDH with a field of view limited to the proximal 15 cm of the femur (REFgroup). For fitting the SSM on REFgroup, 5,000 points were
sampled from the REFgroup femurs to use as inputs for a Bayesian linear regression algorithm. The output of the fitting task were fully reconstructed 3D femurs
(SSMGroup) in point-to-point correspondence, allowing for automatic landmark identification for the SSMGroup femurs. Landmarks had to be manually selected
on the REFgroup femurs, as they were not in correspondence. Five anatomical measures (Table 1) were then automatically calculated from the landmarks using
a custom written Matlab script. The parameters were also manually measured using the 2D radiographs taken during the clinical care of the subjects. The
manual measures were done by an expert clinician, and was taken as the mean of 2 randomized and anonymized measures (2Dgroup). A dependent T-test was
used to compare the error of the manual 2D measurements and shape model generated predictions versus the REFgroup as a ground truth.
RESULTS SECTION: No significant differences were found between the error of the measures of the SSMGroup and 2Dgroup relative to the REFgroup (p>0.05)
(Figure 1, Table 1), however, SSMGroup tended to overestimate the values compared to 2Dgroup. The largest relative error occurred in a large overestimation and
underestimation of the height of the greater trochanter (GT) in the SSMGroup , and 2Dgroup respectively (71.76%, -68.28%). While the relative error in GT was
much larger than that of lesser trochanter in 2D, the absolute error between the measures was only 3 mm. Femoral neck-shaft angle had the lowest error of all
measures, and was predicted to less than 1° of error for both SSMGroup and 2Dgroup.
DISCUSSION: The preliminary results of our model show our error scores and both our reference and predicted values to be generally in good agreement
with the available literature for anatomical measures (1). Similar error scores seen between our model prediction and manually measured 2D values suggest
our tool is within levels of clinical acceptance for the margin of error. However, errors in the predictions of the height of the greater trochanter, femoral head
diameter and femoral offset highlight the effect of both model error due to the lack of DDH femurs in the training population, and the limitations of 2D imaging
methods, which have been shown to have errors due to positional and rotational effects of the leg during image acquisition (5). We aim to improve our results
by incorporating DDH femurs into the model building process in order to reduce the error between SSMGroup and REFgroup.
SIGNIFICANCE/CLINICAL RELEVANCE: This study highlights the limitations of 2D imaging methods in describing morphological parameters of bones
due to positioning errors. The use of 3D models opens possibilities for precise automatic calculation of several parameters used for diagnosis and surgical
planning in several orthopedic pathologies.
REFERENCES: 1) Wells et al. Clin Orthop Relat Res (475) 2017: 1045-54. 2) Ebert et al. Forensic Sci Int. (332) 2022. 3) Kwasny et al. World Congr.
Biomech, 2022:1938-1939. 4) Kistler et al. J Med Internet Res (15) 2013:e245. 5) Guidetti et al. Comp Meth in Biomech & Biomed Eng: Imag & Vis 2022.
ACKNOWLEDGEMENTS: Funding by German Sport University Cologne.
Originalsprache | Englisch |
---|---|
Titel | ORS 2023 Annual Meeting |
Erscheinungsdatum | 2023 |
Aufsatznummer | 730 |
Publikationsstatus | Veröffentlicht - 2023 |
Veranstaltung | ORS 2023 Annual Meeting: Where Science and Community meet - Hilton Anatole, Dallas, USA/Vereinigte Staaten von Amerika Dauer: 10.02.2023 → 14.02.2023 https://www.ors.org/2023annualmeeting/ |