Prediction of patellofemoral instability based on statistical shape analysis and machine learning based on 3D magnetic resonance imaging

Machine Learning

In this study, we created 3D MRI-based geometric models of a femur with PFI and a normal femur and compared the two models using SSA tools. Trochlear elevation in his PFI model compared to the normal model was observed at the central floor of the proximal trochlea. Furthermore, he applied PCA on the GPA-matched coordinate system to evaluate the shape variation of the PFI group and found that several principal components were related to the shape variation of the trochlea base and intercondylar width. Ta. Using multivariate analysis, we showed that these shape components were significantly correlated with his PFI/non-PFI distinction after adjusting for age and gender. Furthermore, using these shape components he developed an ML-based prediction model for PFI and obtained good prediction performance.

Morphological analysis of PFI has been previously reported. Previous studies have considered radiographic signs and several specific measurements such as the crossing sign, trochlear prominence or depth, and patellar alta index.5,27. However, these morphological parameters are mainly qualitative or semi-quantitative, resulting in high inter-reader variability and poor reliability. To overcome these shortcomings and provide a more realistic understanding of complex joint structures, computer-aided morphological analyzes using 3D shape models have been carried out.28,29Using 3D models of bone and articular cartilage, Yamada et al. demonstrated proximalization and lateralization of the trochlear cartilage in PFI and an associated wider convex trochlea.29. However, the approach used in this study was still based on a limited set of discrete geometric variables such as angle, height, and distance. SSA is considered suitable for visualizing and systematically understanding anatomical structures and their changes13. SSA allows you to estimate the variation in shape within a sample, obtain the average shape from groups, and perform tests for clustering and differences between groups. Landmark-based methods, also known as point distribution models, identify anatomical features and use statistics such as Procrustes transformation and PCA to align these point cloud sets to estimate changes in shape. often used in Early two-dimensional studies analyzed the curvature of the trochlear groove in the PFI. 30. In a more recent study, Van Haver et al. reported a CT-based 3D SSA to obtain average shape models of normal and trochlear dysplastic femurs and evaluate their differences.twenty two. This 3D SSA of PFI is mainly based on CT, although some studies have used 3D MRI. Fitzpatrick et al. We created an MRI-based 3D geometric model of the patellofemoral joint and showed that the main components are related to changes in patellar position and groove depth.26. In a more recent study, Yang et al. performed MRI-based 3D SSA and described a shallower trochlear groove and decreased anteroposterior and mediolateral dimensions of the femoral condyle in femurs with PFI. Did.twenty five.

The greatest risk factor for PFI is trochlear dysplasia, which has been shown to have a more anterior proximal trochlea in femurs with PFI than in normal femurs.27. Van Haver et al. We performed 3D SSA on trochlear dysplasia and demonstrated that the greatest differences between the average normal femoral and trochlear dysplastic femoral models were observed in the proximal part of the trochlea.twenty two. This study suggested that the proximal trochlea in trochlear dysplasia cases was elevated anteriorly compared to the normal trochlea. Furthermore, anteriorization of the trochlea was most pronounced in the central floor of the proximal trochlea.twenty two. In our study, similar to these findings, trochlear anteriorization in PFI was mainly observed in the middle of the trochlea base. We showed that anteriorization gradually decreases toward the notch. This is consistent with the results of previous studies. In addition to anteriorization and proximalization of the trochlea, lateralization of the trochlea has also been reported. Van Haver et al. We also demonstrated lateral migration of the trochlea in the trochlear dysplasia model, although this was less obvious than anteriorization and proximalization.twenty two. In the present study, lateralization of the trochlea was suggested by slight lateral movement of the medial and lateral sides of the trochlea. It has been suggested that the articular cartilage of the trochlear may be adapted to come into contact with the articular cartilage of the patella.29. Therefore, anteriorization, proximalization, and lateralization of the trochlea may be associated with patellar high-riding and lateralization. However, in this study, it was not possible to perform a correlation test of patellar position.

In this study, we performed PCA to assess the shape variability of the PFI group. Our findings showed that the first principal component involves the medial and lateral epicondyle and defines the horizontal size of the distal femur rather than the trochlea or condyle. The first factor is usually size variation, but this is not necessarily related to PFI, as described in previous studies.22,26. Our study showed that the second and third principal components were related to the shape changes of the groove angle and intercondylar width, and these were related to the actual shape changes of his PFI. Ta. Similar results were obtained in a previous study, demonstrating that the second and third components were related to the groove angle and intercondylar width, respectively. twenty two. It has been reported that there is a close relationship between the groove angle and the width of the intercondylar notch.27and both are associated with anteriorization of the central trochlea, which is the main factor in trochlear dysplasia, as discussed above.

In this study, we performed multivariate analysis to examine which shape features are correlated with PFI/non-PFI distinction. His second and third principal components above were significantly correlated with the distinction between PFI and non-PFI cases. Additionally, other shape components such as the 5th, 8th, and 10th components were also correlated with the PFI/non-PFI distinction. These shape features also explain variations in the depth of the anterior trochlear floor and the width of the intercondylar notch. Multivariate analysis of shape variance is useful for comparing groups with multiple confounders, just as in regular statistical analysis. In the current study, we selected age and gender as possible confounding parameters other than PFI/non-PFI. The principal components (2nd, 3rd, 5th, 8th, and 10th components) that significantly contribute to the PFI/non-PFI distinction are independent discriminatory factors even after adjusting for age and gender. was confirmed. Additionally, multivariate analysis was used to examine shape components associated with age and gender. No significant correlation between shape factor and age was observed in this study.Previous studies have shown that before osteoarthritis progresses, the shape of the knee does not change significantly over time. 31. Age-related changes may necessarily include changes related to osteoarthritis, but the current study did not include people over 50, where osteoarthritis is common. did. Therefore, there is no need to consider this factor. Regarding gender, this study showed correlations among several shape factors. After adjusting for PFI and age, the second, eighth, and ninth principal components were significantly correlated with gender. There are several reports on gender-related changes in the shape of the knee, with women having a deeper intercondylar fossa and a wider axial width compared to the epicondyle width compared to men. It has been shown that the inferior protrusion of the head is reduced.patellar groove31,32. In our study, the second and her eighth principal components had overlapping correlations not only with her PFI but also with her gender. However, there appeared to be more obvious changes in the width of the epicondyle than in other components. Furthermore, the ninth principal component was only correlated with gender, and its shape change was mainly related to diaphyseal width relative to epicondyle width, which is in line with previous findings regarding gender-related knee shape changes. It was consistent with research.31,32. In our results, low elevation of both condyles and deepening of the intercondylar fossa were unremarkable.

In this study, we evaluated ML-based PFI prediction using discriminative shape features obtained from a prior 3D SSA. Previous studies tested predictive models using discriminant analysis for PFI/non-PFI classification. Van Haver et al. reported automatic classification of trochlear dysplasia cases and normal cases using 3D SSA-based shape components with 85% sensitivity and 95% specificity.twenty two.Their classification model used only a linear discriminant modeltwenty two. To the best of our knowledge, no ML classifier other than LDA has been investigated so far. SVM exhibits high generalizability because it allows the choice of linear or nonlinear kernels. A linear kernel may be the best fit for your model. LDA also gave good results, but other nonlinear classifiers such as SVM with rbf kernel, k-NN, and RF were slightly worse than LDA and SVM with linear kernel. Possibly due to the small scale, this study was unable to build an efficient nonlinear model. Additionally, we used default hyperparameters in the nonlinear classifier without any parameter tuning during training. This may have led to more favorable results for the linear classifier than for the nonlinear classifier.

The current study has several limitations. The sample size of this study was small. Therefore, a full statistical evaluation could not be performed. For accurate comparisons, a control group of healthy volunteers matched for age, gender, and ethnic background would be desirable. Additionally, the patella and proximal tibia could not be evaluated, so only the distal femur was evaluated. Ideally, the overall shape of the knee should be considered. These issues should be resolved in future research.

In conclusion, this study reports a 3D MRI-based SSA of a normal femur with a PFI shape model. Pointwise distance maps showed that the elevation of the trochlea in his PFI model compared to the normal model was observed at the central floor of the proximal trochlea. In the PCA examining shape changes in the PFI group, several principal components showed shape changes in the trochlear floor and intercondylar width. Using multivariate analysis, we showed that these shape components were significantly correlated with her PFI/non-PFI distinction after adjusting for age and gender. Using these shape components he further developed an ML-based prediction model for PFI and obtained good prediction performance. 3D MRI-based SSA can realistically visualize the statistical results of surface models and may facilitate the understanding of complex geometric features. Further studies are needed to confirm the feasibility of 3D SSA and elucidate the disease mechanism of PFI.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *