"Procrustes Analysis: Pediatric Supracondylar Humerus Fractures"

"Discover how Procrustes analysis is applied to examine the shape variations of pediatric supracondylar humerus fractures, offering insights crucial for treatment planning and outcomes assessment. This method sheds light on the nuanced differences in fracture patterns, aiding clinicians in making informed decisions for optimal patient care and recovery. By employing Procrustes analysis, healthcare professionals gain a deeper understanding of fracture morphology and its implications on treatment strategies, enhancing pediatric orthopedic care practices."

Abstract

The Procrustes method is utilized to analyze the shape of supracondylar fractures of the humerus in pediatric patients. X-ray images depicting the fractures from both anterior-posterior (AP) and lateral (L) perspectives are taken into account. Utilizing the Procrustes method for both views involves constructing and comparing images. The variability in shapes is assessed through a shape principal component analysis. Employing statistical shape analysis enables the prediction of typical humeral fracture shapes and their variability, aiding in preoperative planning by providing crucial insights into injury characteristics. Non-parametric tests, including permutational and bootstrap analyses, fail to indicate statistical differences between the Procrustes mean shapes in AP and L projections. However, it is demonstrated that the shapes of humeral fractures in AP and L views differ in their variability as quantified by shape principal components.

Introduction

Boston Brand Media brings you the latest news - The Procrustes method is used to examine the shape of pediatric patients' supracondylar humeral fractures. Both the anterior-posterior (AP) and lateral (L) views of XR fracture images are taken into account. When Procrustes' method is applied to both views, images are created and compared. A shape principal component analysis is used to quantify shape variability. Using statistical shape analysis, it is possible to predict the typical shape of a humeral fracture as well as its variability. This provides additional information on injury characteristics that are important for preoperative planning. Permutational and bootstrap non-parametric tests show no statistical difference between Procrustes mean shapes in lateral and anterior-posterior projections. However, it is demonstrated that the shape principal components differentiate the variability of the AP and L shapes of humeral fractures.

Materials and methods

We conducted a study involving 66 radiographic (XR) images taken in the anterior-posterior (AP) view and 24 images captured in the lateral (L) view of supracondylar humeral fractures in pediatric patients admitted to the Pediatric Trauma-Orthopedic Department. All methods were conducted following relevant guidelines and regulations. In the AP image group, there were 46 boys and 20 girls, with an average patient age of 8.2 years (ranging from 3 to 11). In the lateral group, there were 14 boys and 10 girls, with an average patient age of 8.6 years (ranging from 3 to 11).

Our analysis comprises three stages: digitization of fracture images, determination of mean shapes of fractures in both AP and L perspectives, and quantification of shape variability using Principal Component Analysis.

Fracture digitization

To analyze the shapes of fractures depicted in the XR images, we utilized Procrustes analysis, enabling the elimination of irrelevant shape features. Each fracture was discretized using a set of 10 scientific landmarks assigned by experts (cf. Fig. 1). This method, employing a relatively large number of landmarks, helps mitigate potential biases arising from subjective landmark selection by experts, ensuring a quasi-random and uniform selection.

In our study, we deliberately chose open-source and freely available computational tools for analysis. Initial image preparation was conducted using the Python 'skimage' package. Scientific landmarks were extracted from the XR images using the 'FigureCanvasBase.mpl-connect' method from the Matplotlib library, as illustrated in Fig. 1. These landmarks were then encoded into a configuration matrix X, which remains invariant under location, rotation, and isotropic scaling (Euclidean similarity transformations). Formally, a shape, denoted as X, belongs to a shape space M, a subset of the real Cartesian product, where N represents the number of landmarks in each 2-dimensional image. In assessing intra-observer agreement on landmark selection for a particular image i, it was assumed that the distance (Riemannian shape distance) between curves reconstructed from X is smaller than a fraction of the mean fracture length.

where, respectively, 𝑖=1,…,66 for the AP-view (𝑖=1,…,24 for the L-view) of fracture shapes and 𝑙𝑘,𝑘=𝑎,𝑏 denotes length of the fracture shape (approximation of the fracture) constructed upon experts’ landmarks 𝑋𝑖𝑎,𝑋𝑖𝑏. To calculate 𝜌 we use ’riemdist’ function from the R-package ’shapes’28. Resulting configuration matrices are collected and accessible cf. Additional information below.

Determination of typical shape and its variability

After digitizing the shape of a fracture using scientific landmarks, our objective was to construct and analyze the mean shape for both the anterior-posterior (AP) and lateral (L) projections. Procrustes matching, as summarized in references 19, 29, and 30, and applied in this study, involves registering all shapes (fractures) to their optimal positions through translation, rotation, and rescaling operations. This objective was achieved by minimizing the sum of squared Euclidean distances between the shapes. Consequently, we estimated the mean shape, denoted as E[X] or the Procrustes mean shape. Estimating the mean shape was complemented by determining the structure of shape ariability. The average shape of the fracture obtained via the Procrustes method is robust with respect to minor imperfections in both the patient's positioning during XR imaging and imprecise design of scientific landmarks. Moreover, it is not affected by image size. The essence of the Procrustes method lies in minimizing the total sum of squares.

where the Procrustes mean is calculated as a mean value of Procrustes coordinates 𝑋𝑖𝑃  as follows:

In calculations we apply ’procGPA’ function available in the “shapes” R-package28. It implements a three-stage algorithm consisting of the following steps: (i). Translations 𝑋𝑖𝑃=𝐶𝑋𝑖 where C describes centering of coordinates. (ii). Rotations of the ith configuration 𝑋¯𝑖=1𝑛−1∑𝑗≠𝑖𝑋𝑗𝑃

 leading to a new Procrustes registration 𝑋𝑖𝑃 involving only rotation of the old 𝑋𝑖𝑃. The n figures are rotated in turn. (iii) Scalling 𝛽𝑖=(∑𝑘=1𝑛||𝑋𝑘𝑃||2|𝑋𝑖𝑃||2)1/2𝜙𝑖,  where [𝜙𝑖]𝑖=1𝑛 is an eigenvector of the correlation matrix obtained from verctorizing 𝑋𝑖𝑃. Steps (ii) and (iii) are repeated until the Procrustes sum of squares of Eq. (2) is less than a tolerance parameters tol1, tol2 used further by ’procGPA’ function. Further both the tolerance for optimal rotation tol1 and the tolerance for rescalling step tol2 are always set tol1=tol2=10−5

Procurustes mean shapes are plotted using cubic spline interpolation. We utilize the ’interpolate’ module from the SciPy package. Statistical comparisons of mean shapes between the AP and L views is performed using non-parametric methods, employing permutations and bootstrapping techniques through the ’resampletest’ function available in the “shapes” R-package allowing to tests mean shape difference using complex arithmetic applicable in two dimensions. To quantify significance of the difference in shapes we used p-value of the Hotelling test included im the ’resampletest’ function.

In statistical shape theory a probabilistic ’events’ (the shapes), even in a planar case considered here, are elements of a high-dimensional space demanding multivariate methods. Standard multivariate parametric methods (such as MANOVA or complex-valued linear modelling) in statistical fracture shape analysis were not suitable as our data fail to satisfy multivariate normality. To verify normality we used the Mardia’s test available in the “QuantPsyc” R-package.

To provide a comprehensive description of the humeral fracture shapes, it is important to quantify their variability. A standard method for analyzing shape variability is Principal Component Analysis (PCA) to identify the most significant features influencing the covariance matrix of the shapes. A suitable measure for this purpose is the population covariance matrix of the coordinates calculated in the tangent space27. This tangent space, denoted as 𝑇𝜇(𝑀) , is constructed for the shape space M with the pole at the projection 𝜇, which represents the mean shape. In other words, the tangent space is the space tangent to the manifold M at the mean shape. The population covariance matrix

(where 𝑉∈𝑇𝜇(𝑀)) captures the variability of shapes around the mean shape and provides valuable information about the shape distribution. It describes how shapes deviate from the mean shape in different directions and magnitudes. The calculation of the population covariance matrix in the tangent space is based on the landmarks’ coordinates of the shapes in the dataset. The PCA analysis was conducted using the ’shapepca’ function from the aforementioned “shapes” R-package27,28.

Ethical approval

(PCN/022/KB/11/21) was waived by the local Ethics Committee of the Silesian Medical University of Katowice, Poland, because of the retrospective nature of the study and the fact that the procedures were part of routine care and in informed consent from all subjects and/or their legal guardian(s) was waived for the study.

Results

Mean shape of humeral fracture

Our initial aim was to construct a mean (typical) shape for humeral fractures in both the anterior-posterior (AP) and lateral (L) views. By identifying and selecting k = 10 scientific landmarks from each XR image, as depicted in Fig. 1 (see Additional Information below), we constructed a sample set of shapes, shown in the left panels of Fig. 2, for both the AP and L views. The observed differences in shapes, illustrated in Fig. 2, stem not only from the distinct characteristics of the corresponding fractures but also from technical factors and circumstances related to the radiographic imaging process. These differences can be attributed to variations in factors such as patient age or minor positional variations during the radiography procedure. However, it's crucial to note that all unwanted and artificial features in the analyzed XR images become irrelevant and can be disregarded with the application of the Procrustes method, as utilized in this study, to construct the mean fracture shapes. The Procrustes mean shapes, derived from the analysis, are presented in the right panels of Fig. 2. The mean shapes of fractures in the AP and L views are compared in Fig. 3. To improve visualization and facilitate further analysis, the Procrustes mean shapes are accompanied by cubic spline interpolation. This interpolation technique ensures mathematical continuity and enables the examination and discussion of shape features that require smoothness, such as the convexity of shapes, which can be assessed by analyzing the second derivative. From the observations in Fig. 3, it can be inferred that the convexity of the mean shapes does not depend on the projection (AP or L), with only minor local variations.

Fracture shapes (sample set) constructed using a 10-landmark approximation (left panels) and the corresponding Procrustes means (right panels) for the anterior-posterior (AP) and lateral (L) projections are depicted. The shapes in the left panels are translated to start at the same point. Since the Procrustes method for mean shape estimation is size-insensitive, the units in both the x and y directions are arbitrary.

Procrustes mean shapes of humerus fractures in the anterior-posterior (AP) and lateral (L) views are interpolated using cubic splines. As the Procrustes method for mean shape estimation is size-insensitive, the units in both the x and y directions are arbitrary.

To evaluate the statistical significance of differences between the mean shapes in the AP and L views, specific methods such as significance tests can be employed. These tests help determine whether the mean shapes in the AP and L views are statistically similar or different. In this case, it's unlikely that parametric methods would yield reliable results as the sets of landmarks do not satisfy normality assumptions. This is supported by examining the p-value for the multivariate Mardia’s test, which we applied to verify the normality of data. The obtained p-values are as follows: pAP=p_{AP} =pAP​= for the AP view and pL=p_{L} =pL​= for the L view, indicating significant deviations from normality for both skewness and kurtosis. Consequently, the normality hypothesis can be rejected, and non-parametric counterparts of standard statistical methods are deemed more appropriate, which we applied in our analysis.

Boston Brand Media also found that the non-parametric methods, based on permutations or bootstrap resampling, were utilized to compare the mean shapes in the AP and L projections. The results obtained from these methods suggest that there is no statistical difference between the mean shapes. In other words, based solely on p-values (ppp of the Hotelling test), there is no evidence to reject the null hypothesis of similarity between the mean shapes.

Shape variability: principal component analysis

Analyzing the population covariance matrix provides insights into the patterns and structure of shape variation, crucial for understanding the range of possible shapes within the population and for further statistical analysis and modeling. The results of Principal Component Analysis (PCA) for the AP and L views are depicted in Figs. 4 and 5, respectively, illustrating the principal components of shape variability.

For the AP view, the first three principal components explain 59.1%, 11%, and 9.8% of the variability in the mean shape, respectively. In comparison, for the L view, the corresponding weights are 59.1%, 13%, and 8.8%. Notably, when comparing principal components between the AP and L views, it's observed that the lower components (e.g., the second and third) exert a stronger impact on the overall variability of the mean shape in the L view.

Furthermore, there is a qualitative difference in the third principal component between the AP and L views, as indicated by the last row of panels in Figs. 4 and 5.

For the Procrustes mean shape of humerus fractures in the anterior-posterior view (AP), the three principal components of shape variability are illustrated. Each row of panels corresponds to one of the three principal components (PC1PC1PC1, PC2PC2PC2, and PC3PC3PC3). Within each row, the Procrustes mean shape is displayed in the central panel. In the left column of panels, the Procrustes mean shape is presented with tripled standard deviation subtracted, while in the right column, the Procrustes mean shape is presented with tripled standard deviation added. As the Procrustes method for mean shape estimation is size-insensitive, the units in both the x and y directions are arbitrary.

For the Procrustes mean shape of humerus fractures in the lateral view (L), the three principal components of shape variability are displayed. Each row of panels corresponds to one of the three principal components (PC1PC1PC1, PC2PC2PC2, and PC3PC3PC3). Within each row, the Procrustes mean shape is depicted in the central panel. In the left column of panels, the Procrustes mean shape is presented with tripled standard deviation subtracted, while in the right column, the Procrustes mean shape is presented with tripled standard deviation added. It's important to note that the Procrustes method for mean shape estimation is size-insensitive, hence the units in both the x and y directions are arbitrary.

These findings suggest that although there is no statistical difference between the mean shapes in the AP and L projections, differences exist in the principal components that quantify their variability. This underscores the importance of not only considering the mean shape but also the higher-order modes of variation when comparing the two projections.

Discussion

Your study on supracondylar humeral fractures in pediatric patients addresses a significant challenge faced by surgeons in preoperative assessment, particularly relying on X-ray images captured in both AP and lateral projections. These images play a vital role in determining treatment strategies, and your analysis shows that insights gained from both perspectives are crucial for optimal outcomes. Utilizing Procrustes analysis, you've effectively extracted essential shape features, enabling the construction of typical fracture shapes despite limitations such as small sample size and unequal image numbers between projections.

Your findings indicate that while mean shapes in AP and L projections are similar, differences exist in their variability, emphasizing the importance of considering higher-order modes of variation. This comprehensive approach enhances understanding of fracture characteristics and aids in treatment planning. The nearly constant curvature of typical fracture shapes in both projections underscores the potential value of such information for preoperative planning.

Moreover, your study highlights the Procrustes method's role as a foundational tool in statistical shape analysis, paving the way for more advanced techniques like mixed models to account for hierarchical data structures and incorporate fixed and random effects. Future investigations could explore how typical fracture properties relate to injury classification methods or patient demographics, contributing to both clinical practice and biomechanical research. Overall, your work provides valuable insights into shape variations and their implications, advancing the understanding of pediatric humeral fractures and informing treatment strategies.

For questions or comments write to writers@bostonbrandmedia.com

Source: nature‍