This article has Open Peer Review reports available.
A statistical shape modelling framework to extract 3D shape biomarkers from medical imaging data: assessing arch morphology of repaired coarctation of the aorta
© The Author(s). 2016
Received: 12 November 2015
Accepted: 19 May 2016
Published: 31 May 2016
Medical image analysis in clinical practice is commonly carried out on 2D image data, without fully exploiting the detailed 3D anatomical information that is provided by modern non-invasive medical imaging techniques. In this paper, a statistical shape analysis method is presented, which enables the extraction of 3D anatomical shape features from cardiovascular magnetic resonance (CMR) image data, with no need for manual landmarking. The method was applied to repaired aortic coarctation arches that present complex shapes, with the aim of capturing shape features as biomarkers of potential functional relevance. The method is presented from the user-perspective and is evaluated by comparing results with traditional morphometric measurements.
Steps required to set up the statistical shape modelling analyses, from pre-processing of the CMR images to parameter setting and strategies to account for size differences and outliers, are described in detail. The anatomical mean shape of 20 aortic arches post-aortic coarctation repair (CoA) was computed based on surface models reconstructed from CMR data. By analysing transformations that deform the mean shape towards each of the individual patient’s anatomy, shape patterns related to differences in body surface area (BSA) and ejection fraction (EF) were extracted. The resulting shape vectors, describing shape features in 3D, were compared with traditionally measured 2D and 3D morphometric parameters.
The computed 3D mean shape was close to population mean values of geometric shape descriptors and visually integrated characteristic shape features associated with our population of CoA shapes. After removing size effects due to differences in body surface area (BSA) between patients, distinct 3D shape features of the aortic arch correlated significantly with EF (r = 0.521, p = .022) and were well in agreement with trends as shown by traditional shape descriptors.
The suggested method has the potential to discover previously unknown 3D shape biomarkers from medical imaging data. Thus, it could contribute to improving diagnosis and risk stratification in complex cardiac disease.
Diagnosis and risk stratification of cardiac disease using medical imaging techniques are primarily based on the analysis of anatomy and structure of the heart and vessels. This is particularly true for complex conditions such as congenital heart disease (CHD), where the morphology defines the cardiac defect in the first instance and is subsequently altered by surgical and catheter intervention to improve functionality. In clinical practice, however, anatomical analysis of shape and structure is often carried out via simple morphometry, using parameters measured in 2D (e.g. vessel diameter, area, angulation). This does not fully exploit the abundance of information that current imaging techniques such as cardiovascular magnetic resonance (CMR) or computed tomography (CT) offer [1, 2]. Furthermore, using simple shape descriptors, the relationship between complex global and regional 3D shape features, such as the combination of stenoses, dilations or tortuosity and cardiac function has not been fully explored.
Conversely, statistical shape models (SSM) allow visualisation and analysis of global and regional shape patterns simultaneously and in 3D  as they are constituted by a computational atlas or template, which integrates all anatomical shape information intuitively as a visual and numerical mean shape and its variations in 3D. The template is essentially an anatomical model of the average geometry of a shape population. Based on this template, descriptive or predictive statistical shape models can be built [1, 4], to explore how changes in shape are associated with functional changes.
More recently, a novel non-parametric SSM framework that does not rely on any prior landmarking [9, 10], has been introduced to the cardiac field by Mansi et al. [11–13]. The method is based on a complex mathematical framework, which analyses how a representative template shape deforms into each of the shapes present in the population. In a simplified way, for example, an “ideal” template aorta can be deformed into any possible patient aorta shape by applying the correct transformations. Instead of the shapes themselves, these transformations are analysed  and subsequent shape analysis is carried out robustly within this transformation framework. A key advantage, in addition to neither requiring landmarking nor point-to-point correspondence between input shapes, is that the method is able to handle large variability between shapes, making it an even more attractive tool for investigating 3D cardiovascular anatomical structures in CHD.
The aim of this paper is to present this shape analysis method to the larger clinical and engineering community by describing a step-by-step approach to set up such a SSM and by demonstrating its validity using conventional morphometric parameters. As an example, the study focuses on aortic arch shapes of patients post coarctation repair [15, 16], as they typically present highly variable, complex shapes, which have been extensively described in terms of traditional morphometric analyses [17–19]. To demonstrate the capabilities of the proposed method, we have derived global and regional shape features potentially associated with ejection fraction (EF) as novel 3D shape biomarkers. We hypothesised that low EF, which characterises poor ventricular function, could be associated with distinct shape patterns of the aortic arch that affect cardiac afterload.
Statistical shape modelling framework (SSM)
Extracted shape patterns described by PLS shape modes are visualised by deforming the computed template shape with the transformations along the direction of the mode (Fig. 6a). To determine whether the obtained shape patterns are correlated with a response parameter, shape modes need to be broken down to numbers that allow statistical analysis. This is achieved by mathematically projecting each subject-specific patient transformation onto the found shape mode , which yields the so called shape vector XS (Fig. 6b). Shape vectors are essentially numerical representations of a specific shape mode. Each shape vector entry describes in one subject-specific number how much the template has to be deformed along the derived shape mode in order to match the specific subject shape as well as possible. The shape vector thus represents 3D global and regional shape features associated with a certain subject and response parameter. Further standard correlation analysis between the shape vector and the response parameter reveals how well subject shape features are represented by the derived shape mode (Fig. 6c). A perfect correlation of shape vector and response would imply that the derived shape mode showed exactly those shape patterns associated with low or high response values (such as high or low EF) when moving along the shape mode from low to high shape vector values.
For mathematical details about the shape modelling process as outlined above, we refer to Appendix 1 and the referenced literature. The entire mathematical framework has been published under the name “exoshape” and is publicly available as a Matlab (The MathWorks, Natick, MA) code [12, 22], (https://team.inria.fr/asclepios/software/exoshape/). A similar, open-source code has been recently published in C as “Deformetrica” by Durrleman et al.  (http://www.deformetrica.org/).
After the template is computed, the following post-processing analyses are carried out: i) removing confounders such as size differences between subjects prior to extracting shape features related to the functional parameter EF as they can hide potentially important shape features; ii) accounting for outliers and influential subjects that are common in clinical data of pathological shapes; iii) validating the template as representing the mean shape of the population and as being not substantially affected if any of the shapes that were used to compute it is removed or if a new patient is added iv) analysing associations between extracted shape features (represented by shape vectors as well as by traditional 2D and 3D measured geometric parameters) and demographic (BSA) and functionally relevant parameters (EF) via standard bivariate correlation analysis (Fig. 6c).
Patient population, image data and 2D morphometry
CMR imaging data from 20 CoA patients post-repair (16.5 ± 3.1 years, range 11.1 to 20.1 years; CoA repair performed at 4 days to 5 years of age) were included in the study. Conventional morphological descriptors for this population were previously reported by our group .
Three-dimensional volumes of the left ventricle (LV) and the aorta during mid-diastolic rest period were obtained from CMR using a 1.5 T Avanto MR scanner (Siemens Medical Solutions, Erlangen, Germany) with a 3D balanced, steady-state free precession (bSSFP), whole-heart, free breathing isotropic data acquisition method (iso-volumetric voxel size 1.5 × 1.5 × 1.5 mm) . Ejection fraction (EF) was measured from the CMR data .
Images were segmented using thresholding and region-growing techniques combined with manual editing in commercial software (Mimics, Leuven, Belgium) . A previous study comparing physical objects and their respective 3D segmented and reconstructed computer models found an average operator induced error in the order of 0.75 mm, which equals about half the voxel size in our study . In order to reduce irrelevant shape variability, aortas were cut such that only the root, the arch and the descending aorta up to the diaphragm were kept. As the focus of this analysis lies on the arch shape, coronary arteries and head and neck vessels were cut as close as possible to the arch. This is a common pre-processing step in shape analysis of aortic arches [26–28]. The final segmented surface models of the aortas were stored as computational surface meshes.
Morphometric parameters measured on the 3D surface models of the arches
2D measured parameters (manually)
3D measured parameters (VMTK)
• Arch height A [mm]
• Arch width T [mm]
• Ratio A/T
• Diameters at ascending, transverse, isthmus and descending level of the aorta: Dasc, Dtrans, Disth, Ddesc [mm]
• Ratios Dasc/Ddesc, Dtrans/Ddesc and Disth/Ddesc
• Volume V [mm3]
• Surface Area Asurf [mm2]
• Ratio Asurf/V
• Centreline length LCL [mm]
• Centreline Tortuosity ToCL
• Median curvature along centreline Cmed [1/mm]
• Maximum, minimum and median diameter along centreline Dmax, Dmin, Dmed [mm]
Meshing and smoothing
Alignment of input meshes
To reduce possible bias introduced by misaligned surface models, a two-step approach is proposed. First, each input shape was aligned (i.e. rigidly registered using translation and rotation only) to an initial reference shape using registration functions based on the iterative closest point (ICP) algorithm available in VMTK . The initial reference shape was determined as the closest shape to the centre or “mean” of the population (in this case subject CoA4; Fig. 9a) according to gross geometric parameters (volume V, surface area Asurf, centreline length LCL and median diameter along the centreline Dmed). Point-to-point correspondence between the reference mesh and respective subject meshes is not necessary as the correspondence will be updated at each iteration by finding the closest point.
The input meshes were re-aligned, with the reference shape this time being the computed template
A new template based on the newly aligned meshes was computed
The model compactness was computed as proposed by Styner et al. 
If the compactness was decreased, the reference shape was set to the new template and the procedure continued with step 1, otherwise the meshes were aligned sufficiently.
Here, sufficient alignment was obtained after one iteration of the outlined procedure.
A priori setting of the resolution and stiffness λ parameters
Generally, it is recommended to set the resolution parameter λW in the order of magnitude of the shape features to be captured ; however, clear indication for parameter setting is missing, in particular for the stiffness λV, which cannot be intuitively estimated. Following sensitivity analysis (Appendix 2), λW needs to be small enough to be able to capture all the features of interest, while being large enough to discard noise and to allow efficient template computation.
For the resolution λW, our approach can be interpreted as probing pW % of the smallest aorta surface area if it was cut open and laid out flat. Note that for large aortas the percentage drops below the chosen percentage values as the same parameters are applied to all shape models. Here, we set pW to 2.5 % and pV to 25 %, which yielded an initial λW of 15 mm and a λV of 47 mm, with the minimal surface area present in the set of shapes being Asurf,min = 8825 mm2. Those values were used to compute an initial template based on all 20 subjects. The initial template was then transformed towards the smallest subject (CoA3) while incrementally decreasing λW and λV in 1 mm steps until the matching error between source (initial template) and target (CoA3) was reduced by ≥80 %. A perfect (100 % error reduction) matching is not desired, as for example local shape differences due to segmentation errors or highly localised bulges are not of interest and thus do not need to be modelled. Note that the range of values for λV was fixed from 47 mm down to 40 mm in order to avoid too local deformations (Fig. 5). Starting from λW = 15 mm, transformations were computed in parallel for the range of λV values (47 to 40 mm). If the matching error was not reduced sufficiently by decreasing λV, then λW was decreased by 1 mm. In this way, we prioritised high λW values in order to ensure low runtimes for the final template calculation (Appendix 2). The matching error was determined by calculating the maximum surface distances between the target shape (subject CoA3) and the registered deformed source shape (the initial template). Following this procedure, a resolution of λW = 11 mm and a deformation stiffness of λV = 44 mm were found to sufficiently reduce the matching error and were used for all further template computations. The template, shape modes and shape vectors were then computed in Matlab based on the 20 arch surface models on a 32GB workstation using 10 cores (runtime for simultaneous template computation and transformation estimation approximately 15 h).
Controlling for confounders and influential observations
In this way, 3D shape features most related to size differences could be removed prior to analysing correlations of PLS shape vectors with geometric and clinical parameters normalised to BSA .
Detecting outliers or influential subjects
with yj being the jth fitted response variable and yj (i) being the jth fitted response variable if the fit does not include observation i; p is the number of coefficients in the regression model and MSE is the mean square error. The Cook’s distance was computed for each subject by leaving out the subject and performing PLS regression on the remaining subjects. PLS regression was thus repeated N times, with N being the number of subjects. Here, observations with Cook’s distances exceeding four times the mean Cook’s distance were discarded from the analysis as potentially influential observations.
Validation of the template - geometric approach
The overall deviation ∆Devtotal of the template from population means was calculated as the average of the deviations from each of the above mentioned parameters. A template shape yielding a low overall deviation ∆Devtotal from population mean values of below 5 % was considered to represent a good approximation of the mean shape.
In order to assure that the final template shape is not overly influenced by adding or leaving out a specific subject shape, k-fold cross-validation was performed . The entire dataset was divided into k = 10 randomly assigned subsets. The template calculation was run k times, each time leaving out one of the subsets until each patient had been left out once. As the entire set consists of N = 20 datasets in total, in each of the k runs N/k = 2 patients were left out. The 10 resulting templates should all be close to the template calculated on the full dataset of N = 20 patients. This was assessed visually by overlaying the final template meshes and numerically by measuring the surface distances between each of the 10 templates and the original template.
To back up the findings of the SSM, correlations between the parameters of interest, BSA and EF, with the traditionally measured geometric parameters and demographic parameters (patient age and height) were computed using bivariate correlation analysis. For correlations with BSA, non-indexed geometric shape descriptors were used. In a second step, shape vectors most related to BSA and EF (after removing size effects) were extracted via PLS and were correlated with the response variables BSA, EF, demographic parameters and the 2D and 3D geometric shape descriptors (Table 1). For parameters that were normally distributed, the standard parametric Pearson correlation coefficient r is reported. For non-normally distributed parameters, non-parametric Kendall’s τ is given. Non-normality was assumed if the Shapiro-Wilk test was significant. Parameters were considered significant (2-tailed) for p-values < .05. All statistical tests were performed in SPSS (IBM SPSS Statistics v.22, SPSS Inc., Chicago, IL).
Computed template and validation
Mean geometric parameters of the population compared to geometric parameters of the template
Surface Area ASurf [mm2]
Volume V [mm3]
Centreline Length LCL [mm]
Median Diameter Dmed [mm]
Mean population values
Template values (λW = 11 mm, λv = 44 mm)
Deviation from population values
Shape patterns associated with differences in BSA
Associations of geometric shape descriptors with changes in BSA
Correlations between BSA and BSA Shape Vector and traditionally measured parameters
Body Surface Area
BSA Shape Vector
N = 20
N = 19
Body Surface Area BSA [m2]
r = 0.707** (p = .001)
r = 0.705** (p = .001)
r = 0.696** (p = .001)
Height H [mm]
r = 0.838** (p < .001)
r = 0.872** (p < .001)
Volume V [mm3]
τ = 0.385* (p = .019)
τ = 0.743** (p < .001)
Surface Area Asurf [mm2]
r = 0.537* (p = .015)
r = 0.902** (p < .001)
Centreline length LCL[mm]
r = 0.398 (p = .083)
r = 0.853** (p < .001)
Centreline Tortuosity ToCL
r = 0.022 (p = .928)
r = 0.206 (p = .398)
r = −0.641** (p = .002)
r = −0.787** (p < .001)
Median Curvature Cmed [1/mm]
r = −0.603** (p = .005)
r = −0.718** (p = .001)
Maximum Diameter Dmax [mm]
τ = 0.460** (p = .005)
τ = 0.602** (p < .001)
Minimum Diameter Dmin [mm]
r = 0.628** (p = .003)
r = 0.763** (p < .001)
Median Diameter Dmed [mm]
r = 0.386 (p = .092)
r = 0.709** (p = .001)
Ascending Diameter Dasc [mm]
r = 0.550* (p = .012)
r = 0.708** (p = .001)
Transverse Diameter Dtrans [mm]
r = 0.453* (p = .045)
r = 0.646** (p = .003)
Isthmus Diameter Disth [mm]
r = 0.523* (p = .018)
r = 0.746** (p < .001)
Descending Diameter Ddesc [mm]
r = 0.332 (p = .152)
r = 0.740** (p < .001)
r = 0.264 (p = .260)
r = 0.025 (p = .918)
r = 0.190 (p = .422)
r = 0.315 (p = .189)
r = 0.364 (p = .115)
r = 0.278 (p = .250)
r = 0.074 (p = .757)
r = −0.188 (p = .442)
Arch height A [mm]
r = 0.198 (p = .402)
r = 0.632** (p = .004)
Arch width T [mm]
r = 0.555* (p = .011)
r = 0.626** (p = .004)
r = −0.170 (p = .473)
r = 0.116 (p = .637)
Associations of shape modes and shape vectors with changes in BSA derived from SSM
A first PLS regression of shape features with BSA revealed subject CoA20 to be influential to the regression as CoA20 exceeded the computed Cook’s distance threshold of 0.77. We considered CoA20 as an outlier in terms of its overall shape as it presented with a highly gothic (A/T ratio = 0.94) arch with a bended descending aorta (Fig. 9a) that is considerably larger than other subjects. Thus, CoA20 is likely to skew the subsequent shape feature extraction and was therefore removed from the following analyses.
The correlation between the associated BSA shape vector and BSA was significant with r = 0.707 (p = .001), implying that the BSA shape mode captured shape features associated with differences in BSA well (Fig. 10b). Furthermore, the computed BSA shape vector correlated positively and significantly with age (r = 0.696; p = .001) and height (r = 0.872; p < .001), volume V (τ = 0.743; p < .001) and surface area Asurf (r = 0.902; p < .001), centreline length LCL (r = 0.853; p < .001), diameters Dmax (r = 0.602; p < .001), Dmin (r = 0.763; p < .001), Dmed (r = 0.709; p = .001), Dasc (r = 0.708; p = .001), Dtrans (r = 0.646; p = .003), Disth (r = 0.746; p < .001), Ddesc (r = 0.740; p < .001) and arch height A (r = 0.632; p = .004) and width T (r = 0.626; p = .004) (Table 3). Significant negative correlations were found for the surface volume ratio Asurf/V (r = −0.787; p < .001) and the median curvature Cmed (r = −0.718; p = .001). Those associations were correctly depicted by the BSA shape mode (Fig. 10a).
Shape patterns associated with differences in EF
Associations of indexed geometric shape descriptors with changes in EF
Significant positive correlations were found between EF and the ratio of transverse and descending arch diameter Dtrans/Ddesc (r = 0.456; p = .050). EF correlated negatively and significantly with the indexed arch surface area iAsurf (r = −0.571; p = .011).
Associations of shape modes and shape vectors with changes in EF derived from SSM
Correlations between EF and EF Shape Vector and traditionally measured parameters
Ejection Fraction EF [%]
EF Shape Vector
N = 19
N = 19
Ejection Fraction EF [%]
r = 0.521* (p = .022)
Body Surface Area BSA [m2]
r = −0.147 (p = .548)
r = 0.000 (p = .999)
r = −0.243 (p = .316)
r = −0.142 (p = .561)
Height H [mm]
r = −0.391 (p = .098)
r = −0.246 (p = .310)
Volume iV [mm3/m2]
τ = −0.322 (p = .058)
τ = −0.228 (p = .172)
Surface Area iAsurf [mm2/m2]
r = −0.571*(p = .011)
r = −0.320 (p = .181)
Centreline length iLCL[mm/m2]
r = −0.255 (p = .293)
r = −0.338 (p = .157)
Centreline Tortuosity iToCL
r = −0.039 (p = .874)
r = 0.267 (p = .269)
r = −0.386 (p = .103)
r = 0.072 (p = .770)
Median Curvature iCmed [1/mm m2]
r = 0.269 (p = .265)
τ = 0.158 (p = .345)
Maximum Diameter iDmax [mm/m2]
τ = −0.096 (p = .574)
τ = 0.064 (p = .064)
Minimum Diameter iDmin [mm/m2]
r = −0.409 (p = .082)
r = −0.239 (p = .324)
Median Diameter iDmed [mm/m2]
r = −0.363 (p = .127)
τ = −0.228 (p = .172)
Ascending Diameter iDasc [mm/m2]
τ = 0.000 (p = 0.999)
τ = 0.170 (p = .310)
Transverse Diameter iDtrans [mm/m2]
r = 0.020 (p = .937)
r = −0.018 (p = .942)
Isthmus Diameter iDisth [mm/m2]
r = −0.407 (p = .083)
r = −0.441 (p = .059)
Descending Diameter iDdesc [mm/m2]
r = −0.442 (p = .058)
r = −0.527* (p = .020)
r = 0.312 (p = .193)
r = 0.735** (p < .001)
r = −0.362 (p = .128)
r = −0.453 (p = .052)
r = −0.152 (p = .535)
r = 0.083 (p = .736)
r = 0.456* (p = .050)
r = 0.457* (p = .049)
Arch height iA [mm/m2]
r = −0.154 (p = .529)
τ = −0.146 (p = .382)
Arch width iT [mm/m2]
r = 0.000 (p = .999)
r = 0.018 (p = .943)
r = −0.224 (p = .357)
r = −0.269 (p = .265)
This study describes and verifies a non-parametric statistical shape analysis method in detail and demonstrates its potential for discovering previously unknown 3D shape biomarkers in a complex anatomical shape population. The methodology is comprehensively explained from the user-perspective, with the aim of making the process more accessible to the broader research community. The shape analysis method was applied to CMR images of the aorta from patients post coarctation repair. The method computes a mean shape for this population of patients – the template – that we have shown to have good agreement with the conventional 2D and 3D measurements when averaged across the population (e.g. centreline length of the template = the average of the centreline length measured from each patient). Biomarker information – the shape features – for each individual were then extracted by transforming the mean aorta to each patient’s aorta. These extracted shape features, unique to each individual, were shown to: i) Accurately represent individual characteristics of the arch, as measured by patient-specific 2D/3D morphometric parameters, and ii) Have correlations with body surface area and left ventricular ejection fraction, offering the potential that they may be important biomarkers of biological processes. The found associations of aortic arch shape with ejection fraction were not known previously, which is why we consider the extracted 3D shape features as potential novel shape biomarkers that need to be confirmed by future studies. These results constitute the first statistical shape model of the aorta affected by coarctation.
A description of the statistical shape modelling framework adopted in this study is reported elsewhere in mathematically rather complex terms. Yet, in this paper we present the method from the user’s perspective. Here, we aimed to raise the awareness of the importance of necessary modelling parameters such as the meshing, smoothing and λ parameters for 3D shape analysis of complex anatomical structures. The mesh resolution for the input surfaces mainly affects the computational time needed to compute the template, but does not affect the final template shape substantially. Conversely, the analysis parameters (resolution λW and stiffness λV) affect both computational time and the final template shape considerably, requiring careful setting according to the shape population to be analysed. We provide tips on how to mesh input models and propose a new way of determining the λ parameters, which ensures robust and efficient template computation, even with an increased number of subjects for future studies. Furthermore, a modified PLS regression technique is described, which enables extraction of shape features independent of size differences between subjects. By measuring the Cook’s Distance during PLS regression, we were able to account for outliers such as one subject with an overly large, “abnormal” aortic shape and indeed a highly impaired cardiac function (EF = 17 %) that had to be excluded in order not to affect the shape feature extraction (subject CoA20). This suggests that the methodology could potentially be used to detect outlying shapes in a complex shape population – which, in turn, might be associated with outlying functional behaviour.
The calculated template based on the 20 CoA cohort showed characteristic shape features associated with CoA such as a slightly gothic arch shape, a dilated root, and a distinct narrowing in the transverse and isthmus arch section. The template shape was validated by comparing its geometry with the population average geometric parameters and by applying cross-validation techniques in order to ensure that removing or adding shapes had no influence. Therefore, new patients can be added easily, which involves performing the described pre-processing steps (segmenting, meshing, cutting, registration) and re-computing the template. Such a template could serve as a representative of the “normal of the abnormal”; a reference mean shape that might facilitate the diagnosis of highly abnormal cases within a pathologic shape population.
Three-dimensional global and regional shape features associated with differences in size (represented by BSA) and function (represented by EF) were extracted and found to be well in agreement with trends confirmed by traditional morphometrics. BSA correlated strongly and significantly with conventional geometric parameters, as expected. Those results confirmed the visual results shown by the SSM, whereby an increase in BSA was associated with an overall increase in aorta length and vessel diameters as well as with a shape development towards a slightly dilated root and a more gothic arch shape. For the first time, high EF was associated with a more compact, rounded arch shape with a slightly dilated aortic root and a slim descending aorta, whereas low EF was associated with a more gothic arch shape, a slim ascending aorta and a slightly dilated descending aorta, which may increase flow resistance across the arch and therefore left ventricular afterload.
Note however, that in order not to inflate Type II error of not detecting actual effects, computed correlation significances were not adjusted for multiple comparisons. Therefore, all results have to be considered as exploratory.
Analysing the found correlations in detail
Correlations with traditionally measured geometric parameters
Whereas BSA correlated strongly with multiple measured 2D and 3D shape descriptors, EF correlated significantly only with two geometrical parameters (the ratio of transverse to descending aortic diameter and the indexed surface area). One reason for this may be that the shape of the aortic arch marginally affects EF. However, these discrepancies could also emphasise that complex 3D shapes cannot always be sufficiently described by traditional individual morphometric measurements. Shape features associated with differences in body size between subjects are typically dominant and contribute to the largest portion of shape variability in natural pathologic shape populations . An increase in body size usually results in an overall size increase of the structure of interest, reflected in increased diameters and vessel length in the case of the aorta. This is why shape features associated with size differences are likely to be picked up by traditional 2D and 3D measurements. For the functional parameter EF though, we were interested in shape features independent of size effects, which, however, may be less prominent and may only be captured by a complex combination and collection of different morphometric parameters. Herein lies the power of 3D statistical shape modelling: results such as the mean shape and its variability are derived as visual, intuitively comprehensible and less biased 3D shape representations taking into account the entire 3D shape, instead of an unhandy collection of multiple measured parameters that might miss out crucial shape features.
Correlations with shape vectors describing shape features most related to a specific parameter in 3D
We found a strong significant correlation between the BSA shape vector and BSA, whereas EF correlated less with its EF shape vector. Overall, these results imply that shape features shown by the respective shape modes accounted well for differences in both BSA and EF in our shape population. In a strong correlation between functional parameter and shape vector, all subjects with low EF values would show those shape features given by the EF shape mode for low shape vector values, and vice versa for all subjects with high EF values. Nevertheless, those trends visually confirmed that our method was able to correctly extract 3D shape features from a population of shapes, which are potentially associated with a functional parameter of clinical relevance (Fig. 12). Therefore, the presented method can be used as a research tool to explore a population of 3D shapes, in order to detect where crucial shape changes occur and whether specific geometric parameters are likely to be of functional relevance.
Limitations and future work
The biggest limitation of our study is the small sample size of 20 subjects, with rather inhomogeneous characteristics in terms of age (range 11.1 to 20.1 years), age at arch intervention (4 days to 5 years after birth) and type of surgery . Thus, results presented in this work are primarily meant to demonstrate the potential of the proposed statistical shape modelling method by studying the association of complex 3D shape features with external, functional parameters such as EF. This could improve the derivation of novel shape biomarkers in future studies. In CoA patients, our method applied to a larger cohort of patients could help answer whether specific arch morphologies such as the gothic arch shape are associated with hypertension post-aortic coarctation repair [15, 37].
In this paper, we presented a non-parametric shape analysis method based on CMR data from the user-perspective and applied it to a population of aortic arch shapes of patients post-aortic coarctation repair. The process was described in detail in order to make it more accessible to researchers from both clinical and engineering background. The method has the potential of discovering previously unknown shape biomarkers from medical image databases and could thus provide novel insight into the relation between shape and function. Application to larger cohorts could contribute to a better understanding of complex structural disease, improving diagnosis and risk stratification, and could ultimately assist in the development of new surgical approaches.
2D, two-dimension(al); 3D, three-dimension(al); BSA, body surface area [m2]; CHD, congenital heart disease; CMR, cardiovascular magnetic resonance; CoA, coarctation of the aorta; EDV, end-diastolic volume [ml]; EF, ejection fraction [%]; ICP, iterative closest point algorithm; PDM, point distribution model; SSM, statistical shape model (ling); VMTK, the vascular modelling toolkit.
The authors would like to thank Marc-Michel Rohé from Inria Sophia Antipolis-Méditeranée for his help and assistance regarding the non-parametric shape analysis method.
MOCHA Collaborative Group: Andrew Taylor, Alessandro Giardini, Sachin Khambadkone, Silvia Schievano, Marc de Leval and T. -Y. Hsia (Institute of Cardiovascular Science, UCL, London, UK); Edward Bove and Adam Dorfman (University of Michigan, Ann Arbor, MI, USA); G. Hamilton Baker and Anthony Hlavacek (Medical University of South Carolina, Charleston, SC, USA); Francesco Migliavacca, Giancarlo Pennati and Gabriele Dubini (Politecnico di Milano, Milan, Italy); Alison Marsden (University of California, San Diego, CA, USA); Jeffrey Feinstein (Stanford University, Stanford, CA, USA); Irene Vignon-Clementel (INRIA,Paris,France); Richard Figliola and John McGregor (Clemson University, Clemson, SC, USA).
The authors gratefully acknowledge support from Fondation Leducq, FP7 integrated project MD-Paedigree (partially funded by the European Commission), Commonwealth Scholarships, Heart Research UK and National Institute of Health Research (NIHR).
Availability of data and materials
All relevant data supporting our findings are anonymised and stored at Great Ormond Street Hospital for Children.
JLB, KM and SS designed the study, contributed to the data analysis and interpretation and drafted the manuscript. GB and CC contributed to the data acquisition and revised critically the manuscript. HNN and AMT contributed to the data acquisition, analysis and interpretation and revised critically the manuscript. TYH, MS and XP contributed to the data analysis and interpretation and revised critically the manuscript. All authors read and approved the final manuscript.
All authors declare no relationships or activities that could appear to have influenced the submitted work and declare no competing interests.
This report incorporates independent research from the National Institute for Health Research Biomedical Research Centre Funding Scheme. The views expressed in this publication are those of the author(s) and not necessarily those of the NHS, the National Institute for Health Research or the Department of Health.
Ethics approval and consent to participate
Ethical approval was obtained by the Institute of Child Health/Great Ormond Street Hospital for Children Research Ethics Committee, and all patients or legal parent or guardian gave informed consent for research use of the data.
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
- Lamata P, Casero R, Carapella V, Niederer SA, Bishop MJ, Schneider JE, et al. Images as drivers of progress in cardiac computational modelling. Prog Biophys Mol Biol. 2014;115:198–212.View ArticlePubMedPubMed CentralGoogle Scholar
- Craiem D, Chironi G, Redheuil A, Casciaro M, Mousseaux E, Simon A, et al. Aging Impact on Thoracic Aorta 3D Morphometry in Intermediate-Risk Subjects: Looking Beyond Coronary Arteries with Non-Contrast Cardiac CT. Ann Biomed Eng. 2012;40:1028–38.View ArticlePubMedGoogle Scholar
- Young AA, Frangi AF. Computational cardiac atlases: from patient to population and back. Exp Physiol. 2009;94:578–96.View ArticlePubMedGoogle Scholar
- Lamata P, Lazdam M, Ashcroft A, Lewandowski AJ, Leeson P, Smith N. Computational mesh as a descriptor of left ventricular shape for clinical diagnosis. Computing in Cardiology Conference. 2013;2013:571–4.Google Scholar
- Cootes T, Hill A, Taylor C, Haslam J. Use of active shape models for locating structures in medical images. Image Vision Computing. 1994;12:355–65.View ArticleGoogle Scholar
- Remme E, Young AA, Augenstein KF, Cowan B, Hunter PJ. Extraction and quantification of left ventricular deformation modes. IEEE Trans Biomed Eng. 2004;51:1923–31.View ArticlePubMedGoogle Scholar
- Lewandowski AJ, Augustine D, Lamata P, Davis EF, Lazdam M, Francis J, et al. Preterm Heart in Adult Life Cardiovascular Magnetic Resonance Reveals Distinct Differences in Left Ventricular Mass, Geometry, and Function. Circulation. 2013;127:197–206.View ArticlePubMedGoogle Scholar
- Styner MA, Rajamani KT, Nolte L-P, Zsemlye G, Székely G, Taylor CJ, et al. Evaluation of 3D Correspondence Methods for Model Building. In: Taylor C, Noble JA, editors. Information Processing in Medical Imaging. Berlin: Springer; 2003. p. 63–75. Available from: http://link.springer.com/chapter/10.1007/978-3-540-45087-0_6.View ArticleGoogle Scholar
- Vaillant M, Glaunès J. Surface Matching via Currents. In: Christensen GE, Sonka M, editors. Information Processing in Medical Imaging. Berlin: Springer; 2005. p. 381–92. Available from: http://link.springer.com/chapter/10.1007/11505730_32.View ArticleGoogle Scholar
- Durrleman S, Pennec X, Trouvé A, Ayache N. Measuring brain variability via sulcal lines registration: a diffeomorphic approach. Med Image Comput Comput Assist Interv. 2007;10:675–82.PubMedGoogle Scholar
- Mansi T, Durrleman S, Bernhardt B, Sermesant M, Delingette H, Voigt I, et al. A Statistical Model of Right Ventricle in Tetralogy of Fallot for Prediction of Remodelling and Therapy Planning. In: Yang G-Z, Hawkes D, Rueckert D, Noble A, Taylor C, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2009. Berlin: Springer; 2009. p. 214–21. Available from: http://link.springer.com/chapter/10.1007/978-3-642-04268-3_27.View ArticleGoogle Scholar
- Mansi T, Voigt I, Leonardi B, Pennec X, Durrleman S, Sermesant M, et al. A Statistical Model for Quantification and Prediction of Cardiac Remodelling: Application to Tetralogy of Fallot. IEEE Trans Med Imaging. 2011;30:1605–16.View ArticlePubMedGoogle Scholar
- Leonardi B, Taylor AM, Mansi T, Voigt I, Sermesant M, Pennec X, et al. Computational modelling of the right ventricle in repaired tetralogy of Fallot: can it provide insight into patient treatment? Eur Heart J Cardiovasc Imaging. 2013;14:381–6.View ArticlePubMedPubMed CentralGoogle Scholar
- Durrleman S, Pennec X, Trouvé A, Ayache N. Statistical models of sets of curves and surfaces based on currents. Med Image Anal. 2009;13:793–808.View ArticlePubMedGoogle Scholar
- O’Sullivan J. Late Hypertension in Patients with Repaired Aortic Coarctation. Curr Hypertens Rep. 2014;16:1–6.Google Scholar
- Vergales JE, Gangemi JJ, Rhueban KS, Lim DS. Coarctation of the Aorta - The Current State of Surgical and Transcatheter Therapies. Curr Cardiol Rev. 2013;9:211–9.View ArticlePubMedPubMed CentralGoogle Scholar
- Ou P, Bonnet D, Auriacombe L, Pedroni E, Balleux F, Sidi D, et al. Late systemic hypertension and aortic arch geometry after successful repair of coarctation of the aorta. Eur Heart J. 2004;25:1853–9.View ArticlePubMedGoogle Scholar
- De Caro E, Trocchio G, Smeraldi A, Calevo MG, Pongiglione G. Aortic Arch Geometry and Exercise-Induced Hypertension in Aortic Coarctation. Am J Cardiol. 2007;99:1284–7.View ArticlePubMedGoogle Scholar
- Lee MGY, Kowalski R, Galati JC, Cheung MMH, Jones B, Koleff J, et al. Twenty-four-hour ambulatory blood pressure monitoring detects a high prevalence of hypertension late after coarctation repair in patients with hypoplastic arches. J Thorac Cardiovasc Surg. 2012;144:1110–8.View ArticlePubMedGoogle Scholar
- Durrleman S. Statistical models of currents for measuring the variability of anatomical curves, surfaces and their evolution. University of Nice-Sophia Antipolis; 2010. Available from: https://tel.archives-ouvertes.fr/tel-00631382/.
- Mansi T. Modèles physiologiques et statistiques du cœur guidés par imagerie médicale: application à la tétralogie de Fallot [Internet]. École Nationale Supérieure des Mines de Paris; 2010 [cited 2013 Aug 21]. Available from: http://tel.archives-ouvertes.fr/tel-00530956
- McLeod K, Mansi T, Sermesant M, Pongiglione G, Pennec X. Statistical Shape Analysis of Surfaces in Medical Images Applied to the Tetralogy of Fallot Heart. In: Cazals F, Kornprobst P, editors. Modeling in Computational Biology and Biomedicine. Berlin: Springer; 2013. Available from: http://link.springer.com/content/pdf/10.1007%2F978-3-642-31208-3_5.pdf#page-1.Google Scholar
- Durrleman S, Prastawa M, Charon N, Korenberg JR, Joshi S, Gerig G, et al. Morphometry of anatomical shape complexes with dense deformations and sparse parameters. Neuroimage. 2014;101:35–49.View ArticlePubMedPubMed CentralGoogle Scholar
- Ntsinjana HN, Biglino G, Capelli C, Tann O, Giardini A, Derrick G, et al. Aortic arch shape is not associated with hypertensive response to exercise in patients with repaired congenital heart diseases. J Cardiovasc Magn Reson. 2013;15:101.View ArticlePubMedPubMed CentralGoogle Scholar
- Schievano S, Migliavacca F, Coats L, Khambadkone S, Carminati M, Wilson N, et al. Percutaneous Pulmonary Valve Implantation Based on Rapid Prototyping of Right Ventricular Outflow Tract and Pulmonary Trunk from MR Data. Radiology. 2007;242:490–7.View ArticlePubMedGoogle Scholar
- Casciaro ME, Craiem D, Chironi G, Graf S, Macron L, Mousseaux E, et al. Identifying the principal modes of variation in human thoracic aorta morphology. J Thorac Imaging. 2014;29:224–32.View ArticlePubMedGoogle Scholar
- Bosmans B, Huysmans T, Wirix-Speetjens R, Verschueren P, Sijbers J, Bosmans J, et al. Statistical Shape Modeling and Population Analysis of the Aortic Root of TAVI Patients. J Med Devices. 2013;7:040925.View ArticleGoogle Scholar
- Zhao F, Zhang H, Wahle A, Thomas MT, Stolpen AH, Scholz TD, et al. Congenital Aortic Disease: 4D Magnetic Resonance Segmentation and Quantitative Analysis. Med Image Anal. 2009;13:483–93.View ArticlePubMedPubMed CentralGoogle Scholar
- Antiga L, Piccinelli M, Botti L, Ene-Iordache B, Remuzzi A, Steinman DA. An image-based modeling framework for patient-specific computational hemodynamics. Med Biol Eng Comput. 2008;46:1097–112.View ArticlePubMedGoogle Scholar
- Piccinelli M, Veneziani A, Steinman DA, Remuzzi A, Antiga L. A framework for geometric analysis of vascular structures: application to cerebral aneurysms. IEEE Trans Med Imaging. 2009;28:1141–55.View ArticlePubMedGoogle Scholar
- Antiga L, Steinman DA. Robust and objective decomposition and mapping of bifurcating vessels. IEEE Trans Med Imaging. 2004;23:704–13.View ArticlePubMedGoogle Scholar
- Besl PJ, McKay ND. A method for registration of 3-D shapes. IEEE Trans Pattern Anal Mach Intell. 1992;14:239–56.View ArticleGoogle Scholar
- Heimann T, Meinzer H-P. Statistical shape models for 3D medical image segmentation: A review. Med Image Anal. 2009;13:543–63.View ArticlePubMedGoogle Scholar
- Singh N, Thomas Fletcher P, Samuel Preston J, King RD, Marron JS, Weiner MW, et al. Quantifying anatomical shape variations in neurological disorders. Med Image Anal. 2014;18:616–33.View ArticlePubMedGoogle Scholar
- Mathworks. MATLAB v2014 Documentation - Cook’s Distance. Natick, MA. 2014;Google Scholar
- Joliffe IT. Principal Component Analysis. 2nd ed. Inc.: Springer-Verlag New York; 2002.Google Scholar
- Canniffe C, Ou P, Walsh K, Bonnet D, Celermajer D. Hypertension after repair of aortic coarctation — A systematic review. Int J Cardiol. 2013;167:2456–61.View ArticlePubMedGoogle Scholar
- Lützen J. De Rham’s Currents. In The Prehistory of the Theory of Distributions. [Studies in the History of Mathematics and Physical Sciences, vol. 7]. Springer New York; 1982:144–7.Google Scholar
- Beg MF, Miller MI, Trouvé A, Younes L. Computing Large Deformation Metric Mappings via Geodesic Flows of Diffeomorphisms. Int J Comput Vision. 2005;61:139–57.Google Scholar
- Bookstein FL. Principal warps: thin-plate splines and the decomposition of deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence. 1989;11:567–85.Google Scholar