Generative model of heart growth using Exoshape

The remodeling model

PIC PIC
Rigid alignmentNon-linear registration to the template

Figure 1: 3D RV meshes of 18 young ToF patients. Left panel: The meshes were rigidly registered to a representative patient of the dataset. Observe the extreme variability in shape (see companion video). Right panel: The same meshes registered to the template using the non-linear deformations estimated during the template creation.


1. Characterising Deformation Modes of RV Shapes in ToF

In this work we analyse the deformations ϕi only as we mainly focus on the regional changes of the RV anatomy due to ToF. Principal Component Analysis (PCA) is performed directly on the moments βi to extract the main deformation modes. The elements of the covariance matrix Σ are given by Σij =< v0i -v0,v0j -v0 > V = xk,xl(βi(xk) -β(xk))KV (xk,xl)(βj(xl) -β(xl)), xk being the position of the kth Dirac delta current of T. Then, the moment vector γm of the initial speed vector u0m related to the mth deformation mode is given by γm = β + iVm[i](βi -β). In this equation, Vm is the mth eigenvector of Σ when the eigenvalues are sorted in decreasing order. Finally, the RV shape of each patient i is characterised by the shape vector si = {smi}m=1..M computed from the M first deformation modes, smi =<  v0i,u0m > V = xk,xlβi(xk)KV (xk,xl)γm(xl).

2. Can We Predict the Shape from Clinical Parameters?

First, cross-sectional analysis of the impact of growth on RV shape was performed. Multiple linear regression between the shape vectors si and body surface area (BSA) was carried out to exhibit the effects of BSA on each deformation mode. An optimal set of modes was selected by iteratively removing the modes with lowest significance, until the p-value of the regression overall significance stopped decreasing. Canonical Correlation Analysis (CCA) was then applied to quantify the amount of variation of each mode when BSA varies. Denoting R the overall correlation coefficient between BSA and shape vectors and ρ the correlation vector relating BSA to each deformation mode, the moments μ of the generative deformation Φ are μ = R  kρ[k]γk. Deforming the template T with Φ enables quantifying the average RV remodelling observed in our population.

3. Statistical Model of RV Remodelling in ToF Patients

Patient growth was quantified by body surface area (BSA) index (correlation with age in the data set: R2 > 0.5, p < 0.001). Model reduction discarded all the non-significant modes. The remaining modes were found clinically pertinent by an expert after visual inspection. Mode 1 clearly represented the overall RV dilation. Mode 2 seemed to model the dilation of the tricuspid annulus and of the inflow tract. Mode 3, 6, 7 and 9 exhibited a dilation of a specific RV region: apex (mode 3), basal area under the tricuspid valve (mode 6), apical area of the outflow tract (mode 7) and outflow tract (mode 9), reflecting possible direct impact of regurgitations on the neighbouring tissues.


10 first deformation modes extracted by PCA on a population of 18 patients suffering from repaired Tetralogy of Fallot.


Figure 2: Average template of the right ventricle observed in a population of 18 Tetralogy of Fallot.

Canonical Correlation Analysis (CCA) provided a generative model of the RV remodelling observed in our population. Overall correlation coefficient with BSA was R = 0.87, suggesting a strong correlation between these deformation modes and growth. The correlation vector of the deformation modes was ρ = {-0.56,0.45,-0.35,-0.33,-0.33,-0.37}. When BSA increases by 0.86, each deformation mode m varies by its related coefficient ρ[m]. The model was found clinically realistic by an expert (See video below). As BSA increased, RV volume increased, RV free-wall and valves dilated, and septum was more concave. Indeed, dilation of the valves reduces the remaining pulmonary obstructions, thus decreasing the RV pressure. As a result, left-ventricle pushes the septum towards the right ventricle, making it more concave. However, as regurgitations are still present, the RV still dilates by pushing the RV free wall outwards.

Mean RV remodelling observed in our population when body surface area (BSA, in m2) increases. RV dimensions globally increase while valves dilate. Simultaneously, RV free wall becomes rounder and septum more concave.

4. Validating the Generalisation of the Statistical Models

Generalising a statistical model of RV remodelling is crucial for patient management and therapy planning. We thus tested the robustness of our model on two new patients with matched age (13 and 16). The template was registered to the patients and the related shape vectors s were computed. BSA were calculated from the optimal linear model estimated in Sec. 3.. Results successfully compared with measured values (patient 1: estimated BSA: 1.61, measured BSA: 1.49; patient 2: estimated BSA: 1.29, measured BSA: 1.16). This suggests that the deformation modes involved in the RV remodelling model could be generalised, constituting potential quantitative parameters of remodelling in ToF.


References

Tommaso Mansi, Stanley Durrleman, Boris Bernhardt, Maxime Sermesant, Hervé Delingette, Ingmar Voigt, Philipp Lurz, Andrew M Taylor, Julie Blanc, Younes Boudjemline, Xavier Pennec, and Nicholas Ayache. A Statistical Model of Right Ventricle in Tetralogy of Fallot for Prediction of Remodelling and Therapy Planning. In Proc. Medical Image Computing and Computer Assisted Intervention (MICCAI'09), volume 5761 of Lecture Notes in Computer Science, London, UK, pages 214-221, September 2009. Springer. Download paper