Personalised Simulation of Myocardium Electromechanics
and Pulmonary Valve Replacement Surgery in
Repaired Tetralogy of Fallot: a Case Study
2. Creating a Personalised Anatomical Model of the Heart
Simulating patient cardiac function cannot be performed without considering the geometry of patient heart. Therefore, we need to extract from clinical gated cine-MRI a reliable representation of the biventricular myocardium of the patient. Furthermore, having 3D+t delineation enables us estimating dynamic measurements such as blood pool volume variations, which help us in personalising the electromechanical model.
The anatomical model comprises several information regarding the patient heart. First, it represents the shape of the organ. This can be achieved by segmenting the myocardium from clinical cine-MRI. Next, it defines the anatomical regions that are characterised by different motion patterns, for regional personalisation of the electromechanical model. Finally, it encodes the direction of the myocardium fibres to simulate the anisotropic properties of the myocardium.
2.1. Delineating Myocardium Shape
Numerous methods have been proposed in the literature to efficiently segment the heart from cine MRI (see Part III). However, these methods are not readily adapted to our application. In children, myocardium is thinner, images are less contrasted and motion artifacts are more frequent. Moreover, repaired tetralogy of Fallot leads to abnormal myocardium motion patterns, extreme dilation and tremendous variability of the right ventricle shape. Using atlas-based methods is thus not appropriate, these approaches may fail because of the paediatric image specificities. To tackle these challenges, we devised specific tools that requires few manual interactions.
Delineation of the right-ventricle endocardium
In Health-e-Child, the right-ventricle endocardium is segmented on all the frames of the cardiac sequence by fitting an anatomically accurate geometrical model. Its position, orientation and scale in the images are determined using minimal user interaction. Boundaries are locally adjusted by training a probabilistic boosting tree classifier with steerable features (Zheng et al., ICCV 2007). The resulting delineation is then tracked throughout the cardiac cycle using an optical flow method.
|
|
Delineation of the left-ventricle endocardium and of the epicardium
Interactive delineation of the left endocardium. The surface (in wireframe) is modelled in real-time by interactively placing control points inside (green point), on (red points) and outside (blue points) the surface.
The left-ventricle endocardium and the epicardium are delineated on the first frame of the gated cine MRI by using an interactive tool based on variational implicit functions (Turk et al., Technical Report 1999). The user places control points inside, on and outside the desired surface. The algorithm computes in real-time the implicit function that interpolates those points and extracts its 0-level set (Toussaint et al., VCBM 2008).
Once the desired surface is modelled, diffeomorphic non-linear registration (Vercauteren et al., MICCAI 2007) is applied to automatically propagate the surfaces throughout the cardiac cycle.
Construction of the Dynamic Bi-Ventricular Myocardium
The surfaces computed in the previous steps are combined together to get the dynamic segmentation of the myocardium. To that aim, we combine the binary masks of the surfaces, thus obtaining a dynamic mask of the biventricular myocardium. The valve plane is manually defined using CardioViz3D and muscle consistency is ensured by preserving a minimal thickness of 3 mm (mean thickness of a healthy right ventricle myocardium).
Next, we extract the 3D surface of the myocardium on the first frame. Finally, simplex-based deformable surfaces (Montagnat et al., MedIA 2004) are used to propagate the myocardium mesh throughout the cardiac cycle.
|
|
Quantitative Measurements
The dynamic contours we have obtained enable us to compute quantitative and reproducible measurements related to the cardiac function. In particular, blood pool volumes can be measured at each time frame. For the patient we are considering, ejection fraction is found normal for the left ventricle (61%) but slightly below the normal values for the right ventricle (41%). However, right ventricle volume is twice as great as left ventricle volume (see curve below).
Additionally, the 3D representation of the beating heart enables us to estimate 3D indexes that are not readily available in clinical routine. For instance, radial displacements of the myocardium, computed at each vertex of the 3D mesh, highlight the abnormal motion of the right-ventricle outflow tract. This region dilates when the healthy myocardium contracts. Such observation is of crucial interest when personalising the electromechanical model. We can indeed locally personalise this region by choosing appropriate parameters in order to recover the pathological motion.
Volume variations throughout the cardiac cycle of
the left (in red) and right (in blue) ventricles |
|
2.2. Personalised Myocardium Anatomy
The anatomical model can now be defined. From the volume curves, we select the mid-diastole time frame. The related surface mesh is transformed into a tetrahedral volume mesh. From the endocardia and epicardium surfaces, left and right ventricles are automatically detected and defined in the volumetric mesh. At last, the pathological region is specified in the anatomical model to simulate its pathological motion.
2.3. Myocardium Fibres
Myocardium fibres play a crucial role in cardiac biomechanics. Unfortunately, measuring fibre direction in-vivo is still difficult. To tackle this limitation, we use a computational model based on observations done on anatomical dissections or post-mortem diffusion tensor images. These studies showed that fibre orientation varies from -70° on the epicardium to 0° at mid-wall to +70° on the endocardium (Arts et al., 2001, Peyrat et al., TMI 2007, Sanchez-Quintana et al., 1996). Synthetic fibres are thus created by linearly interpolating their orientation with respect to the short axis plane, from -90° on the epicardium to 0° at mid-wall to +90° on the endocardium. The angles are slightly over-estimated to account for rasterisation effects.
Anatomical model of the patient heart. For regional
personalisation of the electromechanical model, we define in red the left ventricle, in orange the right ventricle and in white the observed dyskinetic area. |
Myocardium fibres estimated using a computational model
adapted to the pathology (Sanchez-Quintana et al., 1996) |
2.4. Conclusions
From the cine-MRI of the patient, we have semi-automatically delineated the biventricular myocardium. Then, from this segmentation, we have built an anatomical model of the patient heart, which represents the shape of the organ, defines some regions of interest in view of regional personalisation of the electromechanical model, and sets the direction of the myocardium fibres. Furthermore, the dynamic segmentation has provided us with volume curves and radial displacements, measurements which will help us in personalising the model, as described in the next section.
References
- Arts, T., Costa, K. D., Covell, J. W., and McCulloch, A. D.: Relating myocardial laminar architecture to shear strain and muscle fiber orientation, Am J Physiol Heart Circ Physiol 280(5), 2001
- Montagnat, J., and Delingette, H.: 4D Deformable Models with temporal constraints: application to 4D cardiac image segmentation, Medical Image Analysis 9(1), 87-100, 2005
- Peyrat, JM., Sermesant, M., Pennec, X., Delingette, H., Xu, C., McVeigh, E. R., and Ayache, N.: A Computational Framework for the Statistical Analysis of Cardiac Diffusion Tensors: Application to a Small Database of Canine Hearts, IEEE TMI 26(11), 1500-1514, 2007

- Sanchez-Quintana, D., Anderson, R. H., and Ho, S. Y.: Ventricular myoarchitecture in tetralogy of Fallot., Heart 76(3), 280-286, 1996
- Toussaint, N., Mansi, T., Delingette, H., Ayache, N., and Sermesant, M.: An Integrated Platform for Dynamic Cardiac Simulation and Image Processing: Application to Personalised Tetralogy of Fallot Simulation, Proc. Eurographics Workshop on Visual Computing for Biomedicine (VCBM), 2008

- Turk, G., and O'Brien, J.: Variational Implicit Surfaces, Georgia Institute of Technology, 1999
- Vercauteren, T., Pennec, X., Perchant, A., and Ayache, N.: Non-parametric Diffeomorphic Image Registration with the Demons Algorithm, Proc. MICCAI 2007
- Zheng, Y., Barbu, A., Georgescu, B., Scheuering, M., and Comaniciu, D.: Fast Automatic Heart Chamber Segmentation from 3D CT Data Using Marginal Space Learning and Steerable Features, Proc. ICCV 2007, 1-8, 2007




