Health-e-Child - IST-2004-027749 - Deliverable D.11.4

Heart Diseases

Towards Patient-Specific Heart Models for Decision Support System

1. Current method: From MRI segmentation to 3-D anatomical model of the heart

Tetrahedral mesh of the bi-ventricular myocardium
obtained in (Sermesant et al., 2006)

To adapt the general disease-based model to the patient's anatomy, the most common approach consists in using 3-D imaging to obtain, through image analysis algorithms, a 3-D mesh of the patient's heart.

In (Sermesant et al., 2006) for instance, the 3-D anatomical model used in the disease-based model is created directly from a high-resolution MRI by segmenting the myocardium through classical segmentation techniques (thresholding and mathematical morphology). The marching cube algorithm (Lorensen et al., 1987) is then applied to get a triangulated surface mesh, which is then converted into a volumetric tetrahedral mesh by using the INRIA software GHS3D (see opposite figure).

Similar methods can be applied to the images that are acquired in Health-e-Child. The figures below show an example of volumetric mesh created from a high-resolution MRI of a patient suffering from right-ventricle overload (University College London, Great Ormond Street Children's Hospital) with a supervised algorithm developed at INRIA (Fernandez., 2006). This algorithm allows the segmentation of the four cardiac chambers. The input MRI is first pre-processed through anisotropic filtering to reduce the noise without blurring the edges. Then, geometrical primitives of the cavities are estimated from a preliminary segmentation of the blood pools (via a region-growth method and mathematical morphology). These primitives are then converted into simplex meshes which are automatically deformed according to internal and image-based forces.

Original cardiac MRI acquired at Great Ormond Street Children's Hospital.
Segmentation of the heart from 3-D MRI.
Left panel: Final segmentation. Right panel: A clipping plane is used to show up the myocardium (semi-transparent).
In red: left ventricle, in blue: right ventricle.
Projection of the final mesh into the input MRI. In red, segmentation of the heart.

INTERACTIVE 3D MESH

Interactive 3-D mesh. Use the mouse to interact with the mesh

A last method would consist in adapting the bi-ventricular model introduced in the disease-based model to the patient's anatomy. In (Sermesant et al., 2005) for example, the bi-ellipsoidal model of the two ventricles was first generated. Then, the ventricular blood pools were segmented by using boundary- and region-based fuzzy classification (Andriantsimiavona et al., 2003). The resulting segmentation was automatically registered to the model through affine transformation. Finally, local adjustments were performed by using a deformable biomechanical model (Sermesant et al., 2003) and by taking as initial estimation the affine transformation previously computed. The following figures present a result obtained by using this technique.

Bi-ellipsoidal model of heart ventricles (Sermesant et al., 2005).
Adjustment of the model to the patient's anatomy.
(a) Semi-automated segmentation of the ventricular blood pools (white lines).
(b) Affine registration of the segmentation result to the synthetic model.
(c) Local adjustment of the model to the segmentation.

All these methods are nowadays able to extract the heart anatomy from 3-D imaging. The results are satisfying and can be used to simulate the heart function of the patient under study. We plan to make use of such methods to adapt the disease-based model to the patient's anatomy when high-resolution MRI are available.

However, these approaches are sensitive to the quality of the underlying images. If the 3-D image is too noisy or has a slice thickness too high for instance, the segmentation algorithms may have difficulties in extracting the blood pools and the myocardium. These techniques are therefore well adapted to high-resolution MRI or CT but cannot be applied on ultrasonography imaging nor in some 2-D MR images. Consequently, other approaches are required to adapt the bi-ventricular geometrical model to the patient's anatomy when such high-resolution images are not available. Relying on this idea, next section presents preliminary concepts about the generation of anatomical model from sparse 1-D clinical parameters.

References