|
We propose here to extract automatically all branches in the aorta (including the two external iliac arteries) in this MR scan. A contrast solution fills the aorta which shows a typical pathology: an abdominal aortic aneurysm. This data was also used for virtual inspection. Our method just need the user to initialize a start point for the tree extraction in the 3D dataset using those 3 orthogonal views.
|
In the first case, the constraint on the front is not sufficient and it evolves outside the aorta. Using our pruning method, the front propagates inside the aorta, without crossing the edges of the object.
|
The front propagates in the aorta, and the edges are pruned using a criterion based on the distance to the propagating front. The total computing cost does not exceed 10 seconds on a 300MHz workstation.
|
The end of the tubular structure is automatically detected using a criterion based on the geodesic distance to the starting point of the front propagation. The slope decreases significantly when the front crosses the walls of the aorta. Representing the maximum distance to the starting point across time enables to accurately determine the stopping time.
|
In order to increase the value of the vessel segmentation, we use a potential based on the hessian eigenvalues (see the work of Alejandro Frangi on the Multiscale vessel enhancement filtering).
|
We compute a measure based on the eigenvalues of the hessian matrix at different scales, and we take as final measure the maximum response of the filter across each scale. This measure is used in Eikonal equation as the speed of the front propagation.
|
|
The end points are automatically extracted by our front propagation approach. The user just need to indicate what is the minimum size of the branches to be detected.
|
If we compute the distance to the object walls, we are now able to use this distance as a new "potential" to extract our trajectories. Thus the final paths are centred inside the aorta. The whole set of paths is represented superimposed on the dataset.
 |
 |
|
| Initialization of the Segmentation with the Fast-Marching | Convergence to the final Segmentation | Restoring the distance function |
The segmentation given by the fast-marching is now used as initialization for a level-set deformable model. This model that uses more complicated schemes and allows curvature dependent speeds will converge to a sub-pixel representation of the contour. The huge amount of computations involved is relaxed by the fast-marching which gives an accurate initialization close to the solution. Therefore, the algorithm converges in a small number of iterations. The final segmentation is a very precise definition of the boundaries of the object, and can lead easily to valuable measures on the volume, or surface of parts or the complete object of interest.
The segmentation obtained is very useful for visual inspection of the inside of the aorta. It allows diagnostic assessment of the extent of an aneurysm and its spatial relationship to branches of the aorta, which is crucial for planning the surgical approach. Virtual endoscopy can be used to assess stenosis and aneurysm from an intraluminal perspective. It is also useful for virtual planning of the placement of endoluminal aortic stents. It allows easy determination of vessel diameters and stent positionning.
|
|
|
| Volume-based virtual endoscopy | Surface-based virtual endoscopy | Wire frame view of the surface |
Volume rendering is usually based on ray-tracing techniques, with a user-defined opacity function. But a threshold on the data is not sufficient when dealing with contrast enhanced images, where the grey level of the aorta critically depends on the contrast product which dilutes during acquisition and is related to the local section of the aorta. This volume-based rendering can produces holes in the walls of the structure of interest.
The segmentation allows us to reduce user interaction, using a surface-based rendering which automatically sets the opacity of the several structures detected. Therefore, the clinician is given a view of the object detected automatically by the algorithm. The automation and the robustness of the method allows the reproducibility of this kind of inspection, leading to reproducibility and assessment in the diagnostics.