Biomedical Engineering Reference
In-Depth Information
For example, Lorigo
et al
. [12, 13] use codimension two GACs for the ex-
traction of the cerebral vasculature. This method successfully obtained a seg-
mentation of the whole tree. However, the final segmentation shows relatively
thinner vessels compared to MIP images and thresholding schemes, and the
model seems to be not able to capture abnormal vessel lumina.
Other authors propose to combine implicit deformable models with a smart
initialization of the evolving surface inside the vessels of interest. Van Bemmel
et al
. [14, 15], for instance, compute the central axis of the artery and use it as
initialization for the GAC model in the segmentation of carotid arteries and the
aorta. Deschamps
et al
. [16] use the output of a vessel enhancement filter [17]
as speed for the Fast Marching method for a fast and rough initialization of the
model. The evolved surface is used as initialization of the implicit deformable
model to obtain a more accurate segmentation.
Classical deformable models depend on the gradient of the image as stopping
force. The front in evolution usually suffers from leakage or bleeding in places
with weak or inhomogeneous image gradient, and does not provide good results
in brain vessels. Since the method of region competition proposed by Zhu and
Yuille [18], there have been several efforts to include statistical region-based
information in the process of segmentation. Paragios made a wide study about
the inclusion of this information in the GAC model [19]. In places with weak
gradient, regional information drives the evolution of the front, thus avoiding
leakage in the edges of the object of interest. Similar work including statistical
information on the implicit model was done by Yezzi
et al
. [20]. Deschamps
successfully used Paragios' Geodesic Active Regions (GAR) model in the seg-
mentation of brain aneurysms [16].
Although a number of algorithms based on implicit deformable models have
addressed the problem of cerebral vessels segmentation [13, 16], these do not
produce satisfactory results when confronted with images of standard qual-
ity in average radiology departments. For example, the work reported by De-
schamps [16] deals with rotational angiography (3DRA) where the background
and bone tissues have a well differentiated contrast with respect to vessels
(Fig. 5.5a). On the other hand, the ranges of vessel and bone intensities in CTA
ususally overlap (Fig. 5.5b). Most of the previous attempts to solve this problem
have presented segmentation results only on few selected images. Unfortunately,
there is a general lack of larger evaluation studies on image databases acquired
under routine clinical conditions.
