Biomedical Engineering Reference
In-Depth Information
segmentation method was very strong for both the entire frontal lobe and only
the GM but the level set algorithm systematically underestimated the frontal
lobe volume with a mean difference of 4%.
4.
In a fourth experiment, regional cortical thickness was quantitatively ana-
lyzed on seven high-resolution MRI data sets acquired with SPGR and isotropic
voxel size of 1.2 mm
3
. Cortical measurements were performed in four lobes to
compare level set segmentation and expert manual tracing. The authors first
compared the mean thickness of each lobe to the results from the study of 63
postmortem males by Pakkenberg and Gundersen [104]. The new segmentation
produced similar results with a frontal cortex thicker than the occipital cortex.
On overall, postmortem data was 5-14% thinner than the new results while vari-
ability of thickness was equal to 1.5 mm for both studies. Statistical tests also
showed that frontal and temporal lobes were thicker than parietal and occipital
lobes.
This method relies on prior statistics for WM, GM, and CSF for the construc-
tion of interface probability maps. The method also requires manual initialization
by clicking center points of concentric spheres on slices. The authors demon-
strated on one example that the number of spheres or their localization does
not have a critical influence on the accuracy or reproducibility of the method.
On the other hand, major advantages of the method include the fact that it does
not require stripping of non-brain data, its performance was demonstrated on a
wide range of MRI image quality and it offers fast computation times compared
to existing methods.
2.4.3.4
Segmentation of RT3D Ultrasound with Geodesic
Deformable Models
This research was published by Corsi
et al.
in [105].
Method:
The proposed method uses a geodesic deformable model as defined
in Eq. (2.17), with weighting parameters associated with the two terms and a
function
g
()defined as in [8]. To handle the noisy nature of the ultrasound data
and the poor definition of the myocardium wall borders in some frames, the
authors did not use any inflationary force in the speed term. This type of model
requires an initialization close to the final endocardial surface to ensure that the
moving front is attracted to local high data gradient locations. The algorithm
was implemented with the narrow-band technique for computation efficiency.
