Biomedical Engineering Reference
In-Depth Information
The use of non-parametric statistical information provides more accurate seg-
mentations with minimal sensitivity to the selection of the parameters. In the
case of Gaussian distributions, the region descriptors are time and front depen-
dent. In contrast, the non-parametric approach presented in this work does not
impose these two constraints.
The features used by the kNN rule are computed at a single scale. It would
seem that, due to the nature of the object to be segmented, a multiscale ap-
proach should provide better results. However, it was observed that results
were worse than when using a single scale. This could be explained by the fact
that as the number of scales increases, the dimensionality of the feature space
also increases. This may deteriorate the performance of the classifier owing to
the
peaking phenomenon
[34]. We are currently working on improvements on
the PDF estimation technique using a multiscale approach and dimensionality
reduction strategies.
The aneurysms involved in the study had a mean size of 2.81 mm for the
neck diameter, 5.40 mm for the width, and 6.44 mm for the depth with standard
deviations of 0.84, 2.95 and 3.10 mm, respectively. The ANOVA reported an intra-
and inter- observer standard deviation of less than 1.50 mm in all the cases.
Results obtained in the Bland-Altman study with the manual method showed
that both observers have a similar performance in independent sessions. The
repeatability study shows a bias less than 0.94 mm in all the cases. The standard
deviation is larger in the measurements of the aneurysm width and depth than
in the neck diameter.
The agreement study indicated a bias less than 0.70 mm in all cases. The
standard deviation is larger in the measurement of the aneurysm width and
depth than in the neck diameter as happened in the repeatability study. This is
logical as minimal variations in the selection of the view angle can yield large
variations in the saccular dimensions of the aneurysm when measured on the
2D projection images. These variations are less significative at the neck due to
its smaller size and symmetry.
When comparing manual and computerized measurements, it can be ob-
served that the bias is, in the worst case, approximately of the order of a voxel
(
−
0.69 mm). The standard deviations are lower than in the agreement study be-
tween observers in all cases. Therefore, the computerized method has a higher
agreement with each observer separately than the agreement achieved between
the observers themselves.
