Information Technology Reference
In-Depth Information
Fig. 19 Average segmentation accuracy of the VB segmentation on 12 CT images (ESP) with
respect to the various noise levels
The results show that the proposed method is robust under various noise levels as
well as faster than other famous alternatives.
The initialization effect is also validated in this experiments. It should be noted that
A1
A3 need perfect manual initializations. However, the method is almost inde-
pendent of the initialization (which is usually required in the registration step). The
segmentation results and the accuracy on the ESP (when the initialization is not
optimal) are shown in Figs.
20
and
21
, respectively. In this
-
figures, the initial point is
chosen not close to the center point of the object of interest. It
s clear that the proposed
method performance is almost constant with different initial points. On the contrary,
the alternative methods are severely suffering from performance degradation.
The effect of each model is validated as shown in Fig.
22
. In the
'
figure, (i) shows
the initialization place for each method. The results which are based on (ii) only the
intensity model, (iii) intensity and spatial interaction, (iv) intensity, spatial inter-
action, and shape models are shown. The intensity based approach is not robust
under various noise levels. After the spatial interaction model is used, the seg-
mentation is getting better and most of the noise is eliminated. However, there are
still missing information and some noise using the two models. With the proposed
approach much better results are obtained compared with other models. The seg-
mentation accuracy with respect to the various noise levels is shown in Fig.
23
.
2.4.10 Experimental Results-Results on Clinical CT Images
In this study, different type of data sets are used. Classi
cation of data sets are
categorized into 3 groups as shown in Table
1
. Classi
cation is based on 6 features.
