Mining Statistical Association Rules to Select the Most Relevant Medical Image Features - Mining Complex Data - page 119

1

Full(30)
StarMiner(21)
StarMiner(20), feature D removed
StarMiner(20), removing
a feature randomly chosen

0.95

0.9

0.85

0.8

0.75

Recall (%)

0.7

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

Fig. 7.6. Precision vs. Recall obtained using 30 features, 21 features selected by
StARMiner, 20 features obtained by removing the feature D from those selected by
StARMiner, and 20 features obtained by randomly removing one feature from those
selected by StARMiner.

queries using 20 features obtained by randomly removing one feature from the
21 features selected by StARMiner, and, as a result of it, the P&R graph always
worsened. Figure 7.6 shows a P&R graph comparing the precision of the queries
using 30 features, 21 features selected by StARMiner, and 20 features obtained
by randomly removing one feature from those selected by StARMiner. The se-
lected features shown in Table 7.3 points out that the feature D has a very low
contribution to represent the images. We redid the similarity queries removing
the feature D of region 1 from the set of 21 selected features to check if it drops
the precision values. The results of this test are also shown in Figure 7.6.
Comparing the curves from Figure 7.6, it is possible to note that removing the
feature D from the 21 features selected by StARMiner, the precision values reduce
much less than removing another feature, since the contribution of the feature D is
small. For Region 1, the feature D is meaningful, but among the selected features,
it is the one that brings the smallest contribution to differentiate the image types.
It is interesting to visualize the images returned when applying similar-
ity queries using the whole set of features and the reduced set selected by
StARMiner. We asked for the 15-nearest neighbors over the same image cen-
ter shown on the left side of the screen of Figures 7.7 and 7.8.
Figure 7.7 shows the results when using the 30 original features, while
Figure 7.7 shows the results using only the 21 selected features. The features