Information Technology Reference
In-Depth Information
be developed to solve such problems. Such techniques might be available in the
Internet, spreading information and knowledge.
It is expected that the progress in the field of mining complex data makes com-
putational devices and programs to go a step further in the task of reproducing
human brain functions. In fact, a human brain analyzes millions of complex data
in a fraction of seconds, bringing to conscience just a small fraction of such data.
This data fraction is the information that is stored in memory. Thus, the human
brain is the perfect machine of mining complex data and, in future, scientists
might be reproducing such machine. The selection of the most representative
features extracted from images, based on association rules, is one of the most
promising approaches that can turn these developments into reality.
7.6
Conclusions
This chapter details two issues related to mining complex data: feature selection
and association rule mining. A new approach to select the most relevant image
features in image datasets has been presented, consequently allowing dimension-
ality reduction of medical image features. The presented method uses statistical
association rules to select the most relevant features. The presented mining al-
gorithm, StARMiner, finds rules involving the attributes that most contribute
to differentiate many classes of medical images. The accuracy of the method
was verified in several case studies, and one representative was discussed in this
chapter. The experiments performed
k
-nearest neighbor queries to measure the
ability of the proposed technique in reducing the number of features needed to
perform similarity queries maintaining the accuracy of the results. The results
show that a significant reduction in the number of features can be obtained im-
proving the retrieval ecacy of the features, leading to an impressive gain in
time. The experiments also indicated that the features selected by StARMiner
are more relevant to discriminate images than those selected by Relief-F and
DTM algorithms. Furthermore, the results indicate that the mining of statis-
tical association rules to select relevant features is an effective approach for
dimensionality reduction in medical image datasets. Future trends in the field
of image mining were discussed and an optimistic description of possible future
scenery of the field was presented.
Acknowledgments
This research was supported by the Sao Paulo State Research Foundation
(FAPESP) and by the Brazilian National Research Council (CNPq).
References
1. Agrawal, R., Imielinski, T., Swami, A.N.: Mining association rules between sets of
items in large databases. In: ACM SIGMOD Intl. Conf. on Management of Data,
Washington, D.C, pp. 207-216 (1993)
