Cluster analysis of the signal curves in perfusion DCE-MRI datasets

Pathological studies show that tumors consist of different
sub-regions with more homogeneous vascular properties during their
growth. In addition, destroying tumor's blood supply is the target
of most cancer therapies. Finding the sub-regions in the tissue of
interest with similar perfusion patterns provides us with valuable
information about tissue structure and angiogenesis. This
information on cancer therapy, for example, can be used in
monitoring the response of the cancer treatment to the drug.
Cluster analysis of perfusion curves assays to find sub-regions
with a similar perfusion pattern. The present work focuses on the
cluster analysis of perfusion curves, measured by dynamic contrast
enhanced magnetic resonance imaging (DCE-MRI). The study, besides
searching for the proper clustering method, follows two other major
topics, the choice of an appropriate similarity measure, and
determining the number of clusters. These three subjects are
connected to each other in such a way that success in one direction
will help solving the other problems. This work introduces a new
similarity measure, parallelism measure (PM), for comparing the
parallelism in the washout phase of the signal curves. Most of the
previous works used the Euclidean distance as the measure of
dissimilarity. However, the Euclidean distance does not take the
patterns of the signal curves into account and therefore for
comparing the signal curves is not sufficient. To combine the
advantages of both measures a two-steps clustering is developed.
The two-steps clustering uses two different similarity measures,
the introduced PM measure and Euclidean distance in two consecutive
steps. The results of two-steps clustering are compared with the
results of other clustering methods. The two-steps clustering
besides good performance has some other advantages. The granularity
and the number of clusters are controlled by thresholds defined by
considering the noise in signal curves. The method is easy to
implement and is robust against noise. The focus of the work is
mainly the cluster analysis of breast tumors in DCE-MRI datasets.
The possibility to adopt the method for liver datasets is studied
as well.