Dynamic Cerebral Contrast-enhanced Perfusion MRI - Biomedical Signal Analysis: Contemporary Methods and Applications - page 322

ful hypothesis to interpret the time series of at least some clusters in the
light of physiological meta knowledge, although a definite proof of such
an interpretation will be missing. Hence, such an approach is certainly
biased by subjective interpretation on the part of the human expert per-
forming this interpretation of the resulting clusters, and thus, may be
subject to error. In summary, it is not claimed that a specific cluster
is well-correlated with physiological phenomena related to changes of
brain perfusion, although one cannot exclude that a subjective inter-
pretation of some of these clusters by human experts may be useful to
generate hypotheses on underlying physiological processes in the sense
of exploratory data analysis. These remarks are in full agreement with
the whole body of literature dealing with unsupervised learning in MRI
time series analysis, such as [84] and [53].
The normalization of signal time-curves represents an important is-
sue where the concrete choice depends on the observer's focus of interest.
If cluster analysis is to be performed with respect to signal dynam-
ics rather than amplitude, clustering should be preceded by time series
normalization. While normalization may lead to noise amplification in
low-amplitude CTCs, in cluster analysis of signal time series, preceding
normalization is an option. However, CTC amplitude unveils important
clinical and physiological information, and therefore it forms the basis
of the reasoning for not normalizing the signal time-curves before they
undergo clustering.
In order to provide a possible explanation of the relatively high MTT
values obtained in the results, the following should be mentioned. The ra-
tionale for using equation (11.3) for computing MTT is that the arterial
input function, which is dicult to obtain in routine clinical diagnosis,
was not determined. The limitations of such an MTT computation have
been addressed in detail in the theoretical literature on this topic (e.g.,
[299]). In particular, it has been pointed out that the signal intensity
changes measured with dynamic MR imaging are related to the amount
of contrast material remaining in the tissue, not to the eux concentra-
tion of contrast material. Therefore, if a deconvolution approach using
the experimentally acquired arterial input function (e.g., according to
[149, 281]), is not performed, equation (11.3) can be used only as an
approximation for MTT. However, this approximation has been widely
used in the literature on both myocardial and cerebral MRI perfusion
studies (e.g., [106, 219, 283]).