@mastersthesis{schoner_extended_1999,
	address = {Berlin, Germany},
	type = {Diploma Thesis},
	title = {Extended Decorrelation Procedures for Source Separation of Biomedical Image Data},
	abstract = {The way the human brain works has fascinated philosophers and scientists since a long time. It
is hard to imitate tasks, which seem to be easy for the human brain, like visually recognizing
faces, with machines or computers. On the other hand, these have capabilities the brain is not
very powerful in, like doing calculations and storing information. It would open up many new
possibilities, if processes like face recognition, language understanding, abstraction and inference
in the human brain were comprehended.
One method applied to reach this comprehension is the analysis of how information is processed
and represented in different parts of the brain. Maps of the activity of neurons in the cortex, for
different stimuli or during certain actions are performed, are very useful for this method. The
optical imaging experiments, which are examined in this thesis, have as a goal the extraction of
such maps. Different signals indicating neural activity are recorded, together with unrelated signals
like blood vessels, biological and recording noise, by these experiments.
Conventional optical imaging mostly uses, among other methods to improve the signal to noise
ratio, bandpass filters to extract the activity maps. The use of bandpass filters is problematic,
because the resulting maps and the statistics of their features (e.g. number of singularities in orientation
preference maps) can be influenced by this.
A different approach recently used is the use of Blind Source Separation ({BSS})methods to separate
signal sources containing the mapping signal from those containing blood vessel artifacts, noise,
etc. This is achieved by learning a linear demixing matrix. When applied to the observed image
stack, the demixing matrix yields the estimated sources. Different methods exist for this learning;
one of them, used in this work, is the Extended Spatial Decorrelation ({ESD}) approach. For this
information about spatially shifted correlations of the mixtures is used.
There are different {BSS} techniques available, and [{SSM}+99] evaluated some of them on an image
stack obtained during an ocular dominance experiment. It is obvious that, although able to extract
activity maps, these algorithms have problems with sensor noise. One of the algorithms, which
yielded the best results there, the {ESD} algorithm, is improved and applied to two data sets in this
thesis.
The goal was to approximately decorrelate the estimated sources for several shifts instead of just
one, as the {ESD} algorithm does. This decreases the influence of sensor noise on the separation
results, and reduces the problem of selecting the right shift for decorrelation. For the optimization
of the extended error function an acccelerated gradient descent is used. Though this algorithm
is dependent on the initialization of its parameters, it is more flexible in learning the demixing
matrix, when compared to other multi-shift algorithms.
An artificial data set is used to control the sensor noise present in the analysed data. Issues analysed
using the artificial data set are a comparison between single- and multi-shift algorithms, the
differences in performance when using noise-robust sphering instead of the standard sphering approach
(sphering is a preprocessing step needed by some algorithms, helpful to others), and the
effects of spatially correlated sensor noise instead of white sensor noise. The results indicate a superior
noise robustness of the algorithm developed in this thesis, when compared to other variants
of the {ESD} algorithm. Evaluation for the second data set, the same ocular dominance experiment
as in [{SSM}+99], shows, that the newly developed algorithm compares favorably to the other {ESD}
variants and is very well able to extract ocular dominance maps. The extracted image maps have
better separation of the mapping signal from other sources like blood vessel artifacts or global
signal than other algorithms, which are currently used.},
	school = {Technical University of Berlin, Department of Computer Science, Neural Information Processing Group},
	author = {Schöner, Holger},
	month = may,
	year = {1999},
	keywords = {blind source separation, optical imaging},
	file = {download/publications/Schoener1999_DiplomaThesis.pdf}
}