Multielectrode-Arrays (MEA) allow the analysis of in-vitro neuronal
networks with a high spatiotemporal resolution [Jon11]. For analyzing the
signals recorded from a MEA and possible interneuronal connections, the
spike positions are to be identified. In this thesis, it is assumed that
the neuron-electrode interface and the electrode are characterized by a
Linear-Time Invariant (LTI) system. The original neuronal signal passes
through this system and is convoluted with an unknown transfer function.
This alters the signal and will make spike detection more difficult. Hence,
the original signal is to be reconstructed from the convoluted signal. The
transfer function of the LTI system is unknown. Therefore, this process is
called Blind Deconvolution. In this thesis, the Cepstrum of Bispectrum
(CoB) algorithmn for blind signal reconstruction introduced by Shahid and
others [SW08] is explained and implemented. This algorithm reconstructs the
original signal from the convoluted signal. The reconstructed signal is
used for spike detection afterwards. A simulated data set is developed.
Therefor a model for the neuron-electrode interface and the electrode
itself according to [Mar04] are implemented. For evaluation, this data set
as well as a second set of simulated data [Qui04] and a real neuronal
dataset are used. A Non-Linear Energy Operator (NEO) with and without CoB
preprocessing and a wavelet detector are used for spike detection. The
results of these three algorithmns are evaluated and compared. In case the
data has passed a LTI system before, the results using CoB algorithmn are
significantly better than results obtained without signal recontruction.
For simulated data which has not passed an LTI-System as well as for the
real dataset, this improvement is not observed. This leads to the
conclusion that the model does not apply to real data which has been used.
An examination of more real data sets and, if necessary, adjustment of the
model are prospective tasks. The wavelet algorithm has a higher
computational complexity than NEO algorithm without CoB preprocessing.
Nevertheless, the average results achieved with wavelet detection are not
significantly better than results produced by NEO. The three algorithmns
possess an error between true spike positions and spike positions
calculated by the three detectors. This error is dynamic and depends on the
length of a single spike in samples. When using this computed spike
positions for further analysis, this error must be taken into
consideration.