Transcript Region growing - Facultad de Ciencias

Segmentación de mapas de
amplitud y sincronía para el
estudio de tareas cognitivas
Alfonso Alba1, José Luis Marroquín2, Edgar Arce1
Facultad de Ciencias, UASLP
2 Centro de Investigación en Matemáticas
1
Introduction
Electroencephalography (EEG) consists
of voltage measurements recorded by
electrodes placed on the scalp surface
or within the cortex.
Electrode cap
• During cognitive tasks, several
areas of the brain are activated
simultaneously and may even
interact together.
Varela et al., 2001
EEG synchrony data
Synchrony is measured at specific frequency
bands for a given pair of electrode signals.
Typical procedure:



Band-pass filter electrode signals Ve1(t) and
Ve2(t) around frequency f.
Compute a correlation/synchrony measure mf,t,e1,e2
between the filtered signals
Test the synchrony measure for statistical
significance
In particular, we obtain a class field cf,t,e1,e2
which indicates if synchrony was significantly
higher (c=1), lower (c=-1) or equal (c=0) than
the average during a neutral condition.
Visualization
(Figure categorization experiment)
The field cf,t,e1,e2 can be partially visualized in various ways:
Multitoposcopic display of the
synchronization pattern (SP) at a
given time and frequency
Time-frequency (TF) map for a
given electrode pair (T4-O2)
Time-frequency-topography (TFT)
histogram of synchrony increases at
each electrode
• The TFT histogram shows regions with homogeneous
synchronization patterns. These may be related to
specific neural processes.
Seeded region growing
TF regions with homogeneous SP’s can be
segmented using a simple region growing algorithm,
which basically:
1.
2.
3.
Computes a representative synchrony pattern (RSP) for
each region (initially the SP corresponding to the seed).
Takes a pixel from some region’s border and compares its
neighbors against the region’s RSP. If they are similar
enough, the neighbors are included in the region and the
RSP is recomputed.
Repeats the process until neither region can be expanded
any further.
Region growing
(Figures experiment)
Automatic seed selection
An unlabeled pixel is a good candidate for a
seed if it is similar to its neighbors, and all of
its neighbors are also unlabeled.
To obtain an automatic segmentation, choose
the seed which best fits the criteria above,
grow the corresponding region, and repeat
the procedure.
Bayesian regularization
The regions obtained by region-growing show
very rough edges and require regularization.
We apply Bayesian regularization by
minimizing the following energy function:
lt,f is the label field
Lt,f is a pseudo-likelihood function
Ns is the number of electrode pairs
V is the Ising potential function
lt and lf are regularization parameters
Results
(Figure categorization experiment)
Automatic segmentation
Regularized segmentation
Results
(Figure categorization experiment)
Results with induced amplitude
Region optimization
Merge regions with similar RSP’s

Two regions i and j are merged if
d ( RSPi , RSPj )
HC i HC j
 em
Delete small regions

After merging, regions whose area is
smaller than some ed are deleted.
Region optimization example
Region optimization example
Conclusions
We have developed a visualization system for
EEG dynamics which



Produces detailed representations of synchrony and
amplitude patterns that may be relevant to the
task.
Helps neurophysiologists determine TF regions of
possible interest.
Can be fully automated and allows for human
interaction.
Future work
Validation
Use of segmented maps for the study
of a psychophysiological experiment.
Segmentation using combined
amplitude+synchrony data?
Homer says thank you!