A dynamic causal modeling study of attention shifting between smooth

pursuit and saccadic targets

Ferenc Acs & Mark W. Greenlee, University of Regensburg

Overview:
The purpose of this project is to explore the possibilities of closing the gap between
fMRI data measurement and the construction of artificial neural networks, especially
mean field networks (Brunel, 2003). These models simulate the behaviour of massive
biological cell assemblies. The system dynamics of interconnected massive biological
cell assemblies in human brains became measurable with functional Magnetic
Resonance Imaging and the data analysis method of Dynamic Causal Modelling
(Friston, 2003 & Fig. 4). DCM requires assumptions about the connectivity between
brain regions, for this purpose, Krauzlis’ (2005) model of attentional eye movement
processing was applied for the human brain (Fig. 1).
Here are the results of a DCM group study, exploring the dynamics of certain brain Figure 1: The connectivity between brain regions relevant to
attentional eye movement processing. This is the schematic illustration
regions during different attention conditions to moving stimuli. of a macaque monkey brain, because effective connectivity data for
humans are still rare. (From Krauzlis, 2005)
Experimental Design:
The experiment was conducted with a 3T Siemens Allegra MRT scanner (TR=2s, 32 I) Visual Stimuli
slices, 3x3x3 mm Voxel Size). A total of 1156 functional scans were acquired, total
running time was 38 minutes. The subjects saw two moving dots and a fixation cross
in his visual field. The upper dot made smooth sinusoidal movements, the lower dot
was the target for triple step saccades in both directions. (Fig. 2,I). A simple block
design was used (Tab. 1), the Blocks differed only in the instruction for the subjects. II) Reaction Time Task
Either the subjects had to follow one of the dots (A,B) or the subjects had to fixate the
yellow cross and shift his attention towards the upper (C) or the lower (D) dot. The rest
condition (R) was used to acquire a baseline, only the fixation cross was visible and the
subjects had to fixate it.
Occasionally, with a probability of 0.05% per frame, both dots changed their color, this
was an reaction time task to test the concentration of the subjects (Fig. 2,II). Figure 2: I) The visual stimuli used in the experiment. Sinusoidal
moving dots with an amplitude of 9° in the visual field and a
frequency of 0.77 Hz. II) Shows the change of the stimuli for the
Table 1: The fMRI Block Design. The duration of reaction time task. This stimulus appeared with a probability of
Pursuit Saccades Attention Attention Relax each block was 30s, 15 volume scans were 0.05% per frame (70Hz) and lasted 200ms. The subjects had to
Pursuit Saccades acquired per block. Each block was preceded by press a button with the index finger upon detection
a display of the visual instructions for 6 seconds.
A B C D R The order of the blocks was permutated by the
method of the latin square.

Pursuit --
Results:
0.45
Saccades
Att. Purs.
0.20
-- Comparing six DCM models after parameter estimation, the model shown on the left
Att. Sacc. 0.15
Pursuit
Saccades
-0.91
0.18
Relax -- provided most information (Fig. 3). The model comparison used pairwise testing of the
--
Att. Purs.
Att. Sacc.
--
0.15 FEF
models, the measures were the Bayesian Information Criterion and Akaikes
Photic
Relax --
Pursuit -- Information Criterion (Penny, 2004). Nearly every connection specified in the model
Saccades -- Pursuit --
Stimulation 0.18 Att. Purs.
Att. Sacc.
--
--
-- Saccades
Att. Purs.
--
--
passed the significance threshold (>90%, Bayesian posterior probablility). The general
-0.98 0.06 0.78 0.73
Relax -- Att. Sacc.
Relax
--
-- excitatory effect of V1 to MT/MST is weakened by the conditions B,C & D and
MT
V1 MST LIP Pursuit
Saccades
--
--
Pursuit
Saccades
--
--
enhanced by conditions A & R. The MT/MST - LIP interaction is enhanced by pursuit
Pursuit -0.49 Pursuit 0.16 Pursuit 0.41
Att. Purs.
Att. Sacc.
--
--
0.27
Att. Purs.
Att. Sacc.
--
--
stimuli (A). Attentional shifting effects (C,D) could be reproduced for the V1 -> MT/MST
Saccades -- Saccades -0.23 Saccades -- Relax --
Att. Purs.
Att. Sacc.
--
--
Att. Purs.
Att. Sacc.
-0.14
-0.23
Att. Purs.
Att. Sacc.
--
--
Relax --
-> FEF pathway. However a modified model (Fig. 3, orange line) leaving LIP away and
0.15 0.29
Relax -- Relax 0.32 Relax --
SEF modelling a direct connection between MT/MST and SEF showed clear inhibitory
Pursuit
Saccades
-0.04
--
Pursuit 0.20 Pursuit -- Pursuit 0.22 effects for the attention to pursuit condition (C). Conditions A & R enhanced the
Saccades -- Saccades 0.18 Saccades --
Att. Purs.
Att. Sacc.
-0.25
--
Att. Purs.
Att. Sacc.
--
--
Att. Purs.
Att. Sacc.
--
--
Att. Purs.
Att. Sacc.
-0.47
--
connectivity.
Relax -0.08 Relax -- Relax -- Relax 0.31

0.38 0.41
Discussion:
The effects of the attention conditions (C,D) to the V1->MT/MST->LIP connections
could be shown. However the assumed MT/MST <-> SEF connection seems to be
Figure 3: The resulting averaged DCM model. Five regions of interest have two-
way connections, indicated by the black arrows connecting the circles. The relevant for paying attention to pursuit stimuli (C). Saccadic stimuli seem to affect the
numbers represent the connection strength between regions, the unit of connections MT/MST -> FEF and the connections LIP -> FEF & LIP -> SEF, the
measurement is Hz. The larger the number, the faster a modulational effect activation effect might be caused by the programming of eye movement vectors.
between the regions will propagate, indicating a stronger connection. The large Finally attentional shifting effects could be demonstrated. Effects for attention to pursuit
numbers are the intrinsic connections strengths. They are modulated by the
different conditions affecting the internal processing state of the system (boxes stimuli (MT/MST -> SEF), evident in a four region model may be due to the fact that
with green arrows). However the box ‘Photic Stimulation’ describes the external one important area for attentional modulation is be missing.
stimuli entering the visual system through retinal activation, these external stimuli
are pertubing the system and are necessary to determine the modulational
effects of internal processing states. (Subjects: SA,SB,SC,SD)
The orange line, connecting MT/MST and SEF directly, lacks of physiological
evidence. But the attentional shifting effect visible here, suggest that either there
is a direct connection or one important area in the pathway between these
regions is missing. (Subjects: SC,SD)
Table 2: The MNI coordinates of the regions of interest analysis for each subject (SA, SB, SC & SD)

V1: data and model predictions

2
hemodynamic responses neuronal responses to Pursuit
1
0.2 0.04
FEF SEF MTMST
0.02
0
-1
0 FEFSEF
V1 LIP -2
0 V1 LIPFEFSEF
MTMST
1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000

-0.2 -0.02 MTMST: data and model predictions

0 10 0 10
1st eigenvariate: MT/MST time {seconds} time {seconds} 2

hemodynamic responses neuronal responses to Saccades

2 0
0.5 0.05
LIP MTMST
-2
1 0 V1 LIP
SEF
FEF 0 V1 LIP
MTMST
FEF
SEF

Fixation
1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000
Tracking -0.5 -0.05 LIP: data and model predictions

15
0
Pursuit Pursuit 0 10
time {seconds}
0 10
time {seconds}
2

MT/MST
Coordinates: -1 Saccadic Saccadic hemodynamic responses
0.5
MTMST
neuronal responses to Relax
0.1
0

10 x: -50 MT
-2
-2 0 SEF
FEF
LIP 0 V1 FEF
LIP
MTMST SEF 1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000
y: -72 V1
MST FEF: data and model predictions
5 z: 4 500 1000 1500 2000 -0.5 -0.1
0 10 0 10 2
(MNI Space) time (seconds}
time {seconds} time {seconds}
0
0 80 voxels in VOI at [-50 -72 4] hemodynamic responses neuronal responses to Att_Stim
0.5 0.1
t-Values Variance: 72.73%
MTMST
-2
1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000
0 LIPFEFSEF 0 V1 LIPFEFSEF
V1 V1
MTMST

2
SEF: data and model predictions

-0.5 -0.1
0 10 0 10 0
time {seconds} time {seconds}
-2
1000 1100 1200 1300 1400 1500 1600 1700 1800 1900 2000
time {seconds}

a) b) c) d) e)

Figure 4: Schematical steps for a DCM fMRI single subject analysis:

a) After a conventional GLM analysis a region of interest has to be marked, in this picture it is the MT/MST area in a human brain. Literature:

b) The volume time course if this region is extracted. It consists of the 1’st eigenvariate of all BOLD signals of the Voxels within this Brunel, N. (2003), ‘Dynamics and Plasticity of Stimulus-selective Persistent Activity in
Cortical Network Models’, Cerebral Cortex, 13(11), 1151-1161
region. Steps a) and b) have to be repeated for every Region of Interest.
Friston, K.J., Harrison, L. & Penny, W. (2003), 'Dynamic causal modelling', NeuroImage,
c) A DCM model has to be specified. The connectivity in the model should be evident from theoretical assumptions and knowledge 19 (4), 1273- 1302

about the effective connectivity between these Regions. Krauzlis, R.J. (2005), ‘The control of voluntary eye movements: New perspectives.’ The
Neuroscientist, 11: 124-137
d) The parameter estimation, estimates the neural activity for each region from the BOLD response. During several iterative
Penny, W.D., Stephan, K.E., Mechelli, A. & Friston, K.J. (2004), 'Comparing dynamic
computing steps the model is fitted on the data until: causal models', NeuroImage, 22, 1157- 1172

e) The model data (red) approximate the measured BOLD time course (blue) according to a gradient descent algorithm using a
Laplace approximation.