生物心理学笔记（Biological Psychology）05~08

Lecture 8/24

Lecture 5: EEG

Lecture 6: ERP

Lecture 7: MRI

Lecture 8: fMRI

Lecture 5: Electroencephalography (EEG) Methods

Electroencephalography (EEG) is a method of detecting neural activity by placing electrodes on the scalp. These electrodes pick up small fluctuations of electrical signals, originating from activity of mostly cortical neurons. The raw signals recorded are very noisy and not look like much, but cognitive process is related to it, so that EEG can help us understand better cognitive processes. EEG recording at the scalp is non-invasive, but it is also possible to record intra-cranial EEG by measuring activities directly at the exposed cortex. Normal scalp EEG is cheap and easy to conduct. Unimelb has 5~6 EEGs currently.

Hans Berger (1873-1941) detected the first EEG signal in 1924 with electrodes attached to the scalp of human and reported the results in 1929. Berger first described α wave, it appears when people closed their eyes, where the electrical signal varied with a characteristic frequency of 8-13 Hz.

Electrode gel / paste is applied to the gap between scalp and electrode to decrease the impedance. Typical systems have 32 / 64 / 128 / 256 channels. The systems that are used include: 1) international 10-20 system; 2) international 10-10 system; 3) modified 10-20 system for infant.

However, the raw-recorded EEG waves contains not only the brain wave informations, but also other electrical information, inclusively eye movements, muscle movements and air and other disruptions.

Ground electrode and reference electrode are also needed, other electrodes are used to measure eye movement and blinks. The reference should be a neutral point where no brain activity was there (e.g., nose, mastoids).

EEG signals are very small and have a typical amplitude of 10 to 100 \muV. So that the signals are amplified by a factor of 1,000 to 100,000. The typical sample frequency is between 256 to 1024 Hz (1000 Hz = 1 data point per millisecond). The spatial resolution for EEG is extremely low. The signal is band-pass filtered to remove low (< 0.5 to 1 Hz) and high (> 35 to 70 Hz) frequencies because they cannot reflect brain activity. The signal is also notch filtered (at 50 or 60 Hz, depending on the country) to remove line noise, which is also not brain activity. Different kinds of artefacts can contaminate the signal and need to be removed. Some of them can be detected automatically, and some need to be identified manually. Artefacts include eye blink, muscle and skin potential.

Eye movements and eye blinks can create very strong artefacts, much stronger than the brain signals we are interested in because the eye is a strong dipole. Since we record directly from electrodes next to and under the eye to capture horizontal and vertical eye movements, respectively, we can identify them easily. By using exclude contaminated trials, or using mathematical algorithms, such as independent component analysis (ICA), to remove just the eye-component.

High cut filter cuts off the 45 Hz signals, and low cut filter cuts off the 0.5 Hz signals. ICA can remove EOG, EMG and ECG components, so that the final graph would be the wanted.

Neurophysiologically, EEG activity originates mostly from post-synaptic potentials – voltages that arise when neurotransmitters bind to receptors on the membrane of the post-synaptic cell, and only to some degree from action potentials. With these electrical changes, the neuron acts as a small dipole, signals from single neurons are not strong enough to be recorded outside of the head, but if many neurons spatially align, then their summed potentials add up and create the signals we can record.

This pooled activity from groups of similarly oriented neurons mostly comes from large cortical pyramid cells. Over 10,000 neurons that activated simultaneously is needed to be recorded.

EEG is biased to signals generated in superficial layers of the cerebral cortex (gyrus and sulcus, gyrus is the outside layer of the brain, sulcus is inside the brain). So that EEG only can measure the activity of neurons at the gyrus, the sulcus part for us is still remained unknown. Signals in the sulk are harder to detect and additionally masked by the signals from the gyri. The Meninges, cerebrospinal fluid (CSF) and skull smear the EEG signal. We cannot easily locate the sources of the signal since it is a mathematical inverse problem. It is technically impossible, or very hard for localisation to occur. If the sources are known, then the resulting scalp configuration can be reconstructed. However, one given scalp configuration of the signals can have multiple dipole solutions.

When looking at frequency information (e.g., sleep research), the raw signal can show systematic variations (i.e., a specific frequency is dominant). Frequencies include: gamma: 30~100+ Hz (excited) beta: 12~30 Hz (relaxed) alpha: 8~12 Hz (drowsy) theta: 4~7 Hz (asleep) delta: 0~4 Hz (deep sleep)

If we look at the frequency related data, spectrogram can be used; if we look at the amplitude, then event-related potential method (ERP) will be significant.

Lecture 6: Electroencephalography (EEG) Methods: Event related potential (ERP) method

ERP relates on averaging amplitude data, instead of frequency data. The analysis of ERP is a method that allows us to investigate fast neural processes related to specific events of interest. We want to study what happens in the brain when participants engage in cognitive processes, such as perceiving, deciding, responding etc. ERPs can be obtained by time-locking the signals to the events we want to study, so we can analyse the signal amplitude at specific channels.

There are four key assumptions for this approach: 1) the event of interest is defined in time; 2) the event consistently evokes the signal; 3) the timing of the signal is consistent; 4) the signal and the noise are uncorrelated (signal is the assumed truth that is hidden, whereas the noise is the information that does not really coming from the brain). The signals and the noise should be assumed as uncorrelated and random, if the noise is correlated, then we cannot tell which is true and which is false. 5) the noise is random with a mean of zero.

We want to know whether there is brain activity reliably related to the cognitive processes of interest. However, usually the single-trial EEG trace is far too noisy to do that. There is a big problem of what are signals and what are the noises. We can align the trial segments from the event and average over the respective trials, thus all noise will average out so that a better estimate of the true neural response to the event of interest would be obtained (The weird thing is that: positive goes down the y axis and negative goes up the x axis, which is totally conversed). The peaks of the polynomials are signified as N_i and P_i. In order to get a useful estimate, many trials of the same type must be averaged. However, even the averaged signals per session for the same participant still look different. There is also a lot of variance between different participants who do the same experiment.

ERPs are described by their polarity and their order. Specific ERP components are measured at specific channels, or groups of channels. There are many well-studied ERP components, many of which do not have the typical labels (P_2 refers to the second positivity, N_400 refers to the negativity occurred at 400 ms; ERN refer to error related negativity; etc.).

One major problem with ERPs is reverse inference: When should we look at which component? What are we looking for? To solve this problem… reading literatures! But the fortunate thing is that, we have over 60 years of ERPs.

There are different options to derive a measure of the amplitude (i.e., the dependent variable for ERPs):

Peak amplitude

Peak to peak

Area under curve

Latency (the onset of the amplitude)

Most studies are interested in a baseline to peak measure; but then researchers have to decide what the best way to estimate this is. One could extract:

The maximum peak

The mean amplitude

The latency refers to the onset of the ERP component. There are mathematical algorithms that help with estimating the latencies. It can be very difficult to see when exactly a component starts. The tail end of the component is often neglected.

A good way of studying ERPs for studying cognitive processes is to subtract the waves from one condition from the waves from a control condition. Gehring and colleagues (1993) investigated whether there is a cognitive mechanism for the detection of and compensation for errors. For this, they measured error related negativity (ERN), a negative deflection of up to \mu V in amplitude observed at central electrodes 80 to 100 ms after an erroneous response. The ERN comes so fast after we have committed an error (but cannot take it back anymore) that is has been termed the brain’s “oh shit” response. Gehring et al asked their participants to emphasise accuracy or speed in a simple Flanker test in which participants had to respond to the central letter on the screen. There are congruent and incongruent series of letters (congruent: HHHHH, SSSSS; incongruent: HHSHH, SSHSS). For the incongruent condition, more errors are found and more errors appears, so that the ERN is recorded. Overall, they found a clear ERN on incorrect trial in comparison to correct trials. They found that the ERN was indeed strongest when people emphasised accuracy, and weakest for speed. This confirmed the hypothesis that participants’ brains only really cared about error detection when this was also important. However, this result does not show whether the ERN is also related to compensating for errors.

They also found that the greater the ERN, the lower the response force. Participants might be trying to correct for the error. The greater the ERN, the higher the probability to get it right on next trial, so that participants might be successful learning from errors. The greater the ERN, the slower the response on next trial; so that participants might be slowing down to avoid a subsequent error.

Lecture 7: Magnetic Resonance Imaging (MRI)

After TMS and EEG, we finally touch the real neuro-imaging technique of MRI and fMRI. For the first lecture we investigate the usage of MRI firsthand then we can focus on the fMRI in Lecture 8.

The reason for psychologists want to use fMRI, and other neuro-imaging techniques is to find out the neural basis for cognition. The assumption is that the cognitive process is based on neural activity, so that through the decoding of neural process, we can find out how people do cognitions. Since the brain region is activated, the specific brain area is considered to be involved in the specific cognitive process. fMRI is a most commonly used method. It works by placing the subject into a strong and static magnetic field generated by a large superconducting electromagnet, cooled by liquid helium. The field strength for fMRI is usually 3 T for experimental research. There are stronger magnets such as the 7 T scanner in our university since Monash has a 3 T scanner. 1.5 T for clinical purpose is also sufficient. The earth magnetic field is 65 microtesla. The structure of a fMRI is like this:

As shown, participants are placed into the scanner, and their head is covered by a coil (RF coil). There are also gradient coils, which are used to modify the magnetic field for short periods of time. Despite it is considerer as harmful in strong magnetic field, it is harmless being exposed to a strong magnetic field. Moving in the field can make some people nauseous. However, the danger of attracting magnetic objects, such as metal, can be fatal, since the field itself is invisible, it is easy to forget that it exists. The strength of magnetic field is not linear increase but exponential increase. The safety restrictions must be notified.

Participants can view experiments, which are controlled from outside the scanner room, via mirrors mounter on the head coil, or via goggles. People need to be stable while they are scanned. Responses can be given via scanner compatible keys, joysticks, or a touchpad. Participants’ head position is fixed to avoid any movement, which would distort the signal. The head coil is used to send radio frequency pulses and also functions as a receiver for the incoming signal.

Before introducing detailedly fMRI, MRI itself need to be introduced at the first place. MRI refers to magnetic resonance imaging, MRI is a structural imaging technique to neurons. MRI is the same with nuclear magnetic resonance (NMR). NMR refers to the atomic nucleus, which contains protons and neutrons. More than 70% of the human brain consists of water, which contains Hydrogen atoms (H^+ protons). These can be thought as small bar magnets precessing like a spinning top about an axis. The Hydrogen atoms often simply are referred to as protons or spins, because their precessing is often referred to as spinning. The spin directions are initially random, but in a string magnetic field, like MRI scanner, they align parallel or antiparallel to the magnetic field (the field is often referred to as B_0 field). In this field, most protons align parallel to the B_0 field. However, they are not perfectly aligned, and they are not static, they still keep precessing in a random fashion. The B_0 field is oriented into the direction of the z-axis in the scanners coordinate system, where the z-axis is parallel to the human body direction lying down in the MRI machine. The precession frequency of the protons is also referred to as Larmor frequency, which is depended on the strength of the magnetic field. So that Larmor frequency = precession frequency = the strength of the B_0 magnetic field. So that an imagination for which protons are being aligned with the B_0 field, but they are in different positions in their precession, can be created. The phase in which they spin / precess is different.

Till now, as long as they are aligned in the direction of the B_0 field, we cannot measure a signal with our head coil, which surrounds the head. The second problem is that the signal from each proton itself is tiny, and they are not precessing in phase (they are not in the same position at the same time). In order to get a signal, we applied a radio frequency pulse perpendicular to magnetic field B_0 using the head coil (considering the z axis as the direction parallel to subject’s body, and that the RF pulse is parallel with the xOy plane).

The first effect of the RF pulse is that all protons will start precessing (spinning) in phase, meaning that their magnetisation will all point to same location in space at the same time. This happens because the protons absorb energy from the RF pulse, which also heats up the tissue a bit. The second effect of the RF pulse is that the magnetisation vector is tilted away from the z axis. This means that the magnetisation is tilted from the longitudinal direction of the field B-0 into the transversal plane (the xOy plane).

This can be stated as following: initially the H^+ protons were aligned parallel to the z axis, as the influence of RF pulse, they become parallel with the xOy plane. The magnetisation vector now rotates around in the transversal plane where our head coil is placed, the head coil will now receive this as a signal. However, this signal is still rather meaningless for us, since it comes from the entire brain. The trick is now to now switch off the RF pulse after which the transversal magnetisation decays very quickly because the protons emit the excess energy. They also lose phase coherence very quickly, which makes the signals disappear.

The first consequence of switching off the RF pulse is that protons align again with the magnetic field, also referred to as longitudinal relaxation, spin-lattice relaxation, or T1 recovery. T1 refers to a time constant of a function, indicating how long the recovery of the longitudinal magnetisation takes. Importantly, this time constant is different for different tissue types. Fat decays very fast, and cerebrospinal fluid (CSF) decays very slowly. Their constants are different. The second consequence of switching off the RF pulse is that the transversal magnetisation decays (it is independent), also referred to as transversal relaxation, spin-spin relaxation (interaction of spins the protons with other protons), or T2 decay (dance and party metaphor). T2 is the time constant indicating how long the transversal decay takes. T2 decay is much faster than T1 recovery, and again different for different tissues. Fat is again, has a short T2, and CSF also has a long T2.

When the signal is measured during this phase of relaxation, different signals will be emitted from protons in different tissues. Depending on when signals are measured, researchers can use T1 and T2 to get differently weighted images of the brain and clearly see the type of tissue. For this, different types of sequence are used that are optimised to capture differences in signal, due to T1 recovery or T2 decay. Structural brain images depend on when signal is recorded during this process.

The last thing we need to do is the reconstruction of brain image. In order to get separate measurements from different locations in the brain, we first need to reconstruct where exactly the signal comes from. For this, we use gradients created by the gradient coils. Protons will absorb energy from RF pulse only when the frequency of the RF pulse matches the protons’ precession (resonance) frequency. Thus by causing the magnetic field to vary linearly,, we can cause the resonance frequency to vary throughout the brain. An RF pulse of a specific frequency will now only excite one slice of the brain, precisely the slice where the resonance frequency of the protons matches the frequency of the RF pulse. The first gradient we use is called slice selecting gradient. It varies the gradient field along the z axis such that different slices are exposed to different field strengths. An RF pulse can now be chosen to match precisely the precession frequency of protons inn one slice of the brain. This gradient is applied during the RF pulse. The second gradient is called the phase encoding gradient. It changes the precession frequency of the excited protons depending on their location in the gradient, causing de-phasing. When removed, the resonance frequencies are the same again, but the differences in phase persist, so that their phase is now informative about their position. This gradient is applied after RF pulse. The third gradient is called the frequency encoding gradient and it changes the magnetic field within the selected slice. This happens during readout of the signal. Because all protons at a certain position in the gradient now have same resonance / precession frequency, they frequency at readout is informative about their position. In classical sequences, one RF pulse is used for each slice, meaning that the time it takes to measure the entire brain depends on the number of slices (TR, time to repeat), usually 1 to 3 seconds.

Other parameters includes: 1) slice thickness and gaps between them; 2) voxel, pixel, 3) field of view, how many voxels we measure per slice; 4) number of slices; 5) orientation of the slices and read out direction.

Lecture 8: functional Magnetic Resonance Imaging (fMRI)

We have talked about MRI, but that is not all psychologists need, what we need is the function of brain. So we will cover today functional Magnetic Resonance Imaging (fMRI). Subtopics include: 1) hemodynamic imaging; 2) blood oxygen and glucose metabolism (BOLD signal); 3) hemodynamic response function (HRF); 4) when & how to measure HRF; 5) statistcal parametric mapping.

Psychologists wants to understand cognitive process not brain per se, fMRI can do this.

fMRI does not measure neural activity directly but is a hemodynamic neuro-imaging method. In 1990s, Seiji Ogawa discovered that large blood vessels cause tighter areas in MRI scans. At first they could not explain it, but it gives better signals. Ogawa’s group investigated this phenomenon in detail. Ogawa changed the blood oxygen level experimentally and found that this ended impacts the signal.

Local neural activity requires energy, which is generated in the brain in the form of adenosine triphosphate (ATP). To produce ATP, glucose is metabolised, for which oxygen is required. Oxygen is constantly transported through the brain arteries and a network of arterioles (small arteries). In the capillaries, oxygen molecules are removed from haemoglobin (Hb), turning oxyhemoglobin into de-oxyhemoglobin.

While the brain is at rest, the blood flow is normal, with normal quantity of oxyhemoglobin and normal quantity of deoxyhemoglobin (paramagnetic). After stimulus, the neural activity increases and thus oxygen and glucose consumption increases immediately, so that initially in the blood vessel, the quantity of deoxyhemoglobin increases, then the cerebral blood flow increases and so that the oxygen is carried to the brain, the oxyhemoglobin increases, so that fMRI BOLD response increases.

The increase in blood oxygenation is much larger than the initial dip, meaning that shortly after the neural activity, there is an oversupply of oxygen in the blood. The increase in blood oxygenation causes our signal too get better – this is the Blood Oxygen Level Dependent (BOLD) signal we measure. The change in signal is described by the Hemodynamic Response function (HRF), which is similar (not identical) in different brain regions (you can never say ‘Region A is more activated than Region B at the same condition ’). The peak of the HRF is reached 4 to 8 secs after the neural activity occurred. It takes the signal up to 16 secs to go back to baseline levels. If several neural events take place, their HRFs will add up linearly.

What is reflected by the BOLD signal? The relationship between the BOLD signal and “neural activity” is complicated, but that should not be a reason to refrain from using it. The BOLD signal correlates best with local field potentials (LFPs), which reflect the neurons’ input at the synapses, but recent studies have shown that it is also correlates with action potentials (i.e., neural firing, the output of neurons). In sum, the BOLD signal might be a mixture of activity within local cortical excitation-inhibition networks (EIN), small and highly interconnected functional micro units, which show massive recurrent feedback.

From measuring the BOLD signal to getting results. Due to the presence of deoxygenated haemoglobin, which is paramagnetic (it has a magnetic momentum), protons experience an additional speeding up of the decay of transferal magnetisation. This means, the total decay is faster than predicted by the T2 constant. The total decay (dephasing) is referred to as T2* decay. BOLD fMRI mostly uses T2* weighted pulse sequences, which give a better signal when blood is more oxygenated. In a typical fMRI experiment, the HRF can only be measured sparsely. The HRF is must therefore be estimated, and timing relative to the event of interest matters. For the analysis, we use statistical parametric mapping, using a general linear modelling (GLM) approach, we search for regions in which the signal increases fits to the predictions of our model.

The program is given the mapping between conditions and recorded brain images and then estimate the model fit. Irrelevant variables, which might impact the estimate of the model, can also be factored in, but are not analysed. This is done first for each voxel in each participant’s data. Group statistical analyses are then performed on the sample (t tests).

The resulting statistical maps (indicating how well the data fit the model) is then overlaid on a structural image of the brain. Significant effects in a brain region for task A compared to a control task B is interpreted as involvement (activation) of this region. We cannot compare activation between regions because the HRF is different. Activation denoted as red, and deactivation denoted as blue.

We need to repeat the measurement of brain activity many times while participants perform the experimental task because the signal is very noisy, and the final brain maps are always based on averaging across many trials.