What are we measuring in fMRI? Ruth Stavy David Carmel

1. What are we measuring in fMRI?Ruth StavyDavid Carmel

2. Outline • Neuronal activity • Physiology and Haemodynamics • Physics of fMRI • Use of physics to “see” neural activity • What fMRI really tells us (?) about neural activity

3. Physiology and haemodynamics functional Magnetic ResonanceImaging (fMRI) measures brain activity indirectly through changes in blood vasculature that accompany neural activity.

4. Neuronal activity results in: • An initial increase in oxygen consumption owing to increased metabolic demand. 2. After a delay of ~2 secs, a large increase in local blood flow. which overcompensates for the amount of oxygen being extracted 3. Local increase in cerebral blood volume The increase in blood oxyhaemoglobin is what we measure in fMRI This is called the BOLD (Blood Oxygen Level Dependent) response

5. MRI Images: What do we see? • MRI images are usually based on the signal from protons • A Proton is the nucleus of the hydrogen atom • Hydrogen is the most common element in tissue • The signal from protons is due to their „spin“

6. Atomic nuclei with an odd number of neutrons and/or protons have: - a small magnetic moment - an angular momentum called nuclear spin Spinning protons are little magnets which are frequently referred to as just spins.

7. If an external magnetic field B0 is applied, they align with it. An MRI machine’s B0 is alwayskept on! M A slight excess number of protons align with B0 producing an additional magnetic field in the direction of B0 - we will refer to this field as M.

8. In addition, when placed in the magnetic field B0 the hydrogen proton can absorb energy and turn (precess) around the applied magnetic fieldmuch like a spinning top when placed in the earth’s gravitational field: Each type of nucleus has a unique frequency of spin. This is also proportional to the strength of the magnetic field, and is known as its resonance.

9. The resonance frequencies characteristic of today’s MRI machines are in the range of FM radio waves, which is why they are referred to as RF (Radio Frequency) Excitation: If a second magnetic field (another RF) is applied, orthogonal to the static field and at the precessing frequency of the atomic nuclei, the nuclei will all precess about the static field together, in-phase. A nearby coil can detect this precession: the precession causes a current through the coil, which has this characteristic frequency.

10. 2. M begins to return to its original orientation (B0) - At an Exponential rate defined by a certain formula - This formula contains a “Time constant” - T1 - Signal intensity decreases with increasing T1 - T1 is unique to every tissue. This is the basis of structural MRI. - T1 is quite long: tens of milisecs to secs. Decay (relaxation):What Happens when the additional magnetic field is turned off? 1. Protons emit the energy they absorbed (they do so at their RF) - This is the "NMR“ (Nuclear Magnetic Resonance) signal detected by the nearby coil- At the radio frequency- Signal proportional to Proton Density

11. T1 Weighted Image Caudate Nucleus (T1=1200ms) CSF (T1=4300ms) T1-Weighted Image Corpus Callosum (T1=700ms)

12. 3. Spins begin to dephase - At an Exponential rate - defined by a certain formula - This formula contains a different “Time constant” – T2 - Signal intensity decreases with decreasing T2 - The T2 rates are different for different tissues - The observed signal decay T2 is fast: a few msec to tens of msec

13. T2 Weighted Image Caudate Nucleus (T2=100ms) CSF (T2=2000ms) T2 Weighted Image Corpus Callosum (T2=90ms)

14. To sum up, Relaxation: • In the real world, the NMR signal decays faster than T2 would predict. • There are many factors creating imperfections in the homogeneity of a magnetic field. • Every tissue has a different magnetic susceptibility which distorts the field at tissue borders, particularly at air/tissue interfaces. • The sum total of all of these random and fixed effects is called T2*

15. BOLD (Blood Oxygen Level Dependent)contrast Neuronal activity local blood flow increases overcompensating for oxygen consumption oxygen level in venous blood is elevated larger MR signal. • Takes advantage of the different magnetic properties of oxyhemoglobin (HbO) and deoxyhemoglobin (Hb) Hb is paramagnetic and introduces an inhomogeneity into the nearby magnetic field, whereas HbO is weakly diamagnetic and has little effect.

16. How do we know where the signal comes from? f = B0  is proportionality constant for specific atomic nuclei B0 is the strength of the static magnetic (e.g.,1.5T) The frequency f emitted by the protons is proportional to the magnetic field B0

17. B B1 B4… B2 B3 Different B values  Different f values Values of frequencies indicate Position. This is done in all three axes to enable 3-dimensional localisation (blobs) If we modify the magnetic field B0 by applying a „gradient field“ (B rises from left to right):

18. altered neuronal activity changes in local haemodynamics What exactly does fMRI tell us? • From sluggish Haemodynamic response to inferences on neural activity • The central assumption: the fMRI signal is approximately proportional to some measure of the local neural activity, averaged over several millimeters and several seconds. This is sometimes referred to as the linear transform model (Buckner, 2002) fMRI signal How the three are related is unclear

19. (Heeger & Ress, 2002) • We know that the fMRI signal is triggered by the metabolic demands of increased neuronal activity – but the details of this process are only partially understood • Altered neuronal activity  Changes in local haemodynamics: - need for glucose? - For oxygen? - Some combination of both? Various conflicting bits of evidence • In this talk: altered neuronal activity  fMRI signal (the linear transform model)

20. The success of fMRI as a research method depends on understanding the relationship between the fMRI signal and the underlying neuronal activity • If the linear transform model were satisfactory, this would be great (!): • it would mean we could reliably estimate the underlying neural activity from the Haemodynamic Response Function (HRF) • But can we? most studies simply assume we can; if the model weren’t a good approximation, this would make HRF measurements worthless • This is important because some fMRI and neurophysiology experiments have yielded conflicting results

21. X For example, the role of V1 in spatial attention: • Visual discrimination improves when cued to attend, without moving the eyes, to the stimulus location • Does the attentional effect occur early in visual pathways (V1)? • Single-cell recording in monkeys: NO • fMRI in humans: YES • attentional effects in further regions (e.g. V4) much larger in human fMRI than in monkey electrophysiology • Is this due to species difference, or a difference in what is being measured?

22. Things the fMRI signal MAY reflect (but we don’t know): • Firing rates of the local neurons • but also activity that doesn’t evoke spikes: • Sub-threshold activity • Simultaneous excitation and inhibition • Modulatory inputs (e.g. top-down and • feedback from higher cortical areas) • and other effects: • Changes in neuronal synchrony without a concomitant • change in mean firing rate • Large changes in the firing rates of a few neurons • OR • small changes in the firing rate of many neurons

23. To estimate the validity of the linear transform model, it is necessary to see how the fMRI signal correlates with measures of neural activity. • But doing this is not so straightforward • The relationship of fMRI data and neural activity depends on a few factors:

24. 1. How ‘Neural activity’ is measured and quantified: • fMRI: simultaneous activity of MANY neurons in a LARGE region of cortex (millimeters) over a LONG period (seconds). What component of the neural activity most predicts the fMRI signal? • Average firing rate of all / a subpopulation of neurons? • Degree of synchronous spiking? • The Local Field Potential (LFP), believed to reflect dendritic currents? • The Multi Unit Activity (MUA), believed to reflect spiking near the electrode tip? • The current source density? • Some measure of local average synaptic activity? • Some measure of subthreshold electrical activity? • All the above may correlate with each other under some circumstances, but can also vary independently of each other. • Logothetis et.al. (Nature, 2001): simultaneous fMRI, LFPs and MUAs in rats. Concluded that BOLD fMRI signals “reflect the input and intracortical processing of a given area rather than its spiking output.”

25. 2. fMRI acquisition technique • BOLD (Blood Oxygen Level Dependent), the most common, provides a mixed signal dependant on: • blood FLOW • blood VOLUME • blood OXYGENATION. • Variations on the technique can be used to emphasize or de-emphasize one or another of these components: • Pefusion-based fMRI blood flow • Injections of various compounds blood volume • Diffusion-based fMRI cell swelling (after excitation) • So far, only little work has been done to quantify the relationship between different fMRI techniques.

26. 3. Experimental protocol & data analysis • Early days of fMRI: worry that the signal arises entirely/mostly from large draining veins  misleading on localization • For example, visual stimuli at nearby locations  activity in nearby (but distinct) loci in V1 • But if the fMRI signal were only evident in a large vessel draining blood from V1, The activity from both points would seem to occur in the location of that vessel

27. The solution: appropriate experimental protocols • For visual stimuli, retinotopic maps are produced by using stimuli that move through the visual field slowly, evoking a traveling wave of neural activity across the grey matter • Large veins draining a large portion of visual cortex are de-emphasized because the blood flow and oxygenation in these vessels are roughly constant throughout the experiment • Hence the experimental design can be crucial for precise localization • So it is important to consider this each time a new protocol is developed

28. Temporal summation • According to the linear transform model, it should be possible to predict the response to a long stimulus presentation by summing the responses to shorter stimuli • For example, the response to a 12-second stimulus should be the same as summing the response to two identical consecutive 6-second stimuli • This temporal summation holds up well in some experiments, but fails in others • Why?

29. Two possibilities: • Blood flow is proportional to neural activity, but BOLD fMRI signals have a non-linear dependence on flow: The BOLD signal might saturate at high levels of blood flow, as further increases in flow would cause negligible decreases in the concentration of deoxyhaemoglobin. So a moderately-strong stimulus could evoke a near-maximal fMRI response, leaving little room for further increase in response (even to a stronger stimulus).

30. Two possibilities: 2. Temporal summation works, but experiments may miss it: These experiments observed primary sensory and motor areas. In such areas, Short-duration stimuli are expected to evoke disproportionately large neural responses for ~3s. This is because such neurons: - Show large transient responses after stimulus onset - Adapt during prolonged stimulation - Their responses are boosted by attention, which is likely to be engaged automatically by stimulus onset Experiments have not been carried out yet to measure the temporal summation of fMRI responses while explicitly controlling for these factors

31. Colocalization of fMRI and neural activity • The linear transform model implies that the fMRI response should be at the same location as the underlying neural activity • However, this may depend on the specific method of data analysis. For example: • The specific HRF used may pinpoint the center of fMRI activity 2mm-1cm from the center of intrinsic optical-imaging activity (Cannestra et.al., 2001). • Dibrow et.al. (2000) report a large discordance between fMRI and electrophysiological recordings in somatosensory cortex of anesthetized monkeys. On the other hand, Wandel et.al. (2000) report excellent agreement between fMRI and the electrophysiological literature for early visual areas. This difference between somatosensory and visual maps might reflect a fundamental difference in the basis of the Haemodynamic response in the two areas.

32. Conclusions • Is the linear transform model a reasonable and useful approximation of the way fMRI data reflects neural activity? • Heeger & Ress (2002): YES, • for some recording sites / measures of neural activity • in some brain areas • using certain experimental protocols • NO, under other circumstances. • We need to know more about the relationship between various measures of neural activity, and determine whether they reflect different aspects of neural function.

33. Thank you