Controversial take? I consider myself a neuroscientist, and I am not able to understand the usefulness of fMRI for cognitive neuroscience studies. (fRMI seems like a great tool to diagnose brain cancer, though.)
In fMRI, every voxel represents several cubic millimetres of brain tissue comprising millions of neurons; the temporal sampling is 2 seconds, when neurons fire action potentials in the ~10 millisecond range, and fast behavioural responses are in the ~300 millisecond range; and the signal measured is blood flow which is somewhat correlated with neural activity at those timescales.
fRMI studies in patients with chronically implanted electrodes (to detect the location of epileptic centres) seem to indicate that areas with low fRMI signal aren't necessarily "unimportant", on the contrary, a small percent of neurons in that area may be critical, yet their activity isn't captured in the fMRI signal as significant. Studies from Ueli Rutishauser and collaborators come to mind.
Then there's the issue of brain "areas". The study of the brain as made of compartments breaks down at close scrutiny. First, monitoring neural activity of the visual cortex in the absence of visual stimulus showed that neuron activity tracks body motion (Carsen Stringer et al. 2019 https://www.science.org/doi/abs/10.1126/science.aav7893 ); in other words multi-sensory integration is the norm. Second, high-functioning hydrocephalic cases present a greatly altered brain architecture with the grey and white matter occupying a tiny fraction of the overall volume. Third, accidents have revealed great plasticity in brain areas, with areas not being spatially stable but rather able to expand over adjacent areas that are less used because of e.g., a missing body part. Even complete absence of the entire cerebellum (cerebellar agenesis) can result in mild phenotypes (Yu et al. 2014 https://www.doi.org/10.1093/brain/awu239 ).
In other words, brain "areas" is not quite the useful abstraction we would want it to be. And therefore, fRMI imaging of blood flow changes over time across coarsely spatially and temporally sampled brains is, at best, too much of a low pass filter over the signal we'd be interested in monitoring.
Are fMRI studies a case of "there's more light here and therefore I look for my wallet here rather than overthere in the shadows where I can't see at all"? I understand that fRMI, and EEG, are all we have to study neural activity in the human brain, so there's a strong incentive to just go with that despite strong shortcomings. Am I missing something fundamental about fRMI?
The only studies using fMRI that make sense to me are longitudinal studies, where the same patient is imaged multiple times and comparisons are like to like, and have more to do with discovering structural issues related to e.g., ageing than assigning function to any subset of the brain, such as in Linda Geerligs' work (Geerligs et al. 2015 https://academic.oup.com/cercor/article-abstract/25/7/1987/462366 ). Are there any other kinds of fMRI studies that beyond doubt have contributed to our understanding of the human brain?
Cortical area instability–experiments in monkeys
On the instability (plasticity?) of cortical areas, just got reminded of this classic work:
"Somatosensory cortical map changes following digit amputation in adult monkeys", Merzenich et al. 1984 https://onlinelibrary.wiley.com/doi/abs/10.1002/cne.902240408
@albertcardona Thank you for this post! I do enjoy thinking about this topic. There's a lot to respond to here, so perhaps I can do it in pieces.
1) The BOLD signal (captured by fMRI) is only "somewhat correlated with neural activity"
It is correlated with neural activity, albeit there is a low-pass filter on top. It is most correlated with LFP signals in monkeys (Logothetis et al, 2001, https://www.nature.com/articles/35084005 ) and humans (Nir et al, 2007, https://www.cell.com/fulltext/S0960-9822(07)01635-1 ).
The BOLD signal is fairly slow, but you can deconvolve it and get temporal precision on the order of 0.5s or so (also subject to your sampling time). The lower temporal precision and indirect aspect of fMRI compared to ephysn is not so different from GCaMP sensors (especially earlier ones), which have been much less controversial.
For studying cognitive neuroscience, 0.5s can actually be really powerful, as you can set up much more complex tasks in humans and jitter stimulus presentation.
2) Each voxel is too large to see anything interesting
If decades of fMRI research have taught us anything, it's that brain areas do get engaged in a grouped enough fashion in order to measure it. Reward signals, for instance, are particularly salient in fMRI. Just that fact has been enough to get researchers excited about connecting economic rewards to their neural instantiation.
One of my favorite studies showing the power of fMRI, despite these bigger voxels, is a clever analysis showing how visual and language representations are right up against each other in the cortex (Popham et al, 2021, https://www.nature.com/articles/s41593-021-00921-6 ). In what other modality could you get such a result??
Besides, MRI technology is getting better and better. You can already get better spatial resolution if you image a specific brain region. There's also efforts to increase the spatial resolution overall by redesigning scanner components and improving signal processing, leading voxels that can image individual columns:
3) Brain areas are plastic, so does it make sense to build a science around regions that generalizes across people?
The most general form of this argument makes neuroscience feel hopeless. Yes, if you wire visual signals to the auditory cortex of a marmoset, it might indeed process these. However, does that mean we should not study the development and fine structure of the visual cortex for processing vision? Surely not!
Plus, it's true that brain regions can be reorganized across people, but to me it amazing at how much is preserved! For instance, Wager et al (2013, https://www.nejm.org/doi/full/10.1056/nejmoa1204471 ) found a signature of pain that could predict a participant's pain even when trained on other subjects. This is not to downplay individual variability, I think all of neuroscience (across all methods and levels) needs to more fully grasp with this. That said, there are common patterns and they are already quite interesting on their own!
I'll note that I think part of the confusion is a cultural one. Having gone from a cognitive neuroscience fMRI/ECoG lab to a drosophila neuroscience lab, I feel there is a disconnect between the goals and methods between the two fields. The drosophila neuroscientists (unfairly) dismiss cognitive neuroscience as not rigorous enough, whereas the cognitive neuroscientists (unfairly) dismiss work on drosophila as they see flies as "too simple".
I think this reflects a larger gap that I've seen of neuroscientists approaching the brain from a cognitive versus biological angle and how this leads to them to pursuing different goals using different methods. Cognitive scientists are often looking at neuroscience for fingerprints that can clarify cognitive concepts, whereas biologists often look at neuroscience trying to understand natural computation and biological processes connecting to the rest of the body.
@neurolili @albertcardona as an in the weeds EM person like Albert, I feel like “fingerprints that clarify cognitive concepts” needs to be unpacked a bit for people who are mostly going after cellular-level mechanisms. One thing that’s hard is feeling like there are so many more details we know about drosophila or mouse visual cortex, and yet the feeling of “understanding” is still fleeting (albeit slowly coming into view in the fly, I think).
To expand on the "cognitive fingerprint" comment, cognitive neuroscientists often treat the pattern of neural activations evoked by a stimulus or mental concept as a "fingerprint" for that concept.
For instance, I remember going to a talk by a social scientist where she showed similarities in the neural activity between two tasks: reading about companies interacting or reading about people interacting. In contrast, reading about objects interacting produced different neural responses. To see whether corporations are seen as people or objects, the specific areas activated are not important, but the similarity of the patterns is. The activation pattern is treated as a fingerprint: the grooves in the finger matter only in that they consistently identify the finger's owner.
Treating the neural activity as a fingerprint and inferring mental concepts is a big part of cognitive neuroscience. However, this can frustrate biologists who care about the specifics of the neural computations. What's more, in systems neuroscience, my impression is that behaviorism rules and it frowns upon inferences about the psychology of non-human animals. Thus the gap is even bigger!
@csdashm @albertcardona Ah, I wanted to add also that I totally agree with you about knowing much more details about the structure of fly brain and nerve cord as well as mouse cortex, yet still grasping at a nice explanation of how it all works.
I have found the integration of connectomics with functional imaging in central complex so satisfying to watch unfold. Certainly you couldn't do this with fMRI! The genetic targeting, spatial resolution of imaging, and connectomics are just not there at this time.
Still in other places my impression is that the strangeness and complexity of the connectome is truly humbling. IMO, we need better constraints from functional studies and from behavior. Then again, my research is in quantifying behavior so I'm probably biased there haha.
@neurolili @albertcardona - Good points Lilli, but I do want to point out that the poor temporal resolution of old calcium indicators was one of the primary reasons why Janelia put so much effort into calcium indicators to get them to be fast, and also that even the slow ones at least have single-cell resolution. And speed is part of why GCaMP6 was so popular.
So I don't think that argument gets very far. (I didn't trust the slow calcium indicator signals all that much either.)
@ichoran @albertcardona Ah yes that's a good point! Yes, perhaps I was overzealous in comparing GCaMP to BOLD. It's more like doing widefield imaging with GCaMP6s.. Even in flies it could be interesting to do this, although there it would be more powerful complemented with genetics and anatomy. Sophie Aimon did some work along these lines, although this whole subfield is still early.
@albertcardona I pretty much agree with you, but I think it depends also on what you're looking for.
To use a programming/computer hardware analogy, if you want to figure out an algorithm, fMRI is useless. If you want to know whether your computation is hitting RAM hard or requires a lot of floating point operations or is engaging the graphics card a lot (and whether shaders are used), it's pretty handy.
Rather than fMRI being a case of looking for the wallet where the light is, it's more like it's showing you where to put the light to look for wallets (if it's done well).
However, the number of overinterpreted fMRI studies does rather disappoint me. The inference "we didn't see a fMRI signal in this area so it's not important in this task" seems to happen far too often.
Then again, this is one of the biggest mistakes of scientists in every area: assuming that absence of evidence is evidence of absence when the tools and/or statistics were not specifically being deployed to address that exact question (i.e. putting bounds on how absent something is).
As a rough heuristic, I find it depressingly accurate to assume that a sentence starting with "This isn't significant, so" is one that will end with an unjustified conclusion.
@ichoran Thanks Rex for your insight. Overinterpretation of data and the pressure to publish positive findings or to simply publish-or-perish has brought us here. I wonder, though, what percentage of #fMRI papers would pass your filter: "if you want to figure out an algorithm, fMRI is useless". Many seem to claim new understanding of brain function.
@albertcardona I guess I could go through a representative sample of recent fMRI papers and check. Knowing whether a process is RAM-intensive or uses shaders does give you new understanding of the function of your computational device, even if it's not an algorithm; there I think the track record is not bad (especially if you ignore the discussion section). But this is just a general impression; to get at anything like "what percentage" we'd have to actually pick out papers and score them, I think.
QOTO: Question Others to Teach Ourselves
An inclusive, Academic Freedom, instance
All cultures welcome.
Hate speech and harassment strictly forbidden.