If you have heard any of the following terms and furrowed your brow, this post is for you:
data-driven, Pain Matrix, neural networks, functional connectivity, small-worldness, multimodal imaging, brain hubs, and more…
A dramatic shift in our understanding of brain physiology has taken place over the past 25 years based on data from functional magnetic resonance imaging (fMRI) studies. In particular, interpretations of fMRI data (and ultimately of brain function in general) have rapidly evolved with the advent of resting state-based analyses of fMRI data. All of this can be a bit overwhelming for individuals who aren’t immersed in the neuroimaging world and want to understand the concepts. In this post, I will attempt to demystify this ideological shift for those who are curious, baffled, inspired, or unaware of the rationale underlying this evolution.
Note that this is an ongoing work in progress, so you’ll find that I’ve added substantial content every week or so. Also note that I am by no means an expert, but I am fortunate enough to have trained with the individual whose work has inspired much of this progress. I will refer to brain “activity,” but keep in mind that fMRI technology is an indirect measure of neural activity (it reflects changes in blood oxygenation related to glucose metabolism, which is required for neurons to fire).
Before we move forward, let’s consider where this all began.
In the 20th century, our theories of brain function were mostly extrapolated from lesion studies. In these studies, the loss (or gain) of function that occurred after a brain region was surgically removed would be used to infer the “job” of that brain region. The prototypical example of the lesion approach is provided by patient H.M. (Henry Molaison), who developed severe anterograde amnesia (inability to form new memories) and moderate retrograde amnesia (inability to recall previous memories) after large portions of his hippocampus, amygdala, and surrounding regions were removed to control his temporal epilepsy. With such dramatic results, it was natural to conclude a one-to-one correspondence between the brain regions that were removed and the “missing” mental functions (e.g., assuming the hippocampus is responsible for recall of past memories and creation of all new memories). While it it true that the hippocampus is an important structure that mediates the formation of new memories and recall of previously formed memories, memories do not reside within the hippocampus. They are distributed throughout the brain. Therefore the importance of an individual brain region is its capacity to communicate, or share, information with other brain regions. As a result, the behavior of a complex system like the brain is dependent on interactions between brain regions.
Let’s start small. Most people are familiar with pixels, the smallest unit of information in digital pictures, computer screens, or digital cameras. Pixels are usually squares organized in a 2-dimensional grid, and the number of pixels in a picture determines its resolution. In brain imaging, the smallest unit of information is called a voxel (volumetric pixel). A voxel is a 3-dimensional pixel (a cube!) that is defined by its position in space, similar to an address that is defined by it’s position along a street, within a city. A single voxel is typically 1 cubic millimeter in size and it represents the space taken up by approximately 100,000 neurons. There are 20,000 voxels in a given brain. No voxel is an island–these populations of neurons are receiving and integrating information from near and distant neurons. We can then assume that the neuronal activity represented within a single voxel depends on the neural environment in which it is embedded.
Two Interpretations of How the Brain Works
A majority of fMRI studies employ a “task-based” study design. Here, “task” means any experimental situation where the participants are asked to do something or pay attention to specific features of sensory, emotional, and/or cognitive stimuli. The fundamental problem with task-based studies is that the design of these experiments and their statistical evaluation all rely on an experimenter’s idea of how the brain works—which is fundamentally incomplete and often wrong. For example, task-based studies are often analyzed by selectively looking at brain activity during or immediately after an event of interest. The assumption here is that there should be a one-to-one correspondence between an event and the related brain activity. For instance, if I look at a picture for 5 seconds, the corresponding statistical analysis would be to look for changes in brain activity that occurred within those 5 seconds and to contrast these changes with “baseline” brain activity (the arbitrary moments immediately before and after the 5 seconds were are interested in). There are many logical reasons why this doesn’t make sense, and many practical reasons why this is the best that researchers can do based on our current understanding of the brain.
Resting State fMRI
The conventional approach to fMRI data analysis, which assigned discrete functions to different brain regions (“blobology”), has evolved into a network-based approach for understanding brain function over the past 5-10 years. Natural fluctuations of brain activity in the absence of a task, or resting state activity, reveal that the brain functions as a network that efficiently traffics information between local neighborhoods of brain regions (nodes) and distant neighborhoods. According to this view, a specific brain region may indeed mediate certain types of information processing, yet it requires broad participation from other brain regions to execute this specialized processing. For example, the primary somatosensory cortex (S1) will reliably exhibit activation when body parts are stimulated, used, or even imagined to be used, with individual body parts topographically represented in Penfield’s homunculus. For instance, stimulation of the vagina, clitoris, and nipples are encoded in distinct regions of S1. However, S1 interactions with other brain regions are needed to mediate anticipation, predictability, intensity, emotional valence, and qualia of the sensory input.
Where is Pain in the Brain?
In my field, there have been 2 decades of controversy regarding “where” pain is located in the brain. Historically, it has been assumed that specialized neurons in the peripheral and central nervous systems uniquely encode nociceptive information in a way that is fundamentally different than the encoding of non-painful sensations. Decades of searching for nociceptive-specific cortical neurons have yielded perhaps a few hundred such neurons, interspersed amidst non-nociceptive neurons. Given the essential adaptive function of pain, the idea of a small number of neurons having such a massive effect on sensation, perception, and behavior does not appear to be a biologically reasonable assumption. Either way it is almost assured that they are too rare to be detected with brain imaging methods because a few hundred neurons will have very limited impact on the neuronal population represented by a single voxel. Therefore neuroimaging studies that claim to yield pain-specific brain activity should be taken with many grains of salt.
Despite this evidence, pain neuroimagers have increasingly interpreted brain activity evoked by acute noxious stimuli as being unique to the pain experience. The idea of a neural “Pain Matrix”—a network of brain regions that are sufficient to produce the conscious experience of pain—has gained strong popularity because it is a convenient heuristic replicated by many international laboratories (and even myself!).
The problem is that such logic is based on reverse inference: “Observations of brain activity in region X must be responsible for the concurrent cognitive process Y.” This is equivalent to the logical fallacy of inferring a relationship between two variables that is in fact driven by a third unseen variable (“correlation does not equal causation”). Moreover, many pain imaging studies that have identified the “Pain Matrix” have lacked control tasks that are needed to statistically account the background cognitive processing required to support a conscious, sensory experience. This includes brain activity associated with working memory, salience/attention, prediction/anticipation of pending sensory events, as well as and sensorimotor processes like motor movement, encoding of stimulus properties, etc. that occur before, during, and after pain perception. An interesting critique has also been raised in relation to the neural activity associated with an individual’s report of her internal experience, which was dissociated from the internal experience when participants were not asked to self-monitor and report their experiences.
In other words, brain regions involved in pain processing serve many functions that are unrelated to pain. Therefore they cannot be specific to its unique experience.
Combining Neuroimaging Techniques
As suggested by the Hebbian axiom, “Neurons that fire together wire together,” multimodal imaging provides a more vivid snapshot of neural activity, the axons that propagate this activity, and the brain regions that determine how the information is used. fMRI data is increasingly combined with other imaging modalities that provide complementary biological information. Other imaging modalities include T1-weighted anatomical images (which measures gray matter, or neurons) and diffusion tensor imaging (which measures white matter, or tracts of axons). fMRI is also used to refine transcranial magnetic stimulation-based therapies.
Brain function and anatomy are not fixed; rather, neural function and structure can dynamically adapt to reflect new learning. As a result, neuroimaging can provide real-time glimpses of the neural machinery related to healthy and disease states, and how they change over time and with treatments.