In 2009, the NIH launched the Human Connectome Project. This endeavor centers on the functional mapping of white matter circuitry in the human brain. Like other esoteric neologisms sandwiched between “Human” and “Project,” we should automatically assume that the summit will prove elusive. Our current technological capabilities have allowed us to capture the form of unknowable biological systems, but the dizzying extent of function is another story. Although the Human Genome Project was completed in 2003, its aftershock is still reverberating. Now researchers are grappling with the epigenome and the transcriptome and non-coding RNA. This is the era of personalized medicine, and it is easy to see parallels here with the experts studying in vivo nuances of diseased, aging, and healthy brains. In theory, a fully-realized connectome could make neurology and CNS surgery a purer science. Despite great leaps in the field of imaging, the project remains incomplete. But that is an unsatisfying characterization. There are many scenic vistas to gawk at on the journey to the top – and we are a mile up from the bottom.
In their 2017 NMR in Biomedicine paper “Diffusion MRI fiber tractography of the brain,” Jeurissen and colleagues discuss the current landscape and limitations of white matter fiber tracking. This burgeoning field has been fruitful for the clinical research and surgical treatment of a variety of neurological conditions like epilepsy and Alzheimer’s, but the authors present a balanced argument that will surely deflate the temerity of those who may consider d-MRI tractography the last frontier of neuroscience. There is a difference between a “tractogram” (awesome reality) and the “connectome” (fantasy?) Let’s say a patient with a neurodegenerative brain disease is looking for answers about his diagnosis. He asks his neurologist: “where?” That question can be addressed with sophisticated neural cartography. He then follows up with “how?” and “why?” The physician pauses. “You are your connectome.”1 Now we are entering murkier, borderline-philosophical waters. Until his aphorism can be boiled down even the most eminent researchers are relegated to educated estimation.
White matter is sheathed in myelin, a fatty coating of sterols and glycolipids that insulates the axon and restricts the movement of water molecules. Myelin water molecules are “bound;” they collide infrequently and have a shorter spin-spin relaxation, yielding a rather long correlation time. Therefore, a very brief echo time is required to detect the myelin water signal. These functions help distinguish white matter from its gray counterpart and cerebrospinal fluid. White matter tractography utilizes a special form of MRI called Diffusion MRI or, specifically, Diffusion Tensor Imaging (DTI). Mori and Tournier state that “the concepts behind DTI are commonly difficult to grasp, even for magnetic resonance physicists.”2 That assertion checks out, but if you can gauge the rate of water diffusion inside the brain, the shape and orientation of fatty axonal fibers can be deduced in any given voxel. A sizeable collection of voxels is then pieced together to create a three-dimensional map of computed vector fields. In DTI, eigenvalues and eigenvectors are used to assess the diffusion properties of tissue. A starting or “seed point” is chosen in the white matter, or alternatively, the white-gray matter barrier and the orientation is calculated in order to depict the streamline or the estimated path of the fiber.
According to the authors, this recently posed a major problem. Older DTI technology was only capable of showing a single fiber population per voxel. That might have been fine for regions of sparsely populated fibers, but it is known that shorter fibers are densely packed and entangled entities. They cross over and under each other. This leads to false positives (assuming the track continues) and false negatives (assuming the track ends when it does not). But over the past few years, advanced modeling methods have managed to represent multiple fiber populations per voxel. This development is key to traversing the subcortical labyrinth. Surprisingly, old-school DTI still reigns supreme in clinical practice. One has to believe that the higher-order models will take the throne soon. Perhaps DTI-based tractography is less expensive, more familiar, and just flat-out good enough for all intents and purposes.
After fiber orientations have been acquired, their trajectories must be established. This is done by Euler integration for outmoded DTI, or Runge-Kutta integration for the latest iterations. The goal is to collate the distances (step size) from one point on the fiber to the next. When employing large step sizes, interpolation errors become more pronounced and track-tracing veers off course. Therefore, higher-order integration methods and small step sizes seem to be the best recipe for accuracy. We can see how technological advancements in these methods – be it in modeling or integration – have launched the connectome industry into a truly HD age. Still, a review of interpolation methods reminds us that these measurements are nevertheless approximations. On one hand, we have the error-prone nearest-neighbor interpolation, and on the other, the more precise trilinear interpolation. Errors from tri-linearly interpolated streamlines do not stray as far from the seed point.
But a tract without a good seed point is like the scribble of a toddler who takes a crayon to his Denny’s menu maze at a spot other than the designated entryway. How do experts know where to begin? Well, they need a deep knowledge of neuroanatomy and a ballpark understanding of functional connectivity. Then they can designate regions of interest (ROIs), which may be local or global. The global method is known as whole-brain tractography, and if one really wants to achieve connectome actualization, it is necessary to visualize this big picture. Researchers should be wary about their method of seeding though. White matter seeding leads to overemphasis of the most prominent bundles. White-gray interface seeding provides better contrast, but due to the relative lack of seeds, may contain errors while failing to capture the entirety of the white matter circuit. In those circumstances, the researcher needs to know when to terminate the track. Older DTI used fractional anisotropy to gauge precipitous drops in white matter; the leading-edge models utilize the continuous fiber orientation function.
Once mapped, a variety of techniques are used to “virtually dissect” and group the white matter bundles into coherent anatomical structures. Algorithms estimate the relatedness of a given tract to its seed of origin, thus making inferences on the degree of functional connectivity between the two points. These algorithms are a topic of contention. A researcher may subscribe to the straightforward deterministic approach. He might believe that the fiber orientation data contain self-evident structural truths. He might not worry about the shortcomings of the local approach to tractography, exemplified by streamlining, which can be done quickly, but is prone to interpolation missteps and may sometimes highlight phantom fiber tracks. A different researcher may idealize a global approach to tractography, but this brand also carries baggage. Global methods offer a better signal-to-noise ratio and minimize artifacts, but their implementation can be infinitely complicated to the point of being impractical. And they, too, can lead to phantom fiber tracks.
Pragmatists though, as the authors claim to be, would likely support a more agnostic ideology. Probabilistic tractography maintains that there is a chance that the fibers have a particular orientation, but also a chance that they do not. The proof is in the k-space distribution and is dependent on the particular faction of probabilistic fiber tracking to which they belong. Some adhere to the notion that there is a “dominant” fiber orientation in a given voxel, while others hold that each voxel contains a diverse array of orientations. The latter school, based on the continuous fODF model, is associated with more anatomically accurate tractograms. That makes sense; white matter fibers are not interstate highways. There are curves and complexity and randomness to account for, so to assume a dominant orientation would be to ignore the fact that fibers can be more like winding country roads. The authors maintain that while the two schools of thought have different aims, they in fact complement each other. I can see that too. There will always be opposing viewpoints in science and academia. To solely endorse one over the other would be like a present-day psychologist claiming to be a Freudian or a Jungian or a Skinnerite. Each pioneer enriched and influenced the field. None of them ran away with it.
No singular method for demystifying the connectome predominates because they are all fraught with quantification problems. There are too many confounding variables. If you’re targeting the frontal pathways, orientation data from the temporal pathway will suffer. And it seems that advanced imaging techniques are still stymied by unpredictable regions of crisscrossed fibers. Track density imaging (TDI) has offered a solution, one that offers vivid anatomical contrast and high resolution. But of course, there is a trade-off. TDI uses millions of seed points, which actually “hides” noise and can therefore give the mere illusion of accurate mapping. It is also biased toward long-fiber tracks. The authors go so far as calling TDI fundamentally non-quantitative, unless used it in short-track mode with globalist filtering techniques. One gets the impression that no tracking program is universally perfect in isolation, but when supplemented with further refinement processes, the end result will be a pretty good set of images.
But those working on the Human Connectome Project seek more than just a static representation of anatomical structure. We are living in the post-structural age, after all. We are looking for function, for connection, for meaning. Will dominant-pathway tracking deliver? Definitely not. It oversimplifies the subcortical terrain. What about global tracking? There’s more potential here, but water follows the path of least hindrance. Two entangled fibers will influence and blur each other’s diffusion pattern. This creates errors over the long game. Diffusion blurring adds ambiguity to white fiber architecture, making it exceedingly difficult to follow distant connections. We should “curb our enthusiasm,” the authors say. One solution is the application of biologically realistic priors, which ideally terminate in the gray matter. These anatomical constraints offer a template that prevents coloring outside the lines, as it were. As the connectome project has matured, researchers have started to think beyond the white matter itself; where tracks start and end are of paramount importance to functional connectivity. The most recent models utilize multiple diffusion weighting strengths to account for the differences in brain matter composition.
For all its ingenuity, dMRI is at its core a qualitative method. That’s not a bad thing. When rendered and colorized, its images are at once masterful artworks and informative atlases worthy of their ubiquitous splashing on the covers of new textbooks and journals. To fault dMRI for its imperfection in quantification is like criticizing the cameras on the International Space Station for not being able to exact the edges of a pebble. White matter fibers branch, kiss, bottleneck, cross over and under each other. Spatial and angular resolution limitations underrepresent their manifold dimensions of repose. False positives and false negatives abound. The truth is that today’s non-invasive technology cannot isolate tangled neurons. At least on the practical level required for a connectome eureka. Even if they could, then what? Alfred Korzybski said it best: “The map is not the territory.”
It is no surprise that the Human Connectome Project is a work in progress. This is new stuff. It is going to take time. And money. Speaking of which, where is the money? Cutting-edge neuroimaging is not a cheap pursuit. For this project, the NIH awarded less than $40 million in grant money across several institutions over five years. That’s $8 million per year, which is about 0.0002% of the budget. Chum change. But it’s understandable. Diabetes and oncology research presents a more immediate public need and benefits from foreseeable interventions. Connectome funding will follow once breakthroughs in technology translate into solutions for medicine. On the HCP’s website, they list as their number one goal to “develop sophisticated tools to process high-angular diffusion (HARDI) and diffusion spectrum imaging (DSI) from normal individuals to provide the foundation for the detailed mapping of the human connectome.”3 The authors of this year-old paper described many of these sophisticated tools, but did not so much as mention HARDI and DSI. Perhaps it was beyond their scope. Or perhaps the sands of neuroimaging technology are shifting quicker than medicine can keep pace.
This has been a review of the following publication:
Jerrissen B, Descoteaux M, Mori S, Leemans A. Diffusion MRI fiber tractography of the brain.
NMR in Biomedicine. 2017;e3785.
Other Citations:
1: Dr. Francis Collins
2: Introduction to Diffusion Tensor Imaging (Second Edition) by Susumu Mori and J-Donald Tournier, 2014
Benjamin Samuel Young 