In recent years, a multitude of debates have arisen to attempt to understand this anatomical gap in between non-human primate and human white matter neuroanatomy, where certain white matter bundles exist in the human brain but do not exist in other primates. For example, the fronto-occipital fasciculi¹, superior fronto-occipital fasciculi (SFOF) and the inferior fronto-occipital fasciculi (IFOF) are human specific bundles that are often debated in literature². Recently discovered or rediscovered bundles, such as the vertical occipital fasciculus (VOF)³ and frontal aslant tract (FAT)⁴, are examples of novel human specific bundles which are controversial in their existence and definition in humans.

In this work, we will add to this controversy by providing novel insights with a highly realistic diffusion MRI phantom with known underlying white matter ground truth anatomy and advanced evaluation techniques. This MRI phantom was used in the ISMRM 2015 Tractography Challenge. For the purposes of this abstract we will focus solely on the false positive connections/bundles found in the 96 submitted tractograms for this challenge and evaluate their seemingly meaningful white matter structures.

The MRI phantom was comprised of 25 well-known human white matter fasciculi^5,6, representing a known ground truth. Raw diffusion data was derived from the 25 bundles for tractography. Hence, the 25 bundles inform on the spatial extent, geometry, length, size and orientations/crossings/branching of the white matter mask of the data both at the local voxel scale and global connectivity scale.

For each of the 96 tractograms, the validity of the tractogram was analyzed in context to the ground truth and the invalid connections were classified into a new tractogram of more than 100,000 streamlines. Then, candidate deterministic and probabilistic tractograms were used to manually dissect anatomically supported bundles which were not included in the phantom’s 25 ground truth bundles. The left and right FAT, SFOF, VOF and inferior and middle longitudinal fasciculi (ILF/MLF) were dissected. An automatic bundle recognition algorithm⁷ based on QuickBundles⁸ clustering was used, to find these bundles from all other 96 submitted tractograms. Clusters of 100 streamlines or more were saved, to ensure that the invalid bundles do not arise from spurious tracts.

Figure 1 reports the percentage of times each invalid bundle was found (out of 96 submissions). Recall that these white matter bundles are NOT in the ground truth 25 valid bundles but are found by tractography algorithms. We note that most invalid bundles reported are frequently found and not found mostly by DTI tractograms. Figures 2 and 3 illustrate representative invalid bundles from DTI streamline tracking, HARDI deterministic and HARDI probabilistic tracking.

Figures 4 and 5 illustrate heat maps of the locations where there are the most invalid and valid bundle overlap. These maps are created by summing binarized images of invalid and valid bundles respectively. Note that the major white matter “bottlenecks” light up such as the temporal stem, locations in the brainstem area, as well as locations in the centrum semiovale and the motor strips.

Regardless of the tractography pipeline used, with or without pre-processing, with or without crossing fiber models, regardless of tractography algorithm used or streamline post-processing, many invalid bundles are found in the tractograms. The presence of invalid bundles in tractograms has been a known fact in the tractography community^{9,10,11,12,13}, where it has been shown in toy examples that white matter ground truth configurations can easily lead to a large number of false positive bundles. However, to the best of our knowledge, this is the first time that invalid bundles are reported in a realistic brain-like phantom. What becomes alarming is how close the invalid bundles approximate true anatomical structures. These invalid bundles connect valid cortical areas and appear to be dense and well structured.

What becomes critical, is that these invalid bundles originate from a handful of “bottleneck “ or “high traffic” locations in the brain. These bottleneck locales are usually where 3 or more valid bundles come together, almost in parallel. These locales are ideal candidate locations to generate all possible endpoint pairs of connections transversing these areas.

Our objective is not to discourage the diffusion community from using tractography, as it is a vital tool that has greatly contributed to our understanding of the brain, but we would encourage critical evaluation of methods and conclusions derived from tractography results¹⁴. More importantly, it is clear that novel tractography methods are needed to better choose what local orientations are to be “connected” together. Hopefully, advanced diffusion microstructure modelling and multi-modality integration in tractography will help overcoming this limitation of tractography.