Diffusion MRI (dMRI) allows us to infer the local orientation of white-matter (WM) pathways, and to derive measures that are thought to reflect the local structure or integrity of those pathways. Reconstructions of WM pathways obtained from dMRI tractography are currently used to study a wide range of neurological and psychiatric diseases, as well as healthy development and aging. However, the lack of ground truth on the connectional diagram of the human brain is a major impediment to the evaluation and optimization of methods for dMRI data acquisition and analysis, reducing confidence in the findings of dMRI studies. In this talk I will review the methods that can be used to validate WM pathway reconstructions derived from dMRI, describe the challenges associated with each approach, and suggest a roadmap for future investigation.

The need for validating dMRI stems from the fact that it provides us with indirect observations of the orientations of axonal bundles, based on the diffusion of water molecules in and around them. Tractography algorithms then combine the local information on the preferential orientation of water diffusion at each voxel in the brain to identify continuous paths of diffusion through the brain. These paths are generally thought to follow the trajectory of WM bundles, but this is not always the case. Due to the limited spatial resolution of typical dMRI acquisitions, most voxels contain more than one preferential diffusion orientation, i.e., they contain multiple axonal bundles that may be crossing, converging, splitting, etc. Even in the absence of measurement error, where we could derive an accurate distribution of the diffusion orientations in a voxel, this would still not be enough to differentiate between all the possible bundle configurations that could give rise to that pattern of diffusion.

Fig. 1 shows three different configurations that could result in a similar orientation distribution function (ODF). It would be impossible for a tractography method to guess which of the diffusion ODF peaks it should follow when it reaches this voxel. Therefore, it would have to make a decision based on an arbitrary rule, such as minimizing the turning angle, i.e., going as straight as possible. However, it is unlikely that any such arbitrary rule would be consistent with WM architecture everywhere in the brain. This often leads to tractography solutions that are anatomically implausible, e.g., merging sections of different WM pathways that have intersecting trajectories but that should not be connected to each other (see Fig. 2). Any network analysis that relies on whole-brain tractography, without making use of prior anatomical knowledge to constrain the results, contains a large number of such spurious connections. Furthermore, the prior anatomical knowledge that is currently available to tractography algorithms applies mostly to a few large WM pathways. Extensive validation work is still needed to establish the true anatomy for all areas of uncertainty like the one shown in Fig. 1.

Validation of dMRI tractography in non-human primates can be performed with the use of invasive tracers that are taken up and transported along axons. The macaque monkey is a common model, due to well-established homologies between the macaque and human brain. Tracer studies are labor-intensive and require specialized neuroanatomical expertise to annotate the axons that have taken up the tracer on histological slides. The few studies that have compared chemical tracing and dMRI data from the same monkey brain have focused on a single injection. Specifically, injections in the primary motor cortex have been compared to in vivo dMRI in the same macaque [1] or to ex vivo dMRI in the same squirrel monkey [2].

Studies that have combined chemical tracing data from multiple injection sites have relied on tracing and dMRI from different sets of macaques [3-6]. When comparing tracer injections and dMRI tractography from different brains, one cannot eliminate the possibility that connections found in one and not the other may be due to inter-individual differences. This potential confound, however, is counterbalanced by the opportunity to take advantage of extensive databases of prior tracer injection studies.

Validating dMRI tractography in the human brain, where invasive tracing techniques cannot be used, is even more challenging. Post mortem gross dissection has been used to establish the main WM pathways of the human brain and compare them to dMRI reconstructions [7]. However, gross dissection does not have the precision required to follow the trajectory of individual axons, or even small axon bundles, all the way from their origin to their terminations. Myelin stains can be used to infer the local properties of myelinated axons, such as in-plane orientation or density [8-10], within a 2D histological slice. However, they do not provide information on how to follow those axons through the slice.

Recently, optical imaging methods that take advantage of the birefringence of myelin sheaths, i.e., the dependence of their refractive index on the polarization angle of incident light, have shown great promise for imaging brain samples at microscopic resolutions. Polarized light imaging (PLI) has been used to estimate the orientation of WM axon bundles at a resolution of 100μm [11,12]. This technique images each slice after sectioning the brain. This causes non-linear distortions (warps and tears) similar to the ones present in histological stains, making alignment of the processed slices challenging. Optical coherence tomography (OCT) methods do not suffer from such distortions, as they make it possible to image each slice of tissue before cutting it. They are very versatile, as they can be adapted to image cell bodies, dendrites, myelinated or unmyelinated fibers. The use of OCT has been demonstrated for imaging orientations of WM axon bundles at a resolution of 15μm [13]. This technique can be applied to small samples. Unfortunately, the acquisition time, processing time, and storage space requirements of OCT are currently prohibitive for imaging the whole human brain.

All of the approaches discussed above have their limitations. Chemical tracing studies are not applicable to the human brain and they involve a cumbersome procedure of injection, perfusion, slicing, and manual annotation. Furthermore, a tracer injection can reveal all projections from/to a location on the cortex, but not all projections going through a location in the WM. Thus, it cannot be used for direct comparison to diffusion ODFs. All techniques that involve the processing of 2D slices after they are sectioned, including chemical tracing, myelin stains, and PLI, require solving a highly non-linear, across-slice registration problem. Although OCT circumvents this registration issue, and can be adapted to estimate both within-slice and through-slice orientation, it is extremely time-consuming, making it applicable only to small samples.

Each of the methods described here can provide invaluable information on WM architecture but cannot tackle the problem of dMRI tractography validation in its entirety. An integrative approach, with converging evidence on WM organization from multiple independent sources, is needed. To this end, chemical tracing in non-human primates, optical imaging in human tissue, as well ex vivo dMRI at microscopic spatial resolutions, will all play an important role.