Reviewer #2 (Public Review):
This work is a cross-validation of an x-ray tomography technique (SAXS) and an optical microscopy technique (SLI) for imaging axonal orientations ex vivo. These innovative methods were introduced in recent papers by the authors, who have teamed up here to compare them side-by-side on the same tissue samples for the first time. The two methods are both label-free (do not require staining) and they are quite complementary. SAXS can provide full 3D orientation measurements on intact tissue, but it operates at a mesoscopic resolution and requires access to a synchrotron. SLI can measure the orientations of multiple fascicles per voxel at a microscopic resolution and relies on more widely accessible equipment, but its accuracy suffers for fiber orientations perpendicular to the imaging plane and it requires tissue to be sectioned before it is imaged. Therefore it makes a lot of sense to explore the complementary strengths of these two techniques, and to use one to "fill in the blanks" of the other. The paper also compares the orientation measurements obtained with SAXS and SLI to those obtained with diffusion MRI. The latter provides only indirect measurements based on water diffusion, at a mesoscopic resolution somewhat lower than that of SAXS, but has the benefit of being feasible in vivo.
A limitation of this study is that conclusions on the comparison between SAXS and SLI are drawn from only 2 sections of a partial monkey brain sample and 2 sections of a partial human brain sample. Conclusions on diffusion MRI are drawn only on the 2 human sample sections. This is particularly an issue for the comparison to diffusion MRI, as the diffusion MRI voxels are wider than the section thickness, hence one cannot preclude that any orientations detected with diffusion MRI but not with SAXS and SLI come from the portion of the voxel that is missing from the corresponding SAXS/SLI section.
The stated aim of the paper is to provide a framework for combining the complementary benefits of SAXS and SLI, rather than simply presenting the results of a cross-validation study. This is a significant and ambitious aim. However, in order for this to serve as a framework, there would have to be clear prescriptions for how researchers interested in obtaining ground-truth measurements of axonal orientations would do so by using these two methods in tandem. This is not adequately developed in the paper in its present form. For example, the results show reasonable agreement between SAXS and SLI orientations when fibers lie within the SLI imaging plane and decreasing agreement for fibers with increasing through-plane inclination. How would the two methods be combined in voxels where they disagree? Would one use SLI orientations in voxels with fewer through-plane fibers and SAXS orientations in voxels with more through-plane fibers? How would voxels be assigned to each category? How would the orientation vectors from the two modalities be composed and how would the resolution difference between the two be handled? When the through-plane measurement of SLI is unreliable, is its in-plane measurement still reliable? That is if there were one mainly in-plane and one mainly through-plane fiber population, would the orientation of the former still be measured correctly by SLI? There is also considerable agreement reported here between through-plane orientations obtained with SAXS and diffusion MRI. Would this mean that diffusion MRI itself could be used to supplement SLI with through-plane orientations? Any clear set of prescriptions along these lines would represent a framework for imaging orientations by combining modalities. This, however, would require detailed steps for how to perform the combination and use the multi- vs. uni-modal framework to reconstruct connectional anatomy.
A key advantage of SAXS is that it can be performed on intact samples, i.e., before any nonlinear distortions of the tissue are introduced by sectioning. Thus it can provide an undistorted reference, with contrast on axonal orientations that would be absent in, say, a structural MRI of comparable resolution. This contrast could be used to drive registration of the distorted SLI sections to an undistorted SAXS volume, and therefore is a key way in which the two techniques can complement each other. Here, however, this is not explored, as SAXS is performed after sectioning. It is not clear if this is the authors' prescription for how a combined SAXS/SLI framework would be implemented, or if it was done specifically for this study. First, it would seem that SAXS on the intact sample would be lower maintenance, requiring less setup time and hence potentially less overall beamtime than performing SAXS on each section separately. This would make it more practical for routine deployment beyond a few sections. Second, because the SAXS data are now nonlinearly distorted, they cannot be affinely aligned to the MRI volumes. While, in principle, performing both SAXS and SLI on the sections may facilitate the comparison between the two, having to unmount, rehydrate, and remount the sections in between may negate this advantage, as now there is no guarantee that SAXS and SLI can be affinely registered to each other. Here all these registration steps are performed affinely, so it is unclear to which extent the computed errors between modalities are characterizing the inherent limitations of the respective contrasts, or limitations of the registration technique. Some of the alignment is performed manually, for example, specific regions of the images are realigned by hand, and the slice of the diffusion MRI volume that is aligned to the SAXS/SLI sections is chosen by hand. Again, for this to serve as a framework that can be deployed on whole samples, there would have to be clear prescriptions for how to perform these steps robustly, how to ensure that the MRI can be acquired in a coordinate frame parallel to the sections, etc.
Finally, the paper puts forth a general conclusion that diffusion MRI overestimates the number of fiber populations per voxel, on the basis of small ODF peaks appearing perpendicular to the main ODF peaks. Of all conclusions in the paper, this is the least convincingly supported by evidence. First, these small perpendicular peaks are a known artifact, which would be typically eliminated by ignoring ODF peaks below a certain amplitude, a common practice in diffusion tractography algorithms. The authors refrain from using an amplitude threshold, with the rationale that it may also remove true diffusion orientations. However, they apply a threshold when they detect SLI peaks (a rather stringent 8% of the maximum). Second, the explanation that these artifactual peaks may appear due to vessel walls is not convincing. Vasculature is sparse. A single vessel wall will not impact the diffusion signal in the same way as a bundle of parallel axons. In an axon bundle, water molecule displacements are restricted in all directions except parallel to the axons. A single vessel wall in a voxel will not have the same effect on displacements (which are much smaller than the size of the voxel). From Figure 5, it looks like there would be at most 1-2 of these vessels in a diffusion MRI voxel, and they would not be in all voxels. This cannot explain the widespread appearance of these small artifactual peaks. Third, many ODF reconstruction methods have parameters that can be adjusted to make these artifactual peaks more or less prominent. The default parameters may be optimal for in vivo but not ex vivo data, due to the effects of fixation. In light of these concerns, I would caution against making such a general statement about all diffusion MRI in the human brain, especially on the basis of a single diffusion reconstruction method applied to a single location in one brain.