Evaluation of surface-based hippocampal registration using ground-truth subfield definitions

Reviewer #2 (Public Review):

DeKraker et al. propose a new method for hippocampal registration using a novel surface-based approach that preserves the topology of the curvature of the hippocampus and boundaries of hippocampal subfields. The surface-based registration method proved to be more precise and resulted in better alignment compared to traditional volumetric-based registration. Moreover, the authors demonstrated that this method can be performed across image modalities by testing the method with seven different histological samples. This work has the potential to be a powerful new registration technique that can enable precise hippocampal registration and alignment across subjects, datasets, and image modalities.

Reviewer #3 (Public Review):

Summary:
In the current manuscript, Dekraker and colleagues have demonstrated the ability to align hippocampal subfield parcellations across disparate 3D histology samples that differ in contrast, resolution, and processing/staining methods. In doing so, they validated the previously generated Big-Brain atlas by comparing across seven different ground-truth subfield definitions. This is an impressive effort that provides important groundwork for future in vivo multi-atlas methods.

Strengths:
DeKraker and colleagues have provided novel evidence for the tremendously complicated curvature/gyrification of the hippocampus. This work underscores the challenge that this complicated anatomy presents in our ability to co-register other types of hippocampal data (e.g. MRI data) to appropriately align and study a structure in which the curvature varies considerably across individuals.

This paper is also important in that it highlights the utility of using post-mortem histological datasets, where ground truth histology is available, to inform our rigorous study of the in vivo brain.

This work may encourage readers to consider the limitations of the current methods that they currently use to co-register and normalize their MRI data and to question whether these methods are adequate for the examination of subfield activity, microstructure, or perfusion in the hippocampal head, for example. Thus the implications of this work could have a broad impact on the study of hippocampal subfield function in humans.

Weaknesses:
As the authors are well aware, hippocampal subfield definitions vary considerably across laboratories. For example, some neuroanatomists (Ding, Palomero-Gallagher, Augustinack) recognize that the prosubiculum is a distinct region from subiculum and CA1 but others (e.g. Insausti, Duvernoy) do not include this as a distinct subregion. Readers should be aware that there is no universal consensus about the definition of certain subfields and that there is still disagreement about some of the boundaries even among the agreed upon regions.

Author Response

The following is the authors’ response to the original reviews.

We would like to thank the Editors for the opportunity to submit a revised manuscript, and the Reviewers for their positive evaluations and constructive comments. We feel that the comments and suggestions significantly improved the quality of our manuscript. We addressed all questions and suggestions in a point-by-point fashion below.

Reviewer #1 (Public Review):

This paper proposes and evaluates a new approach for the registration of human hippocampal anatomy between individuals. Such registration is an essential step in group analysis of hippocampal structure and function, and in most studies to date, volumetric registration of MRI scans has been employed. However, it is known that volumetric deformable registration, due to its formulation as an optimization problem that minimizes the combination of an image similarity term and relatively simple geometric regularization terms, fails to preserve the topology of complex structures. In the cerebral cortex, surface-based registration of inflated cortical surfaces is broadly preferred over volumetric registration, which often causes voxels of different tissue types to be matched (e.g., voxels belonging to a sulcus in one individual mapping onto voxels belonging to a gurys in another). The authors recognize that hippocampal anatomy is similarly complex, and, with proper tools, can benefit from surface-based registration. They propose to first unfold the hippocampus to a two-dimensional rectangle domain using their prior HippUnfold technique, and then to perform deformable registration in this rectangle domain, matching geometric features (curvature, thickness, gyrification) between individuals. This registration approach is evaluated by comparing how well hippocampal subfields traced by experts using cytoarchitectural information align between individuals after registration. The authors indeed show that surface-based registration aligns subfields better than volumetric registration applied to binary segmentations of the hippocampal gray matter.

Overall, I find the methods and results in this paper to be convincing. The authors framed the comparison between surface-based and volumetric registration in a fair way, and the results convincingly show the advantage of surface-based registration. One slight limitation of the current study is that it is uncertain whether the benefits demonstrated here translate to in vivo MRI data for which the authors' HippUnfold algorithm is tailored. The current study utilized the unfolding technique used in HippUnfold on manual segmentations of high-resolution ex vivo MRI and blockface 3D volumes, which are likely closer to anatomical ground truth than automated segmentations of in vivo MRI. However, it is reasonable to assume that given that the volumetric registration to which the proposed approach was compared also used this high-detail data, the advantages of surface-based over volumetric registration would extend to in vivo MRI as well. However, I would encourage the authors to perform future evaluations on datasets with available in vivo and ex vivo MRI from the same individuals.

We thank the Reviewer for the positive evaluation and the thoughtful feedback. We address each comment in the red text below.

We have considered the Reviewer suggestion for a demonstration of the gains from our proposed method in MRI, and decided to include a new analysis of 7T in-vivo MRI data from 10 healthy participants (Supplementary Materials 1: in-vivo MRI demonstration).

It is difficult to assess whether changes to the registration methods are indeed an improvement without same-subject “ground-truth” subfield definitions typically obtained from histology. In this new Supplementary Materials section, we demonstrate an overall sharpening of MRI-mapped features as an indirect indication of better inter-subject alignment (similar to the paper referenced in the comment, below). This is an important proof of concept that demonstrates that the gains made in the current project can be translated to in in-vivo MRI. We did not perform a demonstration of these gains in ex-vivo data, since this also comes with a host of challenges including access to such data and deformations and artifacts associated with ev-vivo scanning. However, we believe that the gains provided by our methods are limited mainly by image resolution and so while we note some concern about the gains from this method at 3T MRI, we expect that in ev-vivo gains provided by our method in higher resolution ex-vivo images should be consistent or better.

We have added the following in-text Discussion of this new analysis (p. 13):

“Ravikumar et al. (2021) recently performed flat mapping of the medial temporal lobe neocortex using a Laplace coordinate system as employed here, and showed sharpening of group-averaged images following deformable registration in unfolded space. This indirectly suggests better intersubject alignment. We perform a similar group-averaged sharpening analysis in Supplementary Materials 1: in-vivo demonstration. Though the gains in image sharpness observed here were modest, we note that current MRI resolution and automated segmentation methods allow for only imperfect hippocampal feature measures. We thus expect unfolded registration to grow in importance as MRI and segmentation methods continue to advance. “

I would also like to point out the relevance of the 2021 paper "Unfolding the Medial Temporal Lobe Cortex to Characterize Neurodegeneration Due to Alzheimer's Disease Pathology Using Ex vivo Imaging" by Ravikumar et al. (https://link.springer.com/chapter/10.1007/978-3-030-87586-2_1) to the current work. This paper applied an earlier version of the unfolding method in HippUnfold to ex vivo extrahippocampal cortex and performed registration using curvature features in the rectangular unfolded space, also finding slight improvement with surface-based vs. volumetric registration, so its findings support the current paper.

Thank you, we agree this is a highly relevant paper and have added a summary of it in the newly added Discussion paragraph which also outlines the new Supplementary Materials section (see previous comment).

Overall, the paper has the potential to significantly influence future research on hippocampal involvement in cognition and disease. Outside of simple volumetry studies, most hippocampal morphometry studies rely on volumetric deformable registration of some kind, typically applied to whole-brain T1-weighted MRI scans. With HippUnfold available for anyone to use and not requiring manual registration, the paper provides a strong impetus for using this approach in future studies, particularly where one is interested in localizing effects of interest to specific areas of the hippocampus. Additional evaluation of in vivo HippUnfold using in vivo / ex vivo datasets, would make the use of this approach even more appealing.

We would like to thank the Reviewer for their enthusiasm for the translatability of this work. We hope they are satisfied with our newly added in-vivo evaluation, and we appreciate the thoughtful suggestions.

Reviewer #1 (Recommendations For The Authors):

No additional recommendations.

Reviewer #2 (Public Review):

DeKraker et al. propose a new method for hippocampal registration using a surface-based approach that preserves the topology of the curvature of the hippocampus and boundaries of hippocampal subfields. The surface-based registration method proved to be more precise and resulted in better alignment compared to traditional volumetric-based registration. Moreover, the authors demonstrated that this method can be performed across image modalities by testing the method with seven different histological samples. While the conclusions of this paper are mostly well supported by data, some aspects of the method need to be clarified. This work has the potential to be a powerful new registration technique that can enable precise hippocampal registration and alignment across subjects, datasets, and image modalities.

We thank the Reviewer for their thoughtful evaluation of our paper and helpful comments. We address them in the red text below each comment.

Regarding the methodological clarification of the surfaced-based registration method, the last step of the process needs further clarification. Specifically, after creating the averaged 2D template, it is unclear how each individual sample is registered to sample1's space. If I understand correctly, after creating the averaged 2D template, each individual sample is then registered to sample1's space via the transform from each sample to the averaged template and then the inverse transform from the template to sample1's space. Samples included both left and right hemispheres, so were all samples being propagated to left hemisphere sample 1 space? The authors also note that a measure of the subfield labels overlap with that sample's ground-truth subfield definitions was calculated. Is this a measure of overlap, for example, between sample 3 (registered to sample 1 space) and the ground-truth (unfolded, not registered) sample 3 labels? It would be beneficial to provide a full walkthrough of one example sample to clarify the steps. Clarification of this aspect of the method is critical for understanding the evaluation of the method.

We would like to thank the Reviewer for the suggestion, and have clarified the passage with the following walkthrough example as suggested by the Reviewer (p. 8):

“For example, sample3 was unfolded and then registered to the unfolded average, making up two transformations. These were then concatenated with the inverse transformation of unfolded sample1 to the same unfolded average, and the inverse transformation of native sample1 to unfolded space. This concatenated transformation was used to project labels from sample3 native space directly to sample1 native space, which should ideally lead to near-perfect subfield alignment in sample1 native space. Dice overlap between sample1 and sample3 registered to sample1 was then calculated in sample1 native space.”

Reviewer #2 (Recommendations For The Authors):

Materials and Methods:

In the Data section, it would be helpful for the authors to clarify whether each hippocampal histology sample is from a different individual or not. Additionally, for the 3D-PLI sample, the authors mention that the anterior/posterior parts of the hippocampus were cut off and the labels were extrapolated over the missing regions. It would be useful to know whether the extrapolation was done manually.

Thank you, we have added separate labels (donors 1-4) for each individual from each dataset. We have also added that the 3D-PLI dataset was extrapolated manually. See the revised Materials and Methods: Data section.

A small clarification, but for the morphological features calculated by HippUnfold, is thickness a measure of how much space each subfield takes up in the 2D unfolded space?

Thickness is measured via HippUnfold, and we have clarified in-text that it is done in each subject’s native space (p. 6):

Results:

In the Results section, a brief summary or description of the Dice overlap metric would be helpful. The authors should also clarify if the Dice metric measures the overlap between an individual sample (e.g., sample3) that has been unfolded and registered/propagated to sample1 compared to the sample1 ground-truth subfields.

We thank the Reviewer, and hope this is now clarified alongside the Reviewer’s Public comment with the addition of the example as quoted in our response to that comment.

We also added to our description of Dice overlap as a measurement (p. 8):

“The Dice overlap metric (Dice, 1945), which can also be considered an overlap fraction ranging from 0-1, was calculated for all subjects’ subfields registered to sample1.”

Figure 3:

In Figure 3A, it is unclear what "moving (sample 3)" refers to. Clarification is needed, and it would be helpful to know if this is sample 3 in native space before it has been unfolded/registered. In Figure 3B, there is a missing "native" before "folded" and "(right)" at the end of the sentence. With these edits, the sentence in the caption would read: "Each measure was calculated in unfolded space (left) and again in the first sample's (BigBrain left hemisphere) native folded space (right)."

We thank the Reviewer, and have now changed “moving” to “sample3 before registration”, and added the suggested caption changes. See the revised Figure 3.

Discussion:

In the introduction, the authors provide a detailed description of the traditional 3D volumetric registration technique that utilizes gyral and sucal patterning as the primary feature for registration, along with other features such as thickness and intracortical myelin. Using their surface-based registration, the authors highlight an interesting finding that hippocampal curvature is the most informative individual feature, and thickness and curvature combined are the most informative features for registration and boundary alignment. In the discussion, it would be beneficial for the authors to discuss the relationship between curvature, thickness, and gyrification (e.g., is there overlapping information across these features) and comment on the reliability of these features observed in the current study compared to past work using traditional methods.

This is an interesting point of discussion, thank you for raising it. We’ve added the following paragraph to the Discussion section (p. 13):

“The feature most strongly driving surface-based registration in the present study was curvature. Many neocortical surface-based registration methods focus on gyral and sulcal patterning at various levels (e.g. strong alignment of primary sulci, with weaker weighting on secondary and tertiary sulci) (Miller et al., 2021). In the present study, hippocampal gyri are variable between samples and so could perhaps be thought of as similar to tertiary neocortical gyri, and this may help explain why gyrification was not the primary driving feature in aligning hippocampal subfields. The methods used to quantify gyrification are often related to curvature, but differ across studies. In the hippocampus, unlike in the neocortex, the mouth of sulci are wide and so sulcal depth, which is often used in defining neocortical gyrification index, is not straightforward to measure. Using HippUnfold, gyrification is defined by the extent of tissue distortion between folded and unfolded space, and individual gyri/sulci are hard to resolve in unfolded gyrification maps, but are readily visible in curvature maps. Thus, hippocampal curvature may be more informative simply due to higher measurement precision. Future work could also employ measures like quantitative T1 relaxometry or other proxies of intracortical myelin content (e.g. Tardif et al., 2015; Glasser et al., 2016; Paquola et al. 2018) for hippocampal alignment, but this is not possible in cross-modal work as in the various histology stains examined here.”

Miscellaneous:

There is a typo on page 11, line 318, with extra parentheses: "(e.g., (Borne et al., 2023;..."

Thank you, we have corrected this error.

Reviewer #3 (Public Review):

Dekraker and colleagues previously developed a new computational tool that creates a "surface representation" of the hippocampal subfields. This surface representation was previously constructed using histology from a single case. However, it was previously unclear how to best register and compare these surface-based representations to other cases with different morphology.

In the current manuscript, Dekraker and colleagues have demonstrated the ability to align hippocampal subfield parcellations across disparate 3D histology samples that differ in contrast, resolution, and processing/staining methods. In doing so, they validated the previously generated Big-Brain atlas by comparing seven different ground-truth subfield definitions. This is an impressive and valuable effort that provides important groundwork for future in vivo multi-atlas methods.

We thank the Reviewer for their positive evaluations, and helpful suggestions. We provide responses to the recommendations in the red text below.

Reviewer #3 (Recommendations For The Authors):

There are a few points I think the authors should address, listed below.

1. As the authors are well aware, subfield definitions vary considerably across laboratories. The current paper states that JD labeled the samples using three different atlas references: Ding & Van Hoesen, 2015; Duvernoy et al. ,2013, and Palomero-Gallagher et al., 2020. This is unclear, however, since these three references differ in their subfield definitions. For example, Ding & Van Hoesen and Palomero-Gallagher define a region called the prosubiculum (area between subiculum and CA1) but Duvernoy does not. Please clarify which boundary rules from which particular references were used here. How were discrepancies across these references resolved when applying labels to the current histological samples?

We thank the Reviewer, and have added the following elaboration (p. 5):

“Since these sources differ slightly in their boundary criteria, and no prior reference perfectly matches the present samples, subjective judgment was used to draw boundaries after considering all three prior works. The “prosubiculum” label used by Ding & Van Hoesen and Palomero-Gallagher et al. was included as part of the subicular complex. See Supplementary Materials 2: ground-truth segmentation for more details.”

1. Another comment has to do more with the "style" of how this paper is written, especially given that this paper was submitted to eLIFE (i.e. not a specialty journal). For example, the motivation for the unfolded with and without registration methods was not well described. Similarly, there was almost no justification for the different methods applied in Figure 4 and I fear that the impact of these results will be lost on a non-expert reader.

We added the following elaboration to the last paragraph of the Introduction section to motivate our benchmark against unfolding without registration (p. 3):

“We benchmark this new method against unfolding alone, which provides some intrinsic alignment between subjects (DeKraker et al., 2018) but which we believe can be further improved with the present methods, and against more conventional 3D volumetric registration approaches.”

We also added a Discussion paragraph on the results shown in Figure 4 which we hope helps to make these results more informative and impactful (p. 13):

“The feature most strongly driving surface-based registration in the present study was curvature. Many neocortical surface-based registration methods focus on gyral and sulcal patterning at various levels (e.g. strong alignment of primary sulci, with weaker weighting on secondary and tertiary sulci) (Miller et al., 2021). In the present study, hippocampal gyri are variable between samples and so could perhaps be thought of as similar to tertiary neocortical gyri, and this may help explain why gyrification was not the primary driving feature in aligning hippocampal subfields. The methods used to quantify gyrification are often related to curvature, but differ across studies. In the hippocampus, unlike in the neocortex, the mouth of sulci are wide and so sulcal depth, which is often used in defining neocortical gyrification index, is not straightforward to measure. Using HippUnfold, gyrification is defined by the extent of tissue distortion between folded and unfolded space, and individual gyri/sulci are hard to resolve in unfolded gyrification maps, but are readily visible in curvature maps. Thus, hippocampal curvature may be more informative simply due to higher measurement precision. Future work could also employ measures like quantitative T1 relaxometry or other proxies of intracortical myelin content (e.g. Tardif et al., 2015; Glasser et al., 2016; Paquola et al. 2018) for hippocampal alignment, but this is not possible in cross-modal work as in the various histology stains examined here.”

1. Finally, the application of the current work beyond the current dataset needs to be made more clear. From what I understand, the discussion says that using a multi-atlas approach with HippUnfold is unfeasible at this point. What kind of computational or technical developments need to take place in order for these labeled datasets to be used for this purpose? How can the current labeled datasets be used in other contexts?

The question of translation to other contexts, namely, in-vivo MRI, was also raised by Reviewer 1, and as such we decided to include an additional analysis to explore this question (Supplementary Materials 1: in-vivo MRI demonstration). Validation using ground-truth subfields is not plausible in MRI, and so we show only an indirect validation of intersubject alignment based on the sharpening of group-averaged features following better alignment using the present methods. We believe this new analysis significantly clarifies the applications we have in mind for this work. See the new Supplementary Section for details, and also a summary of this analysis in the Discussion section (p. 13):

“Ravikumar et al. (2021) recently performed flat mapping of the medial temporal lobe neocortex using a Laplace coordinate system as employed here, and showed sharpening of group-averaged images following deformable registration in unfolded space. This indirectly suggests better intersubject alignment. We perform a similar group-averaged sharpening analysis in Supplementary Materials 1: in-vivo demonstration. Though the gains in image sharpness observed here were modest, we note that current MRI resolution and automated segmentation methods allow for only imperfect hippocampal feature measures. We thus expect unfolded registration to grow in importance as MRI and segmentation methods continue to advance. “

Multi-atlas approaches are also presently possible, but we believe HippUnfold can apply unfolding and subfield definition with even higher validity. Unfolding of the hippocampus was previously possible in-vivo but still showed limited intersubject alignment. The present work validates a novel alignment method ex-vivo, and now additionally shows that this can be translated to better alignment even at the resolution of in-vivo imaging. We hope the above new Discussion paragraph also helps to clarify this.

1. A minor comment is that there are three panels (a,b,c) in Figure 4 but the figure legend does not describe them separately.

We thank the Reviewer, and added a Figure legend for parts B and C.