Article activity feed

  1. Evaluation Summary:

    This paper presents an exciting new automated package to investigate the hippocampal organization in new ways. As such, this package will be equally interesting for the fundamental basic and clinical neurosciences.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #2 agreed to share their name with the authors.)

  2. Reviewer #1 (Public Review):

    I'd first like to congratulate the authors for their impressive work, nicely building on their earlier work, and now representing an automated package which will enable the research community to use their approach widely. It is exciting to see how their approach opens new perspectives to investigate hippocampal organization and I am personally looking forward to all the work that will follow from this.

    One important strength is that their overall approach allows for analyses that have not been possible before in this way. Their topological alignment allows to visualize data from other domains on the folded or unfolded surfaces and enables analyses regarding thickness and curvature. While this was possible using their own earlier work already before, they added one more piece so that the full pipeline can be applied automatically from beginning to end. They had the end user in mind during their development and thus made the pipeline available as BIDS App, provide their pipeline in containers and added a quality assessment step that allows others to flag suspicious results.

    Another strength is that once hippocampal anatomy has been unfolded, their approach allows to label hippocampal subregions based on their own earlier work using the Big Brain 3D histology. While there are already approaches for automated subfield segmentation, one big advantage of their approach is the automated labeling of the hippocampal tail. This is very complicated and difficult to do even with manual segmentation due to the complex appearing anatomy because of the bending of the posterior hippocampus.

    Yet another strength is their careful validation approach regarding their pipeline. They compare their own approach against other popular segmentation tools in the field, ASHS and Freesurfer. They find that their approach is more similar to other histology work, in particular in the hippocampal head, and report that their approach compares very well to their own manual segmentations. Further, they explore how well their approach generalizes to an aging sample as well as to a dataset with different and non-isotropic resolution.

    However, one potential issue relates to the robustness of their approach regarding populations with different hippocampi due to age-related changes or disease. The authors have shown in this work, that their approach can produce solutions in an ageing dataset and a dataset with different resolution that apparently seemed to be correct. However, for example, how would a subject with pronounced thinning or overall volumetric changes in CA1 look like. This makes me wonder whether their approach would be sensitive, for example, to specific changes in one subfield but not others.

    This relates also more broadly to the applied validation measures. For example, the authors state that their approach seems more favorable compared to other approaches because they see continuity of subfields along the long axis of the hippocampus. While I agree that based on the anatomy, subfields should be continuous, image quality does not always allow for segmentation at every voxel even when done manually. I'm wondering in general whether it helps to do so anyways by enforcing subfield labels based on a strong prior (subfield labels defined on Big Brain 3D histology), or whether it would be advantageous to rather not label a subfield in these cases.

    Taken together my point is whether the approach presented here has the risk that it is too dependent on the prior and imposes the same subfield label information on every subject which would produce correct-looking results but would not necessarily be valid. While I appreciate the authors analysis of a dataset from an aged population as well as one with a different resolution, I do not think that these are enough yet to show validity in this respect.

  3. Reviewer #2 (Public Review):

    In this work, a new software package for hippocampus segmentation, unfolding, and subfield labeling is presented. The method is packaged into a BIDS app, in order to use it with standard 3T MRI, but can also accommodate more advanced 7T imaging, and the different steps can be performed independently, for instance when processing post-mortem histology data or incorporating manual delineations.

    The unfolding procedure defines a flat map of the hippocampus, which may be particularly useful for visualization, similarly to the flat maps or partially inflated maps previously built for the cerebral and cerebellar cortices.

    The method is evaluated on high resolution data from the HCP (3T) as well as ultra-high resolution 7T data often acquired for hippocampus morphometry. Comparisons are made with the two other leading software packages for hippocampus segmentation and subfield labeling, showing that the proposed method is more complete, including both head and tail, and arguing that it preserves better the topology of unfolded subfields.

    The software package is distributed in open source, including detailed documentation but unfortunately no actual test data. Multiple outputs options make the software tool very flexible and potentially useful in a large number of data sets.

    Overall, the methods employed are sound and appear both robust and elegant. However, there are a few potential limitations and confusions with regard to the method that needs to be addressed.

    First, the different methods used in the toolbox are not fully described, sending the reader to collate information from multiple sources in order to understand what algorithms are run in the processing pipeline. This article would be the opportunity to summarize the different methods used in sufficient detail, especially as modifications and adjustments have been made from the original works.

    Second, there is a general confusion about topology and topology preservation throughout the paper. The voxel domain and its relatively coarse resolution with regard to the hippocampal formation and its subfields can hardly allow to preserve (digital) topology, and the proposed method in fact does not guarantee it, like the other ones. What it does preserve is the relative arrangement of the subfields in the unfolded plane, which is fixed to match the map obtained from labeling a single post-mortem data set, BigBrain. Comparing the capabilities of other methods to preserve this arrangement is somewhat unfair, and not really relevant. The important topological feature that is actually preserved (or better, estimated) is the hippocampal folding structure, which is conserved independently of the variation in digitation. Separating the two questions (of mapping a folded surface representation and of correctly placing subfields label on it) is important, and somewhat confused in the paper.

    Third, there is an implicit assumption made that unfolded hippocampi should match, which is not tested in experiments, and is arguable: in the same way that cortical maps unfolded into perfect spheres still need to be aligned for establishing proper correspondences (see e.g. the Spherical Demons algorithm in FreeSurfer), hippocampal maps require non-linear alignment in the unfolded plane, unless the unfolding procedure takes into account additional features such as the location of subfields, stable morphometric landmarks, and/or MRI contrasts. While this problem is likely less pronounced here because of the generally less variable shape of the hippocampus, it should be fully acknowledged.

    Fourth, the key segmentation step to obtain the unfolded representation is performed by a U-Net. While such artificial neural networks have generally excellent performance with the type of data they have been trained with, they are often challenged to generalize across different contrasts. The authors provide some results showing a limited yet systematic decrease in performance (Fig.5B), but a discussion of the limitations and important preprocessing steps recommended would be useful for the general user.

  4. Reviewer #3 (Public Review):

    The work proposed by DeKraker and colleagues is the extension of a long-standing research program by the senior author's laboratory to using histologically defined methods to inform surface-based measures that better conform to the unique "rolling" anatomy of the human hippocampus and its subfields. The group has previously the hippocampal "unfolding" technique as a means to capture different metrics and their variation along a two-dimensional surface manifold. The current work improves the implementation of this software using a 'U-Net' deep convolutional neural network as means to successfully identify variation in the subfields that cannot be seen using standard techniques (like atlas or multi-atlas based segmentation techniques. In terms of novelty, and given the previous work from this group its unclear what the novelty of this particular work is. Is it simply the integration of the U-Net into this application? Is it demonstrating generalizability or superior performance to previous pipelines? There are no putative demonstrations of the applicability of the pipeline as well.

    In general, the paper is well written, but there are multiple areas that I have some issues with following the logical flow of what is being proposed. For example, the paper begins by demonstrating multiple metrics that are projected onto the hippocampal flatmap that includes thickness, myelin, curvature, gyrification, etc. It is unclear as to what information the authors want to convey here. This is the first mention of many of these multiple metrics as well and therefore their relevance is ultimately not extremely clear. As a result, it is hard to support their claim that "differences in morphological and quantitative features can be seen across the hippocampus, particularly across the subfields" as the goals of this particular figure are not at all clear.

    Line 147: It is not totally accurate to state that ASHS makes use of multi-atlas registration as it also uses AdaBoost to correct for segmentation inaccuracies.

    For the FreeSufer and ASHS comparisons - is it possible to provide some quantification of errors or anything like that? I think it would be helpful to quantify the differences in a more accurate manner. If this is in a previous publication and I missed it, it could be useful to reiterate here. The qualitative difference is nice - but there is room to compare them more quantitatively to one another.

    For the validation of the U-NET, details on the manual segmentation protocol, who did it, and its reliability are crucial. Training/testing paradigms would be helpful here. So would Bland-Altmann plots. I think in general the validation of these segmentations is quite poor - so more metrics that demonstrate the segmentation beyond dice overlaps would be helpful.

    It is unclear how generalizable the method is outside of HCP acquisitions.