List of Papers Browse by Subject Areas Author List
Abstract
Reliable estimates of cellular composition and anatomical distribution in the human brain are essential for biologically plausible circuit models. In the hippocampus, existing reconstructions rely on low-resolution (LR) data without explicit cell-type-resolved annotations, limiting quantitative maps of excitatory neurons, inhibitory interneurons, and glial cells. Using newly released 1 um/px BigBrain sections of the right hippocampus, we present CALHippo, Cellular Annotation Library for the Hippocampus, a multiscale resource for cell-type-resolved reconstruction of the human CA complex. CALHippo includes the first expert-validated, cell-level annotated dataset spanning all Cornu Ammonis (CA1-CA4) subfields with explicit three-class labels, together with a lower-resolution mesoscale cellular point-cloud map. High-resolution (HR) cell instances are obtained through a human-in-the-loop pipeline combining foundation-model-based segmentation, iterative expert correction, and model ensembling, and are classified as excitatory neurons, inhibitory interneurons, or glial cells. To extend sparse HR annotations to the full volume, we project them into the 20 um/px LR BigBrain space and use the resulting class-specific supervision maps to train a UNet-based density estimation model. The predicted density maps enable slice-by-slice inference across the full CA complex and are sampled to generate a class-resolved mesoscale cellular point cloud. Code (https://github.com/AImageLab-zip/CALHippo-Framework) and dataset (https://ditto.ing.unimore.it/calhippo) are publicly released to support reproducibility.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0727_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AImageLab-zip/CALHippo-Framework
Link to the Dataset(s)
CALHippo: https://ditto.ing.unimore.it/calhippo
BibTex
@InProceedings{CasGio_CALHippo_MICCAI2026,
author = { Casari, Giovanni AND Candeloro, Ettore AND Gandolfi, Daniela AND Mapelli, Jonathan AND Bolelli, Federico AND Grana, Costantino},
title = { { CALHippo: Cell Segmentation for Neuronal Density Inference in the Human Hippocampus } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a multiscale pipeline for reconstructing cell-type-resolved hippocampal organization from histological BigBrain data. Its main contributions are as follows. (i) an expert-validated three-class cell annotation dataset for human hippocampal CA1-CA4 at 1 µm resolution. (ii) a human-in-the-loop annotation/segmentation pipeline built on existing cell segmentation models for this challenging domain. (iii) a compact U-Net-based density estimation model in the 20 µm BigBrain space. (iv) a 3D point-cloud reconstruction of excitatory, inhibitory, and glial cell distributions across the CA complex.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
A major strength of the paper is the creation of an expert-validated three-class hippocampal cell dataset at 1 µm resolution, which is both difficult and valuable. Another strength is the end-to-end multiscale pipeline that uses this dataset to produce a cell-type-resolved 3D point-cloud reconstruction across the CA regions. Overall, the paper provides a potentially useful resource for the community.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Contribution (ii) is less convincing as a methodological novelty. From the current presentation, it appears closer to a careful application of existing HITL segmentation workflows based on models such as Cellpose, rather than a genuinely new hippocampus-specific technical framework. If there are domain-specific annotation difficulties and corresponding technical solutions, these should be made much more explicit.
Although the Otsu-based baseline is recent in terms of publication date, it still seems to represent a relatively simple intensity-thresholding strategy. The paper would benefit from clarifying whether this baseline is included because it reflects the current standard practice in BigBrain hippocampal reconstruction, or merely as a lightweight reference method.
Table 4 is interesting as a potential biological output of the pipeline, but in its current form, it feels somewhat disconnected from the main validation. I fully appreciate that biological validation at this scale is extremely challenging; however, without stronger grounding or discussion, the CA-wise cell counts read more like an auxiliary result than a firmly supported conclusion.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
I provide the following minor comments in the hope of improving this paper. ・I wasn’t sure whether the “ppm” (e. g. , @1 ppm) in Fig. 1 refers to “parts per million”, μm, or something else. Maybe μm. ・Contribution (iv) is better understood as a sampled 3D cellular distribution derived from predicted density maps, rather than a direct geometric reconstruction. ・Since this is presented as an instance segmentation pipeline, the evaluation would benefit from clearer instance-level overlap analysis, e. g. , IoU-threshold-based AP/PQ or class-wise overlap statistics, beyond centroid-based matching alone. ・Fig. 1 would be more informative if it included representative zoomed-in examples of inhibitory and glial cells, not only the dominant excitatory population. Such examples would help readers better understand why these classes are more difficult for Cellpose-SAM and why human-in-the-loop correction and model merging are needed.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I view this paper positively in terms of its potential value to the community, especially as a resource contribution. However, I would like to point out several aspects that currently seem insufficient for a MICCAI paper.
First, the originality of the proposed pipeline is not yet sufficiently clear. At present, the paper mainly reads as a substantial expert-driven annotation effort on a difficult dataset. I acknowledge that this effort is valuable in itself, but I do not clearly see methodological novelty in the pipeline shown in Fig. 1.If there are dataset-specific technical innovations, design choices, or annotation strategies that are particular to hippocampal histology, these should be described much more explicitly.
Second, the comparison in Table 3 feels relatively weak, making the paper’s main contribution appear to be the effort of constructing supervised training data rather than the modeling approach itself. This is not necessarily a problem if the paper is framed primarily as a dataset/application contribution, but the current presentation should better align with that emphasis.
Third, I found Table 4 difficult to interpret. It is not yet clear how the reported CA-wise cell counts should be understood biologically. I appreciate that biological validation at this scale is very challenging. Still, the paper would benefit from stronger contextualization of these numbers, for example, through discussion of consistency with known hippocampal trends, limitations of the estimate, or, if feasible, comparison with external biological reference data. In other histology settings, one possible strategy is to compare adjacent sections stained with alternative protocols for specific cell populations, though I understand this may not be practical for the present dataset.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal substantially addresses my main concerns. In particular, the authors clarified the primary contribution as a dataset/resource and mesoscale atlas rather than a purely methodological advance, provided additional annotation-quality evidence, including inter-/intra-expert AER and final AER within expert variability, and added useful dataset construction/statistical details. The biological interpretation of Table 4 is also better grounded by comparing neuron-like counts with prior estimates. Some details still need to be clearly incorporated in the final manuscript, especially class-wise annotation reliability and consistency of the reported matched HR/LR data, but I think the concerns are fixable and mostly addressed by the rebuttal.
Review #2
- Please describe the contribution of the paper
The paper’s main contribution is a practical pipeline for deriving 3D, cell-type-specific density maps of the human hippocampus from BigBrain histology. The approach combines human-in-the-loop cell segmentation on 1 µm sections, cell-type classification based on foundation-model embeddings and morphological descriptors, and supervised transfer to 20 µm data via a U-Net for dense prediction. This design is meant to make cell-level information available at the scale of the full hippocampus, thereby supporting large-scale anatomical reconstruction and enabling comparison with established neuroanatomical cell-count estimates.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The main strengths of the paper are the novelty of the resulting resource, the relevance of the problem, and the care taken in the methodological design. Most notably, the work appears to provide a novel 3D cell-density map of the human hippocampus with cell-type-specific information and coverage across CA1–CA4 of the right hemisphere, extending beyond what has previously been available in the literature. The technical pipeline is carefully constructed, especially the segmentation stage, which combines expert labels, comparison of multiple models, an ensemble strategy, and analysis of annotator variability. The evaluation is also reasonably thorough, including sensible data splits, cross-validation, ablations for newly introduced loss terms, and comparison of the resulting maps against prior neuroanatomical estimates. In addition, the explicit classification of excitatory, inhibitory, and glial cells increases the biological relevance of the final atlas.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
A major concern is that several important preprocessing and dataset-construction steps are not described clearly enough to make the work fully reproducible. For example, the manuscript states that registration was used to align the 1 µm and 20 µm data, but my understanding is that the BigBrain data already include 1 µm images aligned to the 20 µm reconstruction. It is therefore unclear why an additional registration step was necessary, which data exactly were used, and whether the authors relied on existing aligned data or performed their own alignment. This should be clarified.
Related to this, the creation of the training datasets is insufficiently documented. Section 2.1 does not clearly state how many 1 µm sections were used, how the relevant regions of interest were selected, how patches were extracted, how many patches were generated, or how training, validation, and test splits were defined. While patching details are mentioned later in the density-estimation section, it is not clear whether the same settings also apply to the 1 µm segmentation and classification stages. Likewise, the training parameters for the instance-segmentation models are not provided. If default settings from prior work were used, this should be stated explicitly, and the dataset statistics should still be reported in full.
Another major weakness is that several parts of the method are introduced only at a high level and remain too vague in their actual implementation. The projection of cell densities from 1 µm to 20 µm space is one example: the role of variables such as d and beta is not sufficiently explained, making it difficult to follow how the target density maps are generated. Similarly, a custom Adaptive Thresholding Method (ATM) is introduced but neither properly described nor referenced. If ATM is a novel component, it requires a clear methodological explanation. If it is adapted from prior work, the relevant citation should be provided.
The manuscript is also unclear about the anatomical definition of the training and prediction regions. It is not sufficiently explained how the CA1–CA4 regions of interest were identified, which anatomical reference or annotation source was used, and how whole-slide images were cropped to the corresponding regions for training and inference. Because the final claims concern subfield-specific hippocampal density mapping, the precise derivation of these ROIs is an important part of the method and should be documented more carefully.
The cell-type-classification component is interesting, but several methodological choices are not justified well enough. The authors reduce the embedding dimensionality using PCA, but they do not report the explained variance or justify why n = 20 components were selected. It would also be important to understand the relative contribution of the handcrafted morphological features versus the foundation-model embeddings. In addition, the manuscript does not state whether feature normalization was performed before concatenation, which can materially affect downstream classification performance.
The density-estimation stage is also under-specified. The authors state that nnU-Net is used, but configuration details are missing. Given that nnU-Net is highly configuration-dependent, key settings should be reported so that the model setup can be interpreted and, ideally, reproduced. More generally, several notation issues further reduce clarity. For example, alpha is introduced in the text as the weighting factor for the pixel-wise loss term, but it does not appear in the corresponding equation. The notation should be made consistent throughout.
A further weakness is the limited presentation of qualitative results. The paper contains only a single overview figure, which is useful conceptually but does not allow the reader to assess the quality of the segmentation results, cell-type classification, or the resulting density maps in sufficient detail. Representative close-up examples would be important here, especially given the complexity of the pipeline. In addition, the figure lacks scale bars, which makes it difficult to interpret the spatial extent of the shown fields of view.
I also have concerns about the stability of the reported cell counts. According to the manuscript, centroids are obtained using a sampling-based procedure, specifically multinomial sampling with N trials, but N is not specified. Since final cell counts are derived from sampling, the paper should report variability across repeated sampling runs, for example in the form of standard deviations or confidence intervals, to demonstrate the robustness of the estimates.
I appreciate that the authors made their code available in a well organized structure. Nevertheless, I think the above details regarding the methodology and experimental design should also be addressed in the manuscript, as they fundamentally contribute the the understanding of the results.
Finally, the discussion section remains relatively limited and reads largely as a summary of the paper rather than a broader reflection on the implications and limitations of the approach. For example, it would be valuable to discuss to what extent the proposed pipeline could generalize beyond the hippocampus to other brain regions or even to whole-brain mapping. Related to this, the paper would benefit from a clearer discussion of the annotation effort required for the human-in-the-loop stages, since this is directly relevant for assessing scalability.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I lean weak accept. The main reasons are that the paper addresses a relevant problem, presents a novel result, and appears methodologically sound overall. In particular, the 3D cell-type-specific density mapping of the hippocampus seems to be a meaningful contribution that goes beyond prior literature.
My score is moderated by the fact that several technical details are currently insufficiently explained, which makes parts of the pipeline hard to assess and limits reproducibility. This includes dataset curation, ROI definition, some methodological components, and the sampling-based cell counting procedure. The paper would also benefit from more detailed qualitative results. Overall, I have a positive impression of the work, and if these points are clarified, I would support acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for providing detailed feedback on the reviews. Given the additional details provided, and taking my co-reviewers feedback into account, I recommend reject. While I still view the results of the paper as valuable, the methodology is not described in sufficient detail and design choices in the used method are not sufficiently evaluated and justified (e.g., emprically choosing the PCA dimension). I would recommend the authors to conduct a more in-depth analysis of how the different components influence the results and apply a data-driven approach (where possible) to determine method parameters (e.g., PCA explained variance ratio).
Review #3
- Please describe the contribution of the paper
The paper presents an expert-validated, cell-level-annotated dataset covering the full Cornu Ammonis (CA1–CA4) region with three-class labeling, and a UNet-based density estimation model operates in the 20 μm BigBrain space for estimation of cellular composition
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- To my knowledge, this is one of the first works to provide a cell-level annotated dataset covering the full CA1–CA4 regions, which could be a valuable resource for the community
- The paper proposes a density-estimation model for three cell types, including glial cells. This could be useful for more comprehensive cellular analysis.
- The work presents a 3D cellular reconstruction at mesoscale resolution based on the estimated densities, which could enable large-scale spatial analysis.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Annotation reliability. How many expert annotators were involved? The inter-annotator agreement is reported using AER = (FP+FN)/(TP+FN), with values >0.35 between annotator A and B. I understand this task can be difficult even for experts, but this still seems like a fairly high disagreement. 2.Glial annotation and density estimation. The global SQ is around 0.55, and glia recall is <0.8.This makes me wonder how reliable the glial annotations and density estimates are. Could you comment on how these performance levels affect the glia results, and whether any validation or uncertainty analysis was done? 3.CA4 cell count. In Table 4, CA4 (hilus) seems to have a much higher total cell count than CA2 and CA3, which may be inconsistent with previous studies. For example, West and Gundersen (1990) [1] report mean neuron counts of approximately 16×10⁶ in CA1, 2.7×10⁶ in CA3–CA2, and 2.0×10⁶ in the hilus per hippocampus. I understand those are neuron counts, not total cells, but could you explain this difference? 4.Evaluation metric. Can you explain why you chose NAE for density estimation? Metrics like MSE or RMSE are more commonly used, so it would be helpful to know if your results are consistent under those as well. 5.Table 3 clarity. Table 3 is a bit confusing. “Ours CA1” and “Ours All CA” seem to be evaluated on different datasets, so comparing them directly (with bold/underline) may not be appropriate. It might be clearer to separate them.
[1] West MJ, Gundersen HJ. Unbiased stereological estimation of the number of neurons in the human hippocampus. J Comp Neurol. 1990 Jun 1;296(1):1-22.doi: 10.1002/cne.902960102.PMID: 2358525.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
My main concern is the reliability of the dataset annotations. Given the relatively high inter-annotator disagreement (AER > 0.35), I am unsure how reliable the labels are for training and evaluation. Could the authors provide additional analysis to better characterize annotation quality?
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the paper presents potentially valuable contributions, my main concern is the reliability of the dataset annotations. The relatively high inter-annotator disagreement raises questions about label quality, which directly affects both the segmentation results and the downstream biological conclusions.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I thank the authors for their rebuttal. My major concern regarding annotation quality is partially addressed by their explanation of expert variability, the reported final AER of 0.25, and the statement that all annotations were expert-reviewed. I also appreciate the motivation for ATM.
However, my remaining concern is that the submitted results do not clearly isolate the specific contribution of ATM to the final annotation quality. The rebuttal states that adding ATM improves glial recall measurably, but Table 1 appears to show ATM alone and the final merged result, rather than a direct comparison of the merged pipeline with versus without ATM. Therefore, from the existing results, I cannot fully assess whether the reported improvement is due to ATM, the merging procedure, the other segmentation models, or their combination.
Since the rest of the work appears solid, and this is my only remaining concern, I am inclined to support acceptance. I strongly recommend that the camera-ready version clarify the role and implementation details of ATM, the definition of “Merged-models,” and the interpretation of Table 1.Providing a quantitative ablation study on the contribution of ATM to the merged pipeline would further strengthen the paper.
Author Feedback
The reviewers recognized our work as a valuable resource for the community (R1,R3), highlighting the expert-validated cell-level hippocampal dataset spanning CA1-4 (All), appreciating its potential for comprehensive cellular analysis (R2,R3), as well as the careful multiscale pipeline, sound evaluation, released code, and biologically relevant atlas (R1,R2).
Contribution and scope [R1,R2] This paper’s primary contribution is the first expert-validated three-class cellular dataset for human CA1-4 in Merker-stained BigBrain tissue, together with a class-resolved mesoscale atlas of the CA complex. Methodologically, the key contribution is the HR-to-LR transfer strategy: a few annotated HR images are used to supervise cell prediction across complete LR image sets. Together with HR segmentation ensemble and instance classification, this pipeline is reusable for cell-resolved mapping in other brain regions.
Annotation quality and ATM [All]. Two experts independently annotated two WSIs, yielding an inter-AER of 0.35; re-annotation by the same experts gave an intra-AER of ~0.3, supporting intrinsic tissue variability. Both values and the low prediction SQ are driven primarily by glial cells, which exhibit poorly defined boundaries in Merker silver-stained tissue and are frequently misidentified as tissue artifacts. This ambiguity motivated ATM, which combines adaptive thresholding with filters grounded in glial cytoarchitectonic properties (color, eccentricity, and area) to better capture glial instances. Adding ATM to the ensemble improves glial recall measurably, with the final pipeline achieving AER 0.25, within the range of expert variability. All annotations were expert-reviewed.
Biological interpretation [R1,R3] The concern about CA4 comes from comparing our total-cell estimates with neuron-only stereology. Restricting Table 4 to neuron-like classes (excitatory+inhibitory), our counts closely match prior estimates cited by R3 [1]: CA1 15.2M vs 16M, CA2+CA3 2.8M vs 2.7M, and CA4 2.3M vs 2.0M, supporting biological plausibility. We also observed a ~1:1 glia/neuron-like ratio, consistent with [2] findings. We will revise Table 4, cite these works, and clarify that Gandolfi et al. [2] is included as the only available domain baseline on BigBrain hippocampus data.
Dataset construction and registration [R1,R2] CA1-4 ROIs were taken from the BigBrain hippocampal subfield reference [3*]. To build the LR training data, we first mapped HR coordinates with the provided affine transforms to crop the corresponding LR regions, then applied deformable alignment to correct residual HR-LR tissue mismatch to prevent misaligned density targets. From the available 24 matched HR/LR pairs, we extracted 505 patches and split them 80/20 by WSI ID to avoid leakage.
Method and minor fixes [R1] Instance evaluation is not centroid-only: centroid matching defines TP/FP/FN, but Dice, HD95, and SQ, also reported class-wise, assess overlap and boundary quality; we will additionally report IoU-threshold-based PQ. [R2] PCA with n=20 was chosen empirically to reduce embeddings so they do not dominate morphology/context cues, and both feature groups are normalized. For density-map generation, each cell centroid is projected to LR, then replaced by a Gaussian kernel whose variance is scaled by other local cell distances, controlled by d and fixed beta as in [18]; this smooths sparse regions while preserving cell counts. For the sampled point cloud, N is the predicted region density count; repeated sampling changes only the coordinates, not the expected counts. [R3] NAE was chosen to compare relative count errors across imbalanced classes, while SSIM measures spatial agreement. We will clarify the contribution focus, add missing segmentation/nnU-Net settings, ATM details, and improve tables, figures, notation, and cell visualizations.
[1] doi:10.1002/cne.902960102 [2] doi:10.1002/cne.21974 [3*] doi:10.1016/j.neuroimage.2019.116328
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The review is mixed: one week accept and two week reject. The concerns are serious but fixable.
Points to address:
- Annotation quality
- dataset/statistics details
- dataset vs method contribution
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper is strong as a dataset, atlas, and application/resource paper rather than as a novel segmentation-method paper. The methodological pipeline appears to rely largely on existing components, but the expert annotation effort, multiscale integration, and cell-type-specific hippocampal density mapping are potentially useful to the community.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The final revised recommendations are two Accepts (R1, R3) and one Reject (R2). However, it is worth noting that R2 changed their recommendation from Weak Accept to Reject after seeing the earlier recommendations from the other reviewers, both of whom later changed their recommendations to Accept. All reviewers believe that the paper can be further improved for the camera-ready version without significant additional work.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a multiscale pipeline for reconstructing cell-type-resolved neuronal density maps in the human hippocampus from BigBrain histology. Its main contributions are an expert-validated three-class hippocampal cell annotation dataset, a human-in-the-loop annotation/segmentation pipeline, a compact density estimation model in the 20 µm BigBrain space, and a 3D point-cloud reconstruction of excitatory, inhibitory, and glial cell distributions across the CA complex.
The initial reviews were mixed and overall borderline, with scores of 3 / 4 / 3.Reviewers appreciated the potential value of the work as a resource contribution, especially the expert-validated hippocampal cell dataset and the possibility of producing cell-type-specific density maps at scale. However, the reviewers raised several concerns: limited methodological novelty, unclear positioning of the contribution as a dataset/resource versus a new method, insufficient details on dataset construction and registration, uncertainty about annotation quality and inter-annotator variability, limited instance-level segmentation analysis, and questions about the biological interpretation of the reported CA-wise cell counts.
The rebuttal addressed several of these concerns. The authors clarified that the main contribution should be understood primarily as an expert-validated dataset/resource and mesoscopic atlas, rather than as a purely methodological advance. They provided additional information on annotation quality, including inter- and intra-expert AER, the final AER being within expert variability, and the fact that annotations were expert-reviewed. They also clarified the motivation for the adaptive thresholding module, the HR-to-LR transfer strategy, the construction of matched HR/LR data, and the biological interpretation of the CA-wise counts by comparing neuron-like counts with prior estimates.
Some concerns remain. In particular, the final manuscript should more clearly describe the implementation and contribution of the ATM component, class-wise annotation reliability, the role of merged models, registration details, and the consistency of the matched HR/LR data. The ablation of ATM versus the full merged pipeline remains incomplete, and the biological interpretation should still be carefully framed as an atlas/resource contribution rather than a fully validated biological conclusion.
After rebuttal, the reviewer opinions are A / R / A. Reviewers 1 and 3 moved from weak reject to accept, mainly because the rebuttal clarified the contribution, provided annotation-quality evidence, and better grounded the biological interpretation. Reviewer 2 moved to reject, emphasizing that some methodological details and design choices remain insufficiently justified. I place greater weight on the two positive post-rebuttal assessments and on the value of the dataset/resource contribution, while acknowledging that the methodology itself is not the primary strength.
I recommend Accept. The paper is not a strong methodological contribution, but it provides a useful and difficult-to-obtain resource for the community, and the rebuttal sufficiently mitigated the most serious concerns regarding annotation quality and contribution framing. The remaining issues should be addressed in the camera-ready version, especially the clarification of ATM, merged-model details, annotation reliability, and the interpretation of the density maps.
