Abstract

The hippocampus in the brain performs a pivotal role for memory formation, spa-tial navigation, and emotional regulation. Its volume and morphology are known to change with the progression of neurodegenerative diseases such as Alz-heimer’s disease. Hence, hippocampal atrophy serves as a key biomarker for ear-ly diagnosis and monitoring of such diseases. Whereas MRI has been predomi-nantly employed in that regard due to its excellent soft-tissue contrast, CT-based segmentation of the structure has been relatively far less explored because the modality results in ambiguous boundaries between brain subregions. This study aims to address this technical challenge, achieving accurate segmentation of the hippocampus on CT images. To this end, we develop a deep learning model, termed ‘Hippocampus Dual Decoder Network (HDD-Net)’, characterized by the following four major components: 1) parallel, dual decoders that segment the hippocampal region and its boundaries, respectively, 2) a single, shared encoder in which features combined across multiple blocks are refined via attention, 3) a feature fusion module (FFM) that performs inter-decoder featural supplements, and 4) a cross loss to jointly optimize segmentation and edge predictions. HDD-Net was validated using both internal and external datasets, with its performance assessed using Dice similarity coefficient (DSC) and intersection-over-union (IoU). Our model yielded DSC = 0.823 ± 0.03 and IoU = 0.701 ± 0.04, and DSC = 0.759 ± 0.07 and IoU = 0.617 ± 0.09 for internal and external test da-tasets, respectively, outperforming seven other SOTA methods. Furthermore, volumetric analysis revealed a good agreement between MRI- and CT-derived hippocampal masks. Our findings suggest feasibility of CT-based hippocampal segmentation via HDD-Net, as a cost-effective alternative to MRI. The implementation of HDD-Net is available at https://github.com/sonwonjun103/HDD_Net.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3287_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/sonwonjun103/HDD_Net

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SonWon_CTBased_MICCAI2025,
        author = { Son, Wonjun and Lee, Ji Young and Ahn, Sung Jun and Lee, Hyunyeol},
        title = { { CT-Based Hippocampus Segmentation with Dual-Decoder Network (HDD-Net) } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {142 -- 152}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The premise of this paper is that hippocampal atrophy, as identified by magnetic resonance imaging (MRI), has been a timely and relevant biomarker for Alzheimer’s disease. This study explores CT-based segmentation of the hippocampus as an alternative approach. The primary innovation is the development of a deep learning model termed the Hippocampal Dual Decoder Network (HDD-Net).

  • 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.
    • Ample details are provided about the architecture of the network, including the auxiliary structure on the side of the main pathway.
  • 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 substantial limitation of the proposed approach is that the model was trained only on neurologically healthy individuals. Given the limited soft tissue contrast in CT, the model may be learning to localize the hippocampus based on its typical anatomical position within a healthy brain, rather than identifying hippocampal anatomy directly. To assess generalizability, it is essential to include subjects with both normal and abnormal anatomical variation. Without this, the concern remains that the model is detecting the location of the hippocampus rather than the structure itself. Consideration of atlas-based positioning or comparison to individualized anatomy would help clarify this distinction.
    • It is unclear why cross-validation was not used, particularly given the small dataset.
    • While substantial improvements in performance are reported, the Bland-Altman analysis shows little dependence of error on volume, raising further questions about what is actually being learned.
  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, this is an interesting attempt to approach hippocampal segmentation using CT, in contrast to traditional MRI-based methods. However, the core limitation lies in the lack of exploration into the identifiability of hippocampal variants—especially since the dataset includes only healthy controls and excludes subjects with pathological conditions or correlations with disease severity. Additionally, the paper offers limited discussion comparing the achievable performance bounds in CT with those of MRI-based segmentation methods. It is concerning that cross-validation was not employed, especially considering the small size of the dataset. Given the range of comparative methods, it is difficult to assess whether the contribution lies in more efficient learning from limited data or in a fundamentally new strategy for identifying structures in low-contrast CT images. Finally, the assumption that hippocampal gray/white matter boundaries can be identified via edge detection in CT is questionable. The example image shown demonstrates little brain tissue contrast between gray and white matter, which aligns with existing literature. Therefore, the explicit focus on integrating edge effects along with volumetric effects appears less well-justified than implied.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    The paper proposes HDD-Net, a dual-decoder network for hippocampus segmentation from CT images, addressing the challenge of low soft-tissue contrast in CT. It introduces a Feature Fusion Module (FFM) to enable interaction between segmentation and edge decoders. A cross loss function is designed to jointly optimize volumetric and boundary predictions. The model achieves strong performance on both internal and external datasets, outperforming several U-Net and transformer-based baselines.

  • 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.

    (1) Enables hippocampus analysis using only CT scans, offering a practical alternative to MRI in low-resource settings. (2) Dual-decoder architecture effectively captures both regional structures and boundary information. (3) Demonstrates strong and consistent performance on both internal and external datasets. (4) Includes comprehensive comparisons with both CNN-based and transformer-based segmentation models.

  • 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) Methodological novelty is moderate; core components are adapted from existing techniques. (2) No ablation study to verify the contribution of each module (e.g., FFM, cross loss). (3) Limited dataset diversity; no public datasets used for evaluation. (4) No analysis of model efficiency or failure cases.

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The main reason I rated this paper as a weak accept is based on its significant contribution: enabling hippocampus analysis using only CT scans, which provides a practical and cost-effective alternative to MRI in low-resource settings. This application is especially valuable in clinical environments with limited access to MRI.

    While there are some weaknesses, such as moderate methodological novelty, lack of an ablation study, limited dataset diversity, and no analysis of failure cases, I believe the practical impact of this work outweighs these concerns. The approach demonstrates strong performance and offers a potentially important solution for real-world clinical scenarios.

  • 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 rated this paper as weak accept due to its significant contribution: enabling hippocampus analysis using only CT scans, offering a practical, cost-effective alternative to MRI in low-resource settings, especially valuable in clinical environments with limited MRI access.



Review #3

  • Please describe the contribution of the paper

    The paper advances CT-based hippocampus segmentation based on a dual-decoder approach and channel attention.

  • 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 description of the approach is detailed, the paper is well organized and understandable. The methodology is not terribly innovative, but it seems to be w.r.t. the specific domain of application. Reported scores indicate considerable improvements over previous methods. While CT is less common in cases where the interest is directed specifically at the hippocampus, it could still be interesting when MRI is not indicated or to enable incidental findings.

  • 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.

    While we do have a comparison with other methods we do not have ablation studies in this word, meaning that it’s harder to know how important each of these components is individually.

    This work could benefit from including additionally HD95 as a metric.

    There are some minor mistakes in the language but they do not compromise clarity.

  • 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While this work presents some small weaknesses I do not believe these are enough of a problem to warrant a rejection.

  • Reviewer confidence

    Very confident (4)

  • [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 was already of the opinion of accepting this paper, the follow-up from the authors and other reviewers did not lead me to a change of mind.




Author Feedback

We thank all Reviewers for lending their time and expertise to the review of our paper. Below, we provide our responses to three major critiques raised across the Reviewers.

  1. Unclear what the network learns – location or structure (Reviewer 2 and Meta-Reviewer 1)

Response: Although it is difficult to fully understand how the model extracts hippocampal areas from low-contrast CT images, we conjecture that the model would learn not only hippocampal location, but its structural details. This is supported by Fig. 3 and evaluation scores, which demonstrate a high level of structural fidelity in the CT segmentations in close agreement with the MRI-derived annotations. We believe that this level of anatomical alignment would be difficult to achieve if the model just learned hippocampal locations. Our findings including the ablation study (see our response to #3 below) indicate that the model, with edge constraints, effectively extracts meaningful structural information by leveraging volumetric contextual cues in CT images.

  1. Limited dataset diversity / no public datasets used for evaluation / generalizability to non-healthy subjects (Reviewers 1 and 2, and Meta-Reviewer 1)

Response: To our best knowledge, there is no public sites providing paired CT-MR images in healthy subjects. Furthermore, it is also very challenging to collect paired CT-MRI datasets in patients with hippocampal damages or deformations. Considering these, we initially defined this work as a feasibility study, and confined its scope to neurologically healthy subjects whose datasets had been collected in two local institutions. The model’s generalizability towards diverse populations including patients remains to be explored in future studies. Nonetheless, we expect our model, with its current network parameters, would be applicable to patients with Alzheimer’s disease, where hippocampal atrophy develops gradually with its anatomical configuration largely preserved (References given below). We hope the discussion above is acceptable to the Reviewers.

References: Bremner, J. D. et al. “Hippocampal volume reduction in major depression.” American Journal of Psychiatry 157.1 (2000): 115-118. Henneman, W. J. P., et al. “Hippocampal atrophy rates in Alzheimer disease: added value over whole brain volume measures.” Neurology 72.11 (2009): 999-1007.

  1. No ablation study (Reviewers 1 and 3, and Meta-Reviewer 1) Response: We thank the reviewers for raising this important issue. We had actually performed an ablation study, but did not provide the results because of the page limit. We fully agree with the reviewers that the ablation study should have been included, and apologize for inappropriate organization of the manuscript. Briefly, our ablation study shows that the full model (with the edge decoder, feature fusion module, and cross loss all included) achieves the best performance. The model demonstrates progressive improvements as each component is added sequentially. Furthermore, we observe that the cross loss contributes more to our model’s performance than the feature fusion module does. These findings underscore the value of incorporating edge constraints in segmentation tasks from low-contrast images. We hope to share these results in greater detail at the conference.

  2. Minor changes (such as linguistic errors) will be made in the final version.

  3. Code availability: We will make our model available upon acceptance.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    This work has received mixed reviews. Authors are invited to organize comments and respond via rebuttal.

    Three major concerns that I see are:

    1. it is unclear what the network learns, it might be solely related to hippocampus location, since there are few visual hints available of the hippocampus on CT data
    2. it is unclear if the proposed method generalizes to non-healthy subjects, which is where the clinical relevance lies
    3. there is no ablation study of the proposed components, especially for the edge based loss component this is crucial (since there is limited to no contrast of hippocampus in CT, as can be seen in Fig. 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    The methodological novelty is very limited. The model training, experiments and comparisons are unconvincing. The underlying assumption that hippocampal gray/white matter boundaries can be identified via edge detection in CT is questionable.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    After reading the paper, the original reviewer comments, the author’s feedback and the final comments post-rebuttal, I would like to recommend acceptance for this paper. The main reason for that decision is the novelty of the idea to try and segment the hippocampus in CT images. As far as I am aware, this has not been tried before and it deserves to be presented. I have some concerns about the methodology and whether the model actually learns information related to edges (as raised by reviewer #2) and I am not entirely convinced by the rebuttal. However, I still think that the large method comparison has merit and that the fact that the authors proved that the hippocampus can be segmented with CNN architecture is an important one.

    Nonetheless, I also think that while an interesting result, the authors would benefit from further exploring what is the model actually learning, test the results on abnormal cases and expand on their work on a future publication, regardless of the final decision for this paper.



back to top