Abstract

Hypoxic–ischemic encephalopathy (HIE) is a major cause of neonatal death and lifelong neuro-disability, with limited treatment options beyond therapeutic hypothermia within 6 hours of birth. In translational fetal sheep models, microglia are central to inflammatory injury cascades and are a key immunohistochemical readout for evaluating therapeutic efficacy. Still, their quantification and morphological grading in histological images remains largely manual, limiting throughput, reproducibility, and mechanistic insights. We propose Microglia-RTS, a two-stage pipeline for microglia instance segmentation in high-resolution histology for preclinical drug screening. The first detection stage adapts a modified RT-DETR with a pretrained DINOv3 ConvNeXt backbone that incorporates a novel adaptive query selection module to tune the number of object queries per image based on local cell density. The second segmentation stage fine-tunes the Segment Anything Model (SAM2) to produce cell masks from detector proposals. Patch-based inference with overlap-aware stitching enables high-resolution processing of images, preserving intricate cellular details. We evaluate Microglia-RTS on a new fully annotated dataset comprising ~7,000 cell instances, classified into ameboid, intermediate or ramified, across 45-Iba1 stained hypoxic–ischemic fetal sheep brain images from three white matter regions (IGWM1, IGWM2 and PVWM). The modified detector achieved an F1-Detection score of 0.55±0.01 during validation, considerably surpassing the baseline model, and the fine-tuned SAM2 achieved an average Dice-score of 0.86±0.01. Across 5-fold cross-validation on whole images, Microglia-RTS achieved a mean panoptic quality of 0.48±0.02, supporting scalable microglia analysis for pre-clinical HIE drug discovery by reducing observer bias and accelerating quantitative endpoint extraction.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5949_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LooCal_MicrogliaRTS_MICCAI2026,
        author = { Loomes, Callan AND Davidson, Joanne AND Tian, Grace AND Gunn, Alistair AND Bennet, Laura AND Abbasi, Hamid},
        title = { { Microglia-RTS: Instance Segmentation of Microglia in Hypoxic Ischemic Fetal Sheep Brain Histology } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposed a two-stage pipeline for microglia instance segmentation in high-resolution histology for preclinical drug screening

  • 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 paper proposed a very traditional deep learning method for microglia instance segmentation. The paper is well structured.

  • 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.

    It is more like an experiment report instead of an academic paper. There is no novelty in this paper.

  • Please rate the clarity and organization of this paper

    Poor

  • 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.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    it is more like an experiment report instead of an academic paper

  • Reviewer confidence

    Very confident (4)

  • [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

    1.The authors have created a new, annotated dataset of Iba1-stained microglia over three morphological classes (i.e. ameboid, intermediate, and ramified) in three white-matter regions (i.e. IGWM1, IGWM2, and PVWM) in hypoxic-ischemic sheep brains. The dataset comprises of approximately 7000 images, and uses the three morphological classes instead of four to six alike in other studies in order to preserve biological interpretability and simplify annotation. 2.An instance-segmentation framework was proposed, whereby two models (RT-DETR and SAM2) are combined, along with adapted pretraining, novel modifications to the SAM2 architecture, and an optimised training methodology. The modified RT-DETR surpassed the baseline model with regards to F1-detection score.

  • 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.The paper presents a new, annotated dataset of Iba1-stained microglia whereby there are three morphological classes instead of four to six. This is to preserve biological interpretability, making it less complex and simplifying annotation. 2.The authors combine an RT-DETR and SAM2, in which the RT-DETR contains a novel adaptive query branch which predicts the query count at runtime using an MLP. 3.Overall, the framework produces fast detection of HI in fetal sheep which is beneficial in such a time crucial medical assessment.

  • 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.Some metrics have not been referenced, i.e. Dice and Dice loss. 2.Some details are missing from the paper, for example, P and G in Equation 3 should be defined to guide the reader through the workings of the equations. 3.F1 detection scores are seemingly low, leaving room for improvement, however, this could possibly be the nicheness of the dataset. 4.An older version of RT-DETR (version 2) is being used, when there is a more SOTA version (version 3) that has been out for over a year. 5.There is no 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    1.Please provide references for metrics. I.e. Dice: Dice (1945) Measures of the Amount of Ecologic Association Between Species. Dice loss: Milletari et al. (2016) V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. 2.Please define P and G in Equation 3 to guide the reader through the workings of the equations. 3.Consider using a more SOTA version of RT-DETR (version 3). 4.Please provide a conclusion or rename the discussion to “Discussion & Conclusion”. 5.Ensure you include your Acknowledgements if accepted. 6.Ideally provide links to code and data for reproducibility.

  • 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?

    The paper is well written and the authors have created a new, annotated dataset of Iba1-stained microglia whereby there are three morphological classes instead of four to six. This is to preserve biological interpretability, making it less complex and simplifying annotation. The authors combine two models (RT-DETRv2 and SAM2), whereby RT-DETRv2 isn’t the most SOTA model. The dataset itself is innovative and somewhat should be published as a MICCAI paper, however, the model itself and the slight change to the RT-DETR model for novelty purposes is not quite innovative enough. The figures are very well presented and the models have been very well evaluated. However, there is limited reproducibility as the code nor the dataset have been linked by the authors.

  • Reviewer confidence

    Very confident (4)

  • [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 #3

  • Please describe the contribution of the paper

    The paper proposes a two-stage pipeline for microglia segmentation (ameboid, intermediate, and ramified). The pipeline utilizes a modified RT-DETRv2 for detection and a SAM2-small for segmentation. For RT-DETERv2, the paper utilizes an adaptive query MLP module to predict query count instead of relying on the “objectless” scoring. Figure 4B shows the model performance using fixed query counts vs their adaptive query configuration.

  • 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 paper, especially the introduction section, is very well written. The methods section is described in detail and seems to be well thought out (e.g. three-phase unfreezing strategy, random patch sampling with a greedy selection strategy, etc.) The paper demonstrates the effectiveness of their adaptive query configuration in Figure 4 and Table 2.Qualitative results from different cases (healthy and damaged) are included in Figure 5.

  • 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.

    There is no clear major weakness that stands out to me. While not necessary, the paper could expand and evaluate runtime and efficiency for fixed query counts vs their adaptive query configuration. To nitpick, the paper could explain how their MLP is better than simple density estimation.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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.

    (6) Strong Accept — must be accepted due to excellence

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

    The paper is very well written and provides controlled experiments for the main design choices (backbone and query formulation). The proposed adaptive query configuration is simple yet effective. Quantitative and qualitative results are illustrative. Overall, the method is well carried out and validated. The paper provides great clinical context, demonstrating practical value.

  • Reviewer confidence

    Somewhat confident (2)

  • [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



Author Feedback

We would like to thank the meta-reviewer and the reviewers for their thorough and constructive feedback. In revising the paper, we have strengthened the scientific and structural aspects of the work, and we believe the camera-ready version meets the standards and expectations set out in the reviews. Relevant changes are addressed individually below:

Meta-Reviewer Comments: We would like to thank the meta-reviewer for their positive recommendation and for clearly identifying the two camera-ready conditions, which we have addressed.

A new “Data Availability” paragraph has been added to the manuscript confirming our data policy as follows: “The dataset is not publicly available due to ethical restrictions; data and code may be shared upon reasonable request, subject to formal agreements.”. While we must limit the data at this point, there is intention to possibly release a full extended dataset alongside a forthcoming peer-reviewed journal article, subject to ethical updates.

The version of RT-DETR is now explicitly justified in the Methods section: RT-DETRv2 was chosen over the newer RT-DETRv3 for reproducibility and portability, as it had official support in the HuggingFace Transformers Library at project inception, while RT-DETRv3 was only available through a standalone research repository. The Future Directions paragraph notes that since detection remains the main performance bottleneck, RT-DETRv3 is a natural candidate for further investigation.

The detection values are now contextualised in the Discussion in terms of task-intrinsic difficulty: the relatively low F1 of 0.55 likely reflects the combined challenge of instance detection in clustered microglial environments and three-way morphological classification along a biologically continuous ramified-to-ameboid spectrum.

Reviewer #1

We thank the reviewer for the positive assessment and recognition of the paper’s structure and reproducibility. We hope the revisions made through the review process have brought the paper to the standard required for MICCAI.

Reviewer #2

We sincerely thank Reviewer 2 for their thorough and constructive review, which has substantively improved the camera-ready manuscript.

We have added references for both dice score and dice loss in the methods section, and the variables in Eq. 3 are now defined explicitly in the surrounding text. The discussion section has also been renamed to “Discussion and Conclusions”.

We agree that there is room for improvement with the F1 detection score, likely due to the fact that the dataset is intrinsically challenging. The Discussion has been extended to make this explicit, and we further note in the Discussion that overall performance is currently “detection-limited rather than segmentation-limited”.

As mentioned in the response to the Meta-Review, we have now justified the use of the RT-DETRv2 over RT-DETRv3 in the methods section, and provided a new “Data Availability” paragraph has been added, which outlines our data sharing policy, and encourages readers to contact us if they would like to collaborate

Reviewer #3 We thank Reviewer 3 for their generous and encouraging assessment of the manuscript, and especially for the thoughtful methodological observations.

We fully agree that a runtime comparison would be informative. As the MICCAI 2026 camera-ready guidelines do not permit the addition of new results or experiments, this comparison has been retained as something we will include for the extended journal version of this work that is currently in preparation.

A short justification has also been added to the Methods, noting why a learned MLP was preferred over alternatives for object count estimation.

We thank the meta-reviewer for the time and care they invested in evaluating this work, and for the constructive feedback that shaped our revisions. We hope the changes described above have meaningfully improved the clarity and reproducibility of the camera-ready manuscript in a satisfactory way.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 paper is an interesting contribution and if the dataset is made available (to be confirmed) is likely to have impact in the community. The camera ready version should address the question of the version of the algorithm chosen as well as the obtained detection values



back to top