Abstract

Serial-section electron microscopy (EM) supports quantitative analysis of cellular ultrastructure, yet instance segmentation is hindered by costly expert annotations and strong domain shifts. Given the inter-slice continuity, a practical workflow is ``few-section annotation + sequential propagation.’’ SAM3 provides video-style detection and tracking-based segmentation that fits cross-section propagation, yet its concept recognition and segmentation generalization in the EM domain are unreliable and it still relies on per-frame prompts. We propose Prompt-Only Tuning for SAM3 (POT-SAM3), a parameter-efficient, low-interaction workflow for serial EM segmentation. POT-SAM3 freezes all SAM3 weights and learns only a small set of EM-target concept prompt tokens from a single annotated section, injecting EM target concepts into SAM3 without updating foundation weights. At inference, these tokens replace point/box/text prompts, enabling prompt-free detection and segmentation on each section, while the SAM3 tracker is used for cross-section mask propagation and candidate addition of newly appearing same-category targets. On three serial-section EM datasets (Lucchi, UroCell, and Guay) under one-shot evaluation, POT-SAM3 updates <0.1% parameters and improves the average Dice from 0.464 (SAM3) to 0.770, substantially reducing sequence-level annotation and interaction cost. Project: https://github.com/AI4MyBUPT/POT-SAM3

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0801_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/AI4MyBUPT/POT-SAM3

Link to the Dataset(s)

N/A

BibTex

@InProceedings{GuXun_POTSAM3_MICCAI2026,
        author = { Gu, Xuncheng AND Xiao, Li},
        title = { { POT-SAM3: Prompt-Only Tuning for One-Shot Electron Microscopy Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution is a parameter-efficient adaptation strategy that injects domain-specific concept knowledge into SAM3 through learnable prompt tokens trained on a single annotated slice. This process enables fully automatic 3D propagation without per-frame interaction. The technical novelty is the combination of one-shot token learning and SAM3’s tracking capability for inter-slice continuity, alongside an online instance completion mechanism for handling newly appearing objects.

  • 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 problem tackled by this paper is practical and makes use of the general structure of workflows in EM. Weakly annotating a single slice and automatic dense prediction on future slices is a genuine improvement and speedup compared to traditional workflows. 2.The highly parameter-efficient formulation of the method where only concept tokens are learned and SAM3’s base weights are frozen is elegant. This avoids catastrophic forgetting and keeps compute low for large 3D volumes. 3.The inverse-warp consistency as a self-supervised proxy is a smart step to force robust tokens rather than memorization of a single labeled frame. 4.The new instance discovery on later frames is simple but effective in addressing a common failure mode of tracking-based propagation.

  • 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.Limited baseline comparison: The evaluation omits established EM-specific segmentation methods that represent current practice. Comparing primarily against general-purpose models (SAM2/3, PerSAM, HSNet) and a single supervised baseline (nnU-Net) makes it difficult to assess the method’s practical utility relative to domain-standard pipelines. 2.Ambiguous “one-shot” framing: The paper compares against HSNet as a “one-shot segmentation method” but the meaning of “one-shot” differs between methods. HSNet uses a support/query few-shot learning paradigm, while POT-SAM3 learns tokens from one annotated slice. This distinction should be clarified to avoid misleading comparisons. 3.nnU-Net comparison lacks clarity: The nnU-Net comparison is denoted as 10-shot. This should be explained more clearly. Does 10-shot refer to the nnU-Net learning from 10 fully annotated sections? The first 10 or any 10? Is it exposed to all volumes in the set? 4.An analysis of how performance degrades when instances split or merge across slices, which is a common issue in EM data, would strengthen the contribution.

  • 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

    Clarifying the “one-shot” terminology and adding one or two EM-specific baselines would resolve the main concerns.

  • 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 introduces an efficient and scalable strategy that is valuable in the domain of EM microscopy while remaining highly parameter efficient. The presentation is clear, and the results are convincing. Further comparison to SOTA models would strengthen the contribution, but even in the current state the method is clearly valuable. The practical value for EM workflows and the clean, parameter-efficient formulation outweigh the baseline comparison gaps, which could be addressed in revision.

  • 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 the rebuttal. The response clarifies several important points from my original review, particularly the one-shot setup, the use of the first annotated section for HSNet, and the nnU-Net(10) supervision budget. These clarifications improve the interpretability of the experimental comparison and support the parameter-efficient framing of the method. Some concerns remain. In particular, I remain concerned that the current quantitative evaluation does not fully support all instance segmentation and tracking/propagation claims. This is closely related to my original concern about split/merge behavior across slices. Dice is useful as a foreground-overlap metric, but it can obscure important instance-level and propagation-specific failure modes, including merged instances, split instances, false new-instance discovery, identity inconsistency, and track fragmentation. The rebuttal acknowledges that instance-level and tracking-level metrics would make the evaluation more complete and states that such metrics could be computed from the existing predictions. However, these results are not provided, and the qualitative figure does not replace systematic quantitative evidence, especially if it does not explicitly demonstrate instance identities and tracking continuity. I therefore think the final version should include object-level and tracking-level metrics, such as AP/PQ, split/merge statistics, identity consistency, or track fragmentation, and should temper claims that are supported only by Dice. Nevertheless, I still find the proposed workflow practically relevant and promising. The method addresses a meaningful low-annotation setting in serial EM, the parameter-efficient adaptation strategy is elegant, and the reported results indicate clear improvements over the evaluated baselines. While the missing instance/tracking-level evaluation prevents this from being a strong accept for me, I remain positive overall. In a binary decision, I would vote accept, provided that the final version clearly acknowledges this limitation and includes the clarified experimental details from the rebuttal.



Review #2

  • Please describe the contribution of the paper

    The paper proposes POT-SAM3, a parameter-efficient one-shot adaptation of SAM 3 for instance segmentation and tracking in serial-section electron microscopy. The central idea is to freeze all SAM 3 weights and learn only a small set of class-specific “concept prompt tokens” from a single annotated section.

  • 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 identifies the two real failure modes of vanilla SAM 3 on EM — concept mismatch and the need for per-frame prompts — and the proposed solution addresses both with a single mechanism. The motivation is well-articulated and the design follows directly from it. 2.Reproducibility is good. Hyperparameters are reported in detail (learning rate, schedule, NMS thresholds, novelty thresholds, training duration), and an anonymized code link is provided. 3.Surpassing 10-shot nnU-Net on Lucchi while using only first-frame supervision and updating <0.1% of parameters is a meaningful result, and the consistency across six tasks is reassuring. The improvement over vanilla SAM 3 (0.464 → 0.770 average Dice) is large and clearly attributable to the proposed adaptation.

  • 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 limited. The core mechanism — learning a small set of prompt tokens while freezing the backbone — is a direct application of Visual Prompt Tuning (Jia et al., ECCV 2022, cited as [8]) to SAM 3.The novelty therefore lies in (a) initialization from box-derived prompt embeddings and (b) the auxiliary losses (inverse-warp consistency, token diversity). These are reasonable design choices but should be positioned more honestly as adaptation rather than a new paradigm. 2.Only Dice is reported, despite the paper framing the problem as instance segmentation. Standard instance metrics (AP at multiple IoU thresholds, panoptic quality, split/merge counts) are absent. This is a non-trivial gap because tracking-based methods can produce identity-consistent but topologically wrong masks that Dice cannot distinguish from correct ones. At minimum, AP@0.5 and AP@0.75 should be reported. 3.First-frame sensitivity is not analyzed. The entire approach hinges on the choice of the labeled section. If the first frame contains few or no instances of the target class, token learning degenerates. Reporting performance variance over different first-frame choices (e.g., 5 random selections per dataset) would substantially strengthen the empirical claims. 4.Lack of comparison with relevant SOTA results in the field. Although this paper uses few-shot segmentation, there are related works, and more comparisons should be made. 5.The “1000 epochs” training schedule for tuning three vectors per class on a single image deserves justification — this is a large compute budget for a “parameter-efficient” method, and the early-stopping behavior should be reported.

  • 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

    This is a clean, well-engineered, and timely application of visual prompt tuning to SAM 3 for serial-section EM, and the empirical results are strong enough that the workflow is likely to be useful in practice. The frame-wise analysis and the ablation are well-executed. My main hesitations are the limited methodological novelty (the core technique is VPT applied to a new backbone), the absence of instance-level metrics despite the framing, and the missing first-frame sensitivity analysis. None of these are fatal, but together they keep the paper from being a clear accept. I lean toward acceptance because the practical value to the EM community is real and the experiments are largely sound.

  • 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?

    This is a clean, well-engineered, and timely application of visual prompt tuning to SAM 3 for serial-section EM, and the empirical results are strong enough that the workflow is likely to be useful in practice. The frame-wise analysis and the ablation are well-executed. My main hesitations are the limited methodological novelty (the core technique is VPT applied to a new backbone), the absence of instance-level metrics despite the framing, and the missing first-frame sensitivity analysis. None of these are fatal, but together they keep the paper from being a clear accept. I lean toward acceptance because the practical value to the EM community is real and the experiments are largely sound.

  • 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.

    The authors’ response does not address my concern, and I still believe these weaknesses are too substantial to warrant publication.



Review #3

  • Please describe the contribution of the paper

    The paper proposes POT-SAM3, a prompt-only tuning approach for one-shot serial-section EM instance segmentation. It learns a small set of concept prompt tokens from one annotated section, freezes all SAM3 parameters, and then uses the learned tokens for automatic detection and tracking across the remaining sections.

  • 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 addresses a meaningful low-annotation EM segmentation problem and proposes a workflow that is well aligned with serial-section data. The method is practical, parameter-efficient, and directly targets the need to reduce per-slice interaction during propagation. The empirical results are strong on the reported datasets. POT-SAM3 clearly improves over SAM3, SAM2, PerSAM, HSNet, and one-shot nnU-Net, and even surpasses 10-shot nnU-Net on the Lucchi sequences. This makes the paper practically interesting despite limited methodological novelty.

  • 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.

    The novelty is moderate rather than high. Prompt tuning, one-shot SAM personalization, and tracking-based propagation are all established directions, so the contribution is mainly a targeted adaptation of SAM3 to serial EM rather than a fundamentally new method. The evaluation remains limited in scale and detail. The number of sequences is small, robustness to different annotated slices is not tested, and the paper reports Dice only, without object-level or tracking-level metrics such as instance recall, ID consistency, or track fragmentation.

  • 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 positive because the problem is important, the workflow is practical, and the empirical gains are substantial. The main reasons preventing a stronger score are limited evaluation scale, missing tracking metrics, and incomplete robustness analysis. To improve the chances of acceptance, the paper would benefit from the following revisions: 1.Add stronger tracking and instance-level evaluation. Since the method is framed as detection plus tracking for serial EM, Dice alone is not enough. The paper should report additional metrics such as instance recall, false positives per frame, identity consistency, track fragmentation, or object-level precision/recall. 2.Expand the ablation study to multiple datasets. The current ablation only on Lucchi is not sufficient to show that the same conclusions hold for UroCell and Guay. At least one harder dataset should be included to show that token learning and auxiliary losses help consistently. 3.Discuss the scope of generalization more carefully. The paper should acknowledge that the current evidence is based on a limited number of serial EM sequences, and ideally add a cross-sequence or cross-volume validation experiment to show robustness beyond the current setup.

  • 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



Author Feedback

We thank the AC and reviewers for the insightful feedback. We are glad they appreciated the practical value (R1,R2,R3), parameter-efficient design (R1,R2,R3), reproducibility (R2), and empirical effectiveness (R1,R2,R3) of our work. “Q” denotes the reviewer questions/weaknesses addressed below. Metrics (R1-Q4, R2-Q2, R3-Q2): We chose Dice as the primary metric to use a unified measure of final segmentation quality across different datasets. We agree that instance-level and tracking-level metrics would make the evaluation more complete. These metrics can in principle be computed from the existing predictions without additional experiments, and our checks show trends consistent with Dice. In accordance with the rebuttal guidelines, we cannot present detailed new quantitative results here; if permitted in the final revision, we will include more detailed metric reporting derived from the existing outputs. Fig. 2 provides a qualitative complement: under the same first-section supervision, POT-SAM3 produces more continuous masks that better match the target instances on frame 127, whereas the 10-shot nnU-Net reference shows more fragmented regions. While this cannot replace systematic instance- or tracking-level metrics, it still provides some indication of instance discovery and propagation quality. One-shot setting (R1-Q2, R1-Q3): All one-shot baselines use the same first fully annotated section. For HSNet, this section and its annotation form the support pair, and the remaining frames are queries. nnU-Net(10) is a stronger-supervision reference: for each task sequence, it is trained on 10 randomly sampled fully annotated images, yielding a separate model per task. By comparison, POT-SAM3 stores only three 256-D concept tokens per task, keeping the adaptation lightweight. First-frame sensitivity (R2-Q3): The intended workflow assumes that researchers select an initial section containing target instances. If this section is empty or highly unrepresentative, one-section adaptation becomes ill-posed; we will clarify this assumption and scope. EM-specific baseline (R1-Q1, R2-Q4): We agree that EM-specific methods are an important reference. However, these methods target different scenarios, such as fully supervised 3D segmentation, target-domain fine-tuning, or category-specific organelle models. Under our one-shot setting, directly comparable EM-specific baselines are rare, leaving this application scenario largely unexplored—this is one of the motivations for developing POT-SAM3.Forcing these methods into our one-shot setting would be unfair and would violate their original design assumptions, such as reliance on larger annotation budgets or complex 3D pipelines. Positioning (R2-Q1, R3-Q1): We did not intend to present POT-SAM3 as a new foundation model. Consistent with the manuscript, we position it as a parameter-efficient, low-interaction workflow for serial EM segmentation. Its contribution lies in adapting SAM3 to the one-section EM setting through concept-token initialization, auxiliary regularization, and new-instance discovery during propagation. We will refine the wording to avoid overstatement and make this positioning clearer. Training schedule (R2-Q5): The 1000 epochs refer to the maximum training budget. During training, we apply random augmentations to the original image as validation images and evaluate validation AUC every 10 epochs; early stopping is triggered if the AUC does not improve for 5 consecutive evaluation steps. In practice, early stopping usually terminates optimization before this limit. Thus, the training remains lightweight. Robustness analysis (R3): We agree that broader robustness analysis, including multi-dataset ablations, would further strengthen the conclusion. In accordance with the rebuttal guidelines, we cannot present additional ablation results here, but we will incorporate this suggestion in future work to better assess the robustness of each component.




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 reviews are mixed: two weak accept and one weak reject. The work is practical with strong empirical results with fixable issues.

    Points to address

    • add or justify instance-level metrics
    • one-shot vs. baseline
    • compare with EM-specific methods
  • 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 proposed method effectively adapts visual prompt tuning and SAM3 tracking to serial EM workflows. The final manuscript must temper instance segmentation and tracking claims unless additional metrics are added.



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 paper proposes POT-SAM3, a parameter-efficient one-shot adaptation strategy for serial-section electron microscopy segmentation. The method freezes SAM3 and learns a small set of concept prompt tokens from a single annotated section, then uses the learned tokens for automatic propagation across the remaining sections. The framework also includes auxiliary regularization and a new-instance discovery mechanism to handle propagation over serial EM data.

    The initial reviews were mixed but overall slightly positive, with scores of 4 / 3 / 4.Reviewers appreciated the practical motivation, the low-interaction EM annotation setting, the parameter-efficient design, and the reported empirical improvements over SAM2/3, PerSAM, HSNet, and nnU-Net baselines. The main concerns were limited comparison to EM-specific baselines, ambiguity in the “one-shot” framing, unclear details about the nnU-Net 10-shot comparison, lack of instance-level and tracking-level metrics despite the tracking/propagation framing, and limited robustness analysis across sequences and datasets.

    The rebuttal clarified several important points. The authors explained that all one-shot baselines use the same first annotated section, clarified the HSNet and nnU-Net(10) settings, positioned POT-SAM3 as a parameter-efficient low-interaction workflow rather than a new foundation model, and explained the intended first-section assumption. They also acknowledged that instance-level and tracking-level metrics would make the evaluation more complete, but did not provide new quantitative results, consistent with rebuttal constraints.

    Some limitations remain. The current evidence still relies mostly on Dice-style segmentation metrics, which may obscure instance-level and propagation-specific failures such as split/merge errors, identity inconsistency, false new-instance discovery, and track fragmentation. The comparison to EM-specific methods and the robustness analysis across datasets remain limited. These issues prevent the paper from being a strong accept.

    After rebuttal, the effective reviewer summary is A / R / A, treating the no-update reviewer as retaining the initial weak accept. Reviewer 1’s positive final opinion was based mainly on the clarified one-shot setting, the nnU-Net supervision budget, and the parameter-efficient framing. Reviewer 2 maintained reject, arguing that the remaining weaknesses were still substantial. Reviewer 3 did not update but initially leaned positive due to the practical workflow and substantial empirical gains.

    I recommend Accept. Although the paper has moderate novelty and incomplete instance/tracking-level evaluation, the problem is practically important, the workflow is well aligned with serial EM annotation, and the rebuttal clarified the main experimental-setting ambiguities. The remaining issues should be explicitly acknowledged in the final version, especially the lack of instance-level and tracking-level metrics, limited EM-specific baseline comparison, and limited robustness analysis.



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 reviewers generally agreed that the paper addresses an important and practical problem in serial-section EM segmentation and presents a parameter-efficient adaptation of SAM3 that substantially reduces annotation and interaction requirements while achieving strong performance across multiple public datasets. The rebuttal successfully clarified concerns regarding the one-shot protocol, baseline comparisons, supervision budget, workflow assumptions, and training schedule, resolving several ambiguities raised during review. While concerns remain regarding the absence of instance-level and tracking-level metrics, limited robustness analysis, and moderate methodological novelty, these issues primarily affect evaluation completeness rather than the validity of the proposed approach or the strength of the reported empirical results. Given the consistent performance gains, practical significance, and generally positive reviewer assessments, I believe the strengths outweigh the remaining weaknesses and recommend acceptance.



back to top