Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Establishing large-scale, high-resolution neural connectivity maps is fundamental to elucidating the structural basis of brain function. However, when processing terabyte- or petabyte-scale electron microscopy data, over-segmentation inherent in automated reconstruction algorithms remains a critical bottleneck, requiring extensive manual proofreading spanning person-years. To alleviate the heavy reliance on annotated data and the limited flexibility of conventional tracing methods, we propose a training-free, targeted neuron tracing framework. Specifically, we introduce a skeleton-guided Heuristic Spatial Search paradigm that leverages geometric priors to iteratively reconstruct neuronal morphologies through a probing-verification cycle. To achieve robust zero-shot semantic verification, we further develop a Dimension-Aware Semantic Verification strategy built upon the foundation model NeuroSAM 2. This strategy resolves intra-slice splits via Planar Ensemble Consensus and inter-slice splits via Axial Spatio-Temporal Propagation. Notably, we integrate the proposed workflow into the Neuroglancer visualization platform, enabling an interactive human-in-the-loop proofreading system. Experimental results demonstrate that the proposed method outperforms supervised baselines and reduces manual proofreading time by 33.4%. The source code is publicly available at https://github.com/HeadLiuYun/Probe-EM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3535_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HeadLiuYun/Probe-EM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{JiaLiu_ProbeEM_MICCAI2026,
        author = { Jiang, Liuyun AND Zhang, Yanchao AND Guo, Jinyue AND Chen, Chuanyue AND Yan, Haiyang AND Yuan, Ye AND Liu, Jing AND Han, Hua},
        title = { { Probe-EM: Targeted Neuron Tracing via Training-Free Semantic Verification } },
        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 paper proposes a training-free, seed-driven framework for targeted neuron tracing in EM volumes. The system is integrated into Neuroglancer for human-in-the-loop proofreading, and is evaluated on a serial-section EM dataset of mouse SCN against three supervised connectivity-discrimination 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.The problem framing is well-motivated and aligned with practical proofreading workflows: targeted, seed-driven tracing is genuinely more useful than global pairwise merge classification when researchers want to follow specific neuronal populations rather than correct the entire volume. 2.The geometric prompt sampling design (anchor / probe / rear points) is principled. The ablation showing that removing GPS reduces F1 from 0.586 to 0.328 confirms that center-only prompting causes mask leakage, and the rear-point construction is a sensible safeguard against over-expansion. 3.The framework is integrated into Neuroglancer and evaluated with a human-in-the-loop study involving three annotators, which is uncommon in method-focused papers and provides useful evidence of practical utility.

  • 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.Missing comparison against the most relevant SOTA. eg. RoboEM is cited but not included as a baseline. The paper should at minimum justify why RoboEM cannot be compared on the SCN data, ideally include it. Also SAM3 is actually developed to solve the continuity problem, but this paper only uses SAM2. 2.”Training-free” framing is overstated. DASV depends on NeuroSAM 2, which is fine-tuned on the EMNeuron dataset. The method is training-free with respect to connectivity supervision but not with respect to neuronal EM imagery. The abstract and contributions should distinguish these clearly, otherwise the comparison against learning-based baselines (which only see SCN training labels) is not fully fair. 3.The evaluation should be more comprehensive. Only one dataset (SCN) is used, with 24 axon fascicles and 8 soma-containing neurons in the test set. Soma-seeded tracing in particular is evaluated on a very small sample. No standard deviations or per-neuron variance are reported, and no statistical tests support the comparisons in Table 1. 4.What fraction of HSS-pruned branches were actually correct (false positives of the topological pruning)? When does ASP propagation drift? The paper would benefit from a qualitative discussion of where the method breaks. 5.Skeleton-node-based recall/precision/F1 is reasonable but does not align with metrics commonly used in connectomics benchmarking such as expected run length (ERL), split rate, and merge rate. Without these, the results are not directly comparable to the broader literature on EM segmentation correction. Also the hyperparameter robustness is not analyzed. δ = 500 nm, τ_high = 0.5, τ_low = 0.1, τ_occ = 0.6, K = 5, N = 5, and the topological pruning thresholds are all fixed empirically. Only N is partially ablated. Cross-dataset generalization, which is one of the main motivations for going training-free, is not evaluated. 6.Minor: Equation (1) introduces ε_s without explicitly defining it as the terminal endpoint set (this is only clear from context), and the symbol for the search radius (δ) collides with the distance threshold notation used elsewhere in the literature.

  • 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 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

    The core idea is good and the engineering effort is real. To strengthen the paper, I would prioritize: (a) adding RoboEM as a baseline, (b) clearly separating “no connectivity supervision” from “training-free,” (c) reporting ERL or split/merge counts, (d) at least one cross-dataset evaluation (e. g. , a small slice from FAFB or H01), and (e) wall-clock numbers for throughput. The Neuroglancer integration is a real strength worth expanding on in the camera-ready.

  • 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?

    This work tackles a real and important bottleneck in connectomics proofreading. The ablation study convincingly isolates the contribution of each component, and the human-in-the-loop user study, despite its small scale, demonstrates practical value. However, the paper is held back by an undersized evaluation, the absence of some methods (eg. RoboEM and SAM3) as baselines, and somewhat overstated “training-free” claims. On balance, the contribution is meaningful enough to be valuable to the connectomics community, but the experimental rigor is below what I would expect for a strong accept. I lean toward acceptance dependent on the authors clarifying the points above in the rebuttal.

  • 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.

    Although the paper still has some unresolved issues, such as the lack of comparison with RoboEM, I believe it is still worthy of acceptance.



Review #2

  • Please describe the contribution of the paper

    The authors proposed probe-EM, a train free method to utilitize a probe-then-verify method to identify true merges.

    For the probe part, the authors proposed HSS(heuristic spatial search) to find most likely neighbor candidates at skeleton end points, filtering out spurious morphologies that likely come from glia cells, and add such candidats to a queue

    For the verification part, the authors used DASV(dimension aware semantic verification), based on a finetuned NeuroSAM2.Essenetially, for intra-slice splits, it tries to expand from both parts to check eventual overlap on expanded mask; for inter-slice splits, they used NeuroSAM2 as a video extrapolator to predict likely next frame masks, taking axial direction and trajectory information implicitly, and check actual overlap with next z-frame.

  • 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.
    • Overall idea is elegant, in EM segmentation, it’s typically more difficult to predict correct continuity through thin neurite instances than to verify if two splits belong to the same neuron. The authors successfully taps into that observation and strategically combined a simple prober with a powerful verification engine to address this limitation.
    • The train-free paradigm is notably gonna be a lot more transferable to new datasets and new problems easier and is highly practical.
    • Fully exploited the capabilities of SAM2, could provide useful data point for the industry on how to utilizes ever improving foundation 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.
    • I initially expected the prober to be more intelligent, e.g. using some sort of direction / local shape priors to get some initial confidence score but it seems quite basic, just taking everything within a distance
    • notably the method still requires a human annotator in the loop, but didn’t go one step further to try to automate this step as well, I’d be curious what the benchmark looks like (e.g. in table 2, a “automated annotator” group with much lower time but likely lower F1 score)
  • 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

    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?

    I find the methodology addressing some pain points in the EM connectomics field and the methods, although not fully automated, can be quite useful, especially if integrated well into neuroglancer GUI.

  • 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.

    No major concern from my end. The method in this paper may not be as deep or fancy but should be quite practical and finds a middle ground between full automation and assisted human intervention



Review #3

  • Please describe the contribution of the paper

    This paper presents Probe-EM, a segmentation proofreading framework for connectomics in electron microscopy (EM) data. The method performs iterative neuron tracing by combining geometric priors with a semantic verification step. A key aspect of the approach is that it does not require task-specific training; instead, it leverages a pretrained foundation model (referred to as NeuroSAM 2) for semantic validation within a reconstruction loop.

    The method is integrated into Neuroglancer, enabling interactive proofreading of automated segmentations. Experimental evaluation is conducted on a subset of the SCN dataset, comparing Probe-EM against several learning-based baselines (including a voxel-based CNN, a geometric PointNet approach, and a multimodal fusion model) on axon fascicle tracing and soma-seeded neuron reconstruction tasks. The results show improved segmentation accuracy over the compared methods. Additionally, a user study with three annotators indicates that the proposed tool can improve annotation speed while maintaining or improving segmentation quality compared to manual proofreading.

  • 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.

    • Practical and training-free framework. A key strength of the paper is the proposal of a method that does not require task-specific training. By leveraging a pretrained foundation model within a geometric tracing framework, Probe-EM can be directly deployed, which is highly appealing in connectomics workflows where annotation and training data are expensive to obtain.

    • Integration into an interactive tool. The integration with Neuroglancer significantly enhances the practical relevance of the work. Providing an interactive proofreading tool increases the potential impact of the method for real-world annotation pipelines and user-assisted reconstruction tasks.

    • Strong empirical performance under the presented setup. Under the experimental conditions described, Probe-EM outperforms several learning-based baselines. This is particularly noteworthy given that the competing methods require training, whereas the proposed approach does not.

    • Ablation study supporting method design. The paper includes a detailed ablation study analyzing the contribution of key components (Planar Ensemble Consensus, Axial Spatio-Temporal Propagation, and Geometric Point Sampling). The results show clear performance degradation when these components are removed, supporting the design choices, with ASP appearing particularly critical.

    • User study demonstrating practical benefits. The evaluation with multiple annotators shows improvements in annotation speed while maintaining or improving segmentation quality. This is an important aspect for connectomics applications, where manual proofreading is a major bottleneck.

    • Use of public data and potential reproducibility. The experiments are conducted on a subset of a public dataset (SCN), which facilitates reproducibility. The authors also state that code and data will be released, which, if fulfilled, would further strengthen the contribution.

  • 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.

    • Insufficient description of the foundation model (NeuroSAM 2). The proposed method critically depends on the pretrained model used for semantic verification. However, NeuroSAM 2 is not described in sufficient detail. It is unclear how this model was obtained (e.g., architecture, training data, fine-tuning procedure), and it does not appear to be publicly available. Moreover, despite being listed as a main contribution (together with the dimension-aware semantic verification module), its role and implementation are not adequately explained. This limits the interpretability and reproducibility of the approach.

    • Limited transparency of the experimental setup. While the SCN dataset is described, the specific subset used for training and evaluation is not sufficiently detailed. The manuscript mentions the use of patches from a limited number of axon fascicles and soma-containing neurons, but does not report the total number of samples, patch size, etc. These details are important for assessing the robustness and generalizability of the results.

    • Incomplete description of baseline training protocols. The training setup for the learning-based baselines lacks important details (e.g., learning rate scheduling, validation strategy, data augmentation, convergence behavior). Given that Probe-EM is compared against these methods, it is important to ensure that the baselines are properly optimized. Providing a more detailed description (e.g., in a supplementary table) would strengthen the validity of the comparison.

    • Limited analysis of precision–recall trade-offs. The reported results show noticeable imbalances between precision and recall in several cases. However, the paper does not provide sufficient analysis or discussion of these trade-offs. A deeper examination of failure modes (e.g., over-segmentation vs under-segmentation) would help better understand the behavior of the method and its practical implications.

  • 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

    The reference used to support the distance transform watershed method appears to be incorrect. The cited work (Holler et al., Nature 2021) does not describe this technique. Please verify and replace it with an appropriate reference.

    The manuscript would benefit from clarifying the data used in the user annotation experiment. It is currently unclear whether this evaluation is performed on the same test split used for benchmarking the learning-based methods or on a different subset. Providing this information would improve the transparency and reproducibility of the study.

  • 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 paper presents a practically relevant and technically interesting approach to segmentation proofreading in connectomics. The idea of combining geometric tracing with semantic verification, and doing so in a training-free manner by leveraging a pretrained foundation model, is appealing. The integration into Neuroglancer and the inclusion of a user study further strengthen the practical impact of the work. In particular, the reported improvements in annotation efficiency, together with competitive or superior segmentation accuracy under the presented setup, highlight the potential usefulness of the method in real-world workflows. However, several aspects of the paper limit my confidence in the strength of the contribution in its current form. A central concern is the lack of detail regarding the pretrained model (NeuroSAM 2), which plays a critical role in the proposed framework but is not sufficiently described. This makes it difficult to assess the novelty of the approach and raises concerns about reproducibility. In addition, important details of the experimental setup are missing, including the exact composition of the dataset splits (validation) and the training configuration of the learning-based baselines. Without this information, it is challenging to determine whether the comparisons are fully fair and whether the reported improvements would generalize beyond the specific setup. Furthermore, while the quantitative results are promising, the analysis could be strengthened, for instance by providing a clearer discussion of precision–recall trade-offs and potential failure modes. Addressing these points would improve both the technical clarity and the overall reliability of the conclusions. In summary, I find the problem and approach relevant and potentially impactful, but the current presentation lacks sufficient methodological transparency and detail to fully support the claims. For this reason, I place the paper slightly below the acceptance threshold, although I would be open to a more positive assessment depending on how these concerns are addressed in the rebuttal.

  • 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 Meta-Reviewer (MR) and reviewers (R1, R2, R3) for their constructive feedback and address the major concerns below: 1.”Training-Free” Formulation & NeuroSAM 2 Details (MR, R1, R3) We will refine “training-free” to “zero-shot connectivity inference” to avoid ambiguity. NeuroSAM 2 adapts SAM 2 by fine-tuning the prompt encoder and mask decoder (frozen image encoder) on EMNeuron to learn generalizable EM features. This utilizes Focal/Dice/IoU losses (25k iters, AdamW, batch=16, lr=1e-5, decay=4e-5). Model weights will be open-sourced. 2.Evaluation Comprehensiveness (MR, R1) Zero-Shot Generalization: Evaluated directly on the H01 dataset (6 test neurons, soma-based tracing), Probe-EM achieved Precision=0.447, Recall=0.556, and F1=0.450 without retraining, demonstrating robust cross-dataset transferability. Variance & Significance: Our per-neuron F1 for soma-seeded tracing is 0.544 ± 0.202, demonstrating higher accuracy and stability than the top baseline (0.293 ± 0.320) with statistical significance (p=0.011, paired t-test). Throughput: As total tracing time fluctuates with morphological complexity, we report operational throughput on a single RTX 4090: on average, it takes ~0.19s to search for a candidate pair and ~1.89s to infer its connectivity. 3.Missing Baselines & Metrics (MR, R1) RoboEM: While RoboEM is a pioneering method, fundamental design differences preclude a direct comparison. First, paradigms differ: RoboEM extracts skeletons directly from raw EM images, making it sensitive to local alignment offsets and requiring high spatial continuity. Conversely, Probe-EM starts from segmentation results, utilizing an exploratory single-seed search to connect over-segmented blocks. Second, architectural focuses diverge: RoboEM targets the connectivity of axons and dendritic spines, whereas Probe-EM focuses on reconstructing connections involving axons and complex soma regions. Regarding ERL: Evaluated on the Axon Fascicle Tracing task, Probe-EM achieves an ERL of 222.46 µm, outperforming the top baseline (157.86 µm). SAM 3: We fine-tuned SAM 2 as SAM 3’s Promptable Concept Segmentation (PCS) introduces a language modality unnecessary for our geometry-driven tracing, while SAM 2’s memory already effectively handles cross-slice continuity. Evaluating SAM 3 and exploring language-guided connectomics will be discussed as future work. 4.”Automated Annotator” Benchmark (R2) Table 1 already serves as the fully automated baseline (zero manual time). Its naturally lower F1 score compared to the human-assisted workflow (Table 2) highlights why human-in-the-loop proofreading remains essential to achieve the near-perfect fidelity required in connectomics. 5.Failure Modes & PR Trade-offs (R1, R3) Failure Modes: Across 32 test neurons, HSS pruned 5 complex candidates, mistakenly discarding only one true segment. For ASP, increasing window N from 5 to 8 degrades F1 (Table 3) as cumulative tracking errors cause mask drift into parallel structures. Future work will explore trajectory-guided memory constraints to mitigate these errors. PR Trade-offs: The Recall > Precision imbalance deliberately targets human-in-the-loop workflows. Rejecting false proposals takes seconds, whereas manually searching in PB-scale volumes demands massive labor. Conversely, supervised baselines typically miss rare connections, sacrificing Recall for Precision. 6.HSS Search Strategy (R2) HSS adopts a “lenient-in, strict-out” strategy: all nearby segments are included regardless of morphology to ensure high recall, while the DASV module subsequently ensures precision through rigorous filtering. 7.Experimental Details (R3) As per Sec. 3.2, baselines were trained on 433 positive/8,699 negative pairs. Input patches of 120x120x30 were centered at contact interfaces. Optimization: 50 epochs (converged), Adam (lr=1e-4), spatial augmentations, and WeightedRandomSampler. Models were selected via max validation F1 (80/20 split).



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.

    There are discrepancies among the three reviewers, consisting of strong accept, weak accept, and weak reject. The paper tackles the important problem of neuron tracing in EM volumes, in which a raining-free method is proposed to probe-then-verify the plausible merges, with favorable empirical results. Meanwhile, there are a number of aspects of the work that raise concerns. It would be beneficial if the authors provide further input and clarifications in the areas specifically brought out by the reviewers, including but not limited to, insufficient details of NeuroSAM 2 (an important ingredient of the proposed method), insufficient details of empirical setup, empirical evaluation on limited datasets, limited comparison baselines, and concerns of overstated claims.

  • 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.

    Most of the concerns have been addressed by the rebuttal. That said, there remain concerns such as empirical comparison with RoboEM. The authors are encouraged to carefully follow the reviews and revise the draft accordingly, if it is eventually accepted by this round of MICCAI.



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 decision is to accept the paper. During the rebuttal phase, the authors effectively addressed critical reviewer concerns by clarifying foundation model details, providing cross dataset generalization results, and reporting additional evaluation metrics. Although opinions were initially mixed regarding methodological transparency, the thorough clarifications and the clear utility of the interactive proofreading tool confirm that this work is a valuable contribution to the community.



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.

    I recommend acceptance for this paper. I think the problem is meaningful and the approach is reasonable. In particular, the probing–verification cycle is interesting. The use of foundation models is also appreciated and appears useful.

    Please make sure that the revised submission is within the 8-page limit. The current version is slightly over 8 pages, and I am not sure why it was not desk-rejected, as it appears to exceed the 8-page limitation policy.



back to top