Abstract

Accurate neuron segmentation in electron microscopy (EM) images is fundamental for mapping neural circuits and advancing neuroscience research. While current affinity-based paradigms have achieved notable progress, they persistently struggle with blurred membranes and intricate topologies, yielding severe split and merge errors. To tackle this, we propose a two-stage collaborative topology and connectivity learning framework that explicitly injects EM-specific semantic priors into affinity prediction. Specifically, we leverage neuron skeletons for macroscopic topological guidance and distance transform maps for precise connectivity constraints. Since the direct aggregation of these auxiliary features frequently provokes negative transfer in ambiguous regions, our framework strategically decouples multi-task extraction from feature fusion. In Stage-I, we employ an Affinity-Collaborative Weighting Mechanism (ACWM) to stabilize joint multi-task feature extraction. Crucially, rather than applying a naive hard fusion, Stage-II introduces a cascaded refinement scheme featuring a Block-wise Representation Routing (BRR) strategy. BRR adaptively evaluates and routes only the most reliable feature streams for block-level conditional fusion, inherently preventing noise propagation. To satisfy the rigorous efficiency demands of terabyte-scale inference, we compress the Stage-I backbone via gradient-based structural filter pruning. Extensive experiments across three benchmarks demonstrate that our method establishes state-of-the-art topological correctness while securing a highly efficient deployment architecture. Code is available at https://github.com/vshy-dream/CTCL-NeuronSeg.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2843_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/vshy-dream/CTCL-NeuronSeg

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ShiHao_Collaborative_MICCAI2026,
        author = { Shi, Haoyuan AND Liu, Xiaoyu AND Chen, Yinda AND Xie, Jingcheng AND Xiong, Zhiwei},
        title = { { Collaborative Topology and Connectivity Learning for EM Neuron Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16884},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper aims to address the issue of topological errors in EM neuron segmentation. It proposes incorporating neuronal skeletons and distance transform maps as prior knowledge into the affinity-based segmentation framework. The proposed method has been validated on three datasets, achieving promising results.

  • 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.Proposed a two-stage collaborative framework that incorporates neuron skeletons and distance transform maps as explicit prior knowledge, effectively resolving split and merge errors caused by ambiguous boundaries and complex topologies in EM images. 2.Designed the ACWM module to stabilize multi-task feature extraction, and utilized the BRR module for intelligent feature routing, fundamentally preventing the negative transfer issue prevalent in traditional multi-task learning.

  • 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.Abrupt Introduction and Insufficient Justification of the Pruning Module: The primary narrative of the paper focuses on addressing the pain point of “improving neuron topology and connectivity accuracy. “ Therefore, the sudden introduction of the “Gradient-based Structural Filter Pruning” module appears abrupt and somewhat disconnected from the core motivation. If the goal of this pruning is to handle latency constraints in large-scale applications, Table 4 should report the actual inference time required per 3D block. Reductions in parameter count and FLOPs do not inherently guarantee faster practical inference speeds. Furthermore, the authors should provide a direct comparison of inference latency against other baseline methods to convincingly demonstrate the true effectiveness of this pruning module. 2.Hidden Overhead of Auxiliary Label Generation: The proposed network relies heavily on the “skeleton” and “distance transform map” as auxiliary supervision signals. However, in real-world, ultra-large-scale EM connectomics research, computing 3D skeletons offline is an exceptionally time-consuming process. The authors should explicitly report the time overhead required for generating these labels and critically discuss whether this dependency limits the scalability of the proposed method for massive neuron segmentation tasks. 3.Training Overhead of the Multi-Task Architecture: While the designs of the ACWM and BRR mechanisms are highly ingenious, does the simultaneous prediction of three targets (affinity, skeleton, and distance map) significantly inflate the computational overhead and GPU memory consumption during the training phase? 4.Discrepancy Regarding Parameter Counts (Table 4 vs. Fig. 1): I noticed a significant discrepancy regarding the model size. According to Table 4, the parameter count for “Stage 1” is exactly identical to that of “Stage 2”. However, as illustrated in the architecture diagram (Fig. 1), Stage 2 incorporates an additional BRR module and subsequent refinement steps compared to Stage 1.Logically, adding these extra fusion and refinement modules should increase the total number of trainable parameters. Could the authors explain this inconsistency? 5.Ambiguity in Experimental Evaluation Protocols (Tables 1 & 4): There is considerable ambiguity regarding the evaluation settings presented in Tables 1 and 4.Based on the proposed architecture, my understanding is that Stage 2 sequentially depends on Stage 1.However, the descriptions and formatting in the tables make it appear as though Stage 1 and Stage 2 were separated and evaluated completely independently. The authors must explicitly clarify this evaluation protocol to prevent reader confusion and ensure reproducibility.

  • 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.

    (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?

    Although the paper introduces interesting mechanisms like ACWM and BRR for EM neuron segmentation, several critical issues remain unresolved, as detailed in the weaknesses.

  • 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

    The authors present a two-stage collaborative framework that utilizes a few EM specific priors to improve segmentation performance.

    Stage-I (Joint Extraction):

    • Predicts neighbor affinity + skeleton + distance-transform maps together
    • Uses ACWM (Affinity-Collaborative Weighting Mechanism) to stabilize multi-task learning
    • Skeleton provides macroscopic topological guidance
    • Distance transform ensures precise connectivity constraints Stage-II (Refinement):
    • BRR (Block-wise Representation Routing) adaptively routes only reliable features to help predict the residual of affinity prediction
    • Avoids naive fusion that causes negative transfer
  • 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 explicit split of two stages and introduction of meaningful long-range/local shape priors seems quite novel and useful.
    • Neighbor affinity and DT map are canonical tasks in connectomics field, but using a shared backbone to predict also a centerline/skeleton map is quite interesting.
    • The system explicitly models “confidence” in a specific channel of EM visual representation, which seems to address a key issue in connectomics dataset(blurry and long tail corner cases are harder vs clear/sharp boundaries, and different types of corner case may require different cues)
    • The second stage uses a residual pathway to isolate stage-1’s affinity map and effectively trains the BRR module to learn the finer grained details. This seems quite effective according to the reported ablations.
  • 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.
    • ACWM seems to allow better convergence, but the lambda weights are manually determined, wondering if there was a more principled method, or would be better if trainable/gated in stage 1 as well
    • Evaluation is constrained on smaller test data cubes that has been around for a while, would be nice to see some benchmarks on larger scale, especially long range neurite continuity type of benchmarks
  • 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.

    (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?

    Sound methodology and thorough ablations on the key components. The part I find most useful for the broader field is the formulation of the second stage, where using a top k stream that excludes some less confident sources of signal out performs naive ensemble, which might be an interesting direction for future work

  • 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 presents a two-stage framework for 3D neuron instance segmentation that combines affinity learning with auxiliary structural priors, including skeleton and distance-transform supervision, followed by a cascaded refinement stage with block-wise representation routing and a lightweight pruned backbone for improved efficiency.

  • 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 an important and practical problem in large-scale connectomics. The overall framework is fairly complete and integrates affinity learning, structure-aware auxiliary supervision, refinement, and model compression in a coherent way. The focus on stability and efficiency is relevant.

  • 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 methodological novelty appears limited, as several proposed components seem closer to standard engineering integrations than to clearly distinct methodological advances. Some parts of the pipeline are also not described clearly enough, and the absence of a validation set raises concern that some reported settings may depend on fortunate hyperparameter choices. More explicit reporting of hyperparameter settings would also improve reproducibility.

  • 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.The methodological novelty appears limited, as several components, including the affinity-collaborative weighting, cascaded refinement, and gradient-based pruning, seem closer to standard engineering integrations than to clearly distinct methodological advances. 2.Some modules are not defined clearly enough for the reader to fully follow the pipeline. In particular, the meaning of F_raw, its distinction from the Stage-I affinity prediction A ̂, and its role in the Stage-II refinement process should be clarified explicitly. 3.The manuscript should clarify how hyperparameters (e.g Top-k, pruning ratio, λ) and model selection were performed given the train/test split, since a validation protocol is important to avoid potential test-set tuning. 4.Could the authors clarify what value ranges were explored for the two loss weights λ? More generally, the final settings of the main hyperparameters should be reported more explicitly in the implementation details to improve 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 addresses an important problem with a reasonably complete and practically motivated framework, although the novelty and experimental justification of several components could be strengthened.

  • 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



Author Feedback

We sincerely thank the Area Chair and reviewers for their careful evaluation of our manuscript. We are grateful that the paper was selected for provisional acceptance. We provide below a concise point-by-point clarification of the main methodological, experimental, and presentation issues that may benefit from a more precise statement in the camera-ready version. R1-W1: We clarify that the pruning module aims to eliminate inter-feature redundancy to satisfy the rigorous deployment demands of terabyte-scale connectomics. To better justify its practical value beyond theoretical FLOPs, we will discuss the FLOPs-latency distinction and add practical inference latency comparisons to Table 4 in the camera-ready version. R1-W2: We thank the reviewer for this insightful observation. While 3D skeleton computation is indeed intensive, it is a one-time offline preprocessing step applied only to annotated training data and fully bypassed during inference on massive unlabeled EM volumes, thus preserving terabyte-scale scalability. Given the critical need for topological correctness, this one-time training overhead is a worthwhile trade-off. We will discuss it on massive neuron segmentation tasks in the camera-ready version. R1-W3: We sincerely thank the reviewer for appreciating the ingenuity of our ACWM and BRR designs. We respectfully clarify that our multi-task architecture introduces minimal training overhead. The predictions share a unified backbone with lightweight parallel output heads, and the BRR module adds only marginal parameters. We will update Table 4 and the Implementation Details with comprehensive metrics to empirically demonstrate this efficiency. R1-W4&5: Our framework is a single cascaded pipeline, not two independent models. “Ours (S-I)” evaluates the intermediate output directly from the Stage-I backbone, while “Ours (S-II)” evaluates the final output of the complete framework. The identical parameter counts in Table 4 are unfortunately a clerical error, which we will correct with accurate parameters and FLOPs in the camera-ready version. R2-W1: We sincerely thank the reviewer for this insightful suggestion and apologize for the confusion caused by our terminology. When we used the term “fixed scalars” in the manuscript, we intended to describe the stabilizing and normalizing role of these parameters during the training process, rather than implying they were manually selected. Inspired by the work [1], our task weights are formulated as learnable uncertainty variables to automatically balance multiple loss terms. We will correct this terminology in the camera-ready version to eliminate any ambiguity. R2-W2: We fully agree. While current standard benchmarks enable fair comparisons, they do not capture the challenges of brain-scale, long-range neurite continuity. We will acknowledge this limitation and highlight large-scale continuity evaluation as an important future direction in the revision. R3-W1&2: To clarify, F_raw denotes the implicit features from the raw image, whereas A^ is the explicit Stage-I affinity prediction. Our design is not a standard engineering combination but a synergistic cross-modal fusion. The cascaded refinement uses the explicit structural prior to spatially guide and filter the implicit features, while the gradient-based pruning further exploits refined affinity gradients to eliminate false-positive connections, ensuring strict topological correctness. R3-W3&4: To prevent test-set leakage, our original implementation employed a separate validation set for model selection and tuning (the last 20 images of the 100-image AC4 dataset). Since the loss weights dynamically adapt to dataset variations to balance auxiliary losses, they lack a fixed range [1]. However, for exact reproducibility, both were initialized to 0.5.We will detail these settings in the camera-ready version.

References: [1] Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, In CVPR 2018.




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.

    This paper addresses an important problem with a reasonably complete and practically motivated framework. Sound methodology and thorough ablations on the key components.



back to top