acceptodds
Under review as a conference paper at ICLR 2027

On the Topological Sculpting for Vessel-like Structure Segmentation

Abstract

Thin 3D structures such as vessels, airways, and other branching tubular anatomies demand not only voxel-level accuracy but also correct global connectivity. Standard overlap-driven objectives often tolerate small breaks or spurious connections that severely damage topology, while many topology-aware losses are local, patch-based, or only influence training and thus cannot reliably repair errors at inference time. We present TopoSculpt, a generic refinement framework that improves the topology of predicted 3D tubular segmentations by explicitly enforcing global topological priors. TopoSculpt introduces a Topological Integrity Betti (TIB) loss that combines (i) Betti-number guidance toward a desired topology and (ii) an integrity regularizer that discourages destructive deviations from the initial prediction. To make persistent-homology computations tractable, we propose a curriculum refinement strategy that progressively transitions from coarse to fine corrections with adaptive sampling. Experiments on three challenging 3D benchmarks (pulmonary airway, Circle of Willis, and coronary artery) show consistent gains in both geometric accuracy and topology fidelity, reducing large connectivity errors by an order of magnitude and improving tree/branch detection rates. These results indicate that TopoSculpt is an effective, broadly applicable mechanism for topology-preserving refinement of 3D thin-structure segmentation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Topology Preserving, Curriculum Learning, Vessel-like Structure Segmentation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_1qpSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a new method to improve the topology of 3D tubular segmentations by explicitely enforcing global topological priors. It introduces TIB loss which combines Betti-number guidance toward a desired topology, and an integrity regularizer that discourages destructive deviation from the initial prediction. It then proposes a curriculum refinement strategyoto progressively transition from coarse to fine corrections with adaptive sampling. Experiments on multiple public benchmarks demonstarte the effectiveness of the proposed method.

Critical strengths
  1. The test-time refinement design is interesting. It does not require ground truth labels, and perform the segmentation optimization based on structural priors.
  2. This work identifies the degeneration of Betti loss, which cannot differentiate connecting or removing broken vessels. They propose integrity regularization to mitigate this problem.
Critical weaknesses
  1. The computation cost for the test time optimization is not evaluated. It's better to include the inference time for each 3D volume, the GPU memory overhead, and the efficiency improvement based on curriculum strategy compared with repetive PH computation.
  2. β0\beta_0=1 is a simple prior, which can not handle more complex situations. For example, there can be redundant/wrong vessel loops, which can only be captured by β1\beta_1. Is that possible to extend this method to higher topology constrains?
Questions for authors

Please check the weakness part.

Reviewer_2yFKOpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an approach to improve 3d segmentation of vessels and airways. The proposed approach employs a refinement network trained on a mix of two losses post hoc to a regular Unet trained in a supervised fashion. The two losses employed on the post-hoc network are similar to the losses developed by Clough et al and shit et al. The method is evaluated on three datasets of different structures.

Critical strengths

the paper contributes to an interesting but very niche field of research

Critical weaknesses
  • the motivation is partly ill framed. The paper claims that most topological approaches rely on patching. This is not necessarily the case nowadays. Methods like SkeletonRecall-Loss or clDice-Loss or even newer implementations of PH Loss can be trained on whole images. More importantly, in the inference stage basically all of these models can be applied to the whole image (if the network is relatively small and you have the right GPU). The loss does not effect if inference is possible at the whole image level or patches are needed. Furthermore, PH based losses that work on patches provably cover the global topology. The concept is called relative persistence homology and is described in references the authors provide themselves, e.g. Edelsbrunner et al. 2022

  • I find the paper hard to follow at times

  • the experimental evaluation is limited, there is many more interesting 3D segmentation tasks that could be taken into account also several high-performing methods that are mentioned in the references are not implemented.

  • there is a lack of novelty in my opinion. In my view the work is a combination of what Clough et al did to use Betti numbers to refine segmentations post-hoc and the clDice formulation by shit et al. In my opinion this is a very limited novelty compared to these works.

Questions for authors
  • For the circle of willis segmentation there are many many individual variants which also describe different number of connected components. how would one know in a deployment (inference) scenario which variant is at hand and needs to be segmented? If I dont know this, how would I be able to use this method to enforce the correct topology?

  • Why is there no evaluation of the Betti numbers in dimension 1? It could be relevant for a few cases of the circle of willis which is forming a whole loop.

Reviewer_4gSFOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes TopoSculpt, a framework for improving the connectivity of 3D vessel-like structure segmentation. The approach first produces a segmentation using a lightweight 3D U-Net operating on the whole input region. It then refines the network separately for each test image, without using test annotations. The refinement objective combines a persistent-homology-based correction term, which encourages a prescribed number of connected components, with an integrity-preservation term based on prediction consistency and skeleton overlap. A curriculum changes the frequency of persistent homology computation and the strength of the preservation term during refinement.

Experiments on airway, Circle of Willis, and coronary artery segmentation report improvements in connectivity, centerline coverage, and geometric accuracy. The central contribution is the integration of whole-region prediction and topology-guided test-time refinement for 3D tubular structures.

Critical strengths
  1. Connectivity errors matter for tubular structures, and the motivation explains why voxel overlap is insufficient. Test-time correction is a sensible way to address errors specific to an individual scan.
  2. Improvements extend beyond connected-component counts to clDice, branch detection, tree-length detection, and surface-based metrics. These joint improvements suggest that the method often recovers useful structure rather than merely optimizing a topology statistic.
  3. The correction–preservation tradeoff is thoughtfully motivated. A topology objective can reduce disconnected components by removing them, which would be undesirable when they correspond to genuine branches. The integrity-preservation term directly addresses this failure mode. The whole-view baseline, the ablation removing the preservation term, and the additional backbone comparisons provide useful supporting evidence, although further controls are needed to establish individual contributions.
  4. Presentation is well organized and visualization is clear.
Critical weaknesses

Major weakness 1: The ablations do not clearly identify the contribution of each design choice.

The comparison between the whole-view baseline and TopoSculpt establishes the benefit of the overall refinement procedure. Table 2 also shows that removing the integrity-preservation objective generally degrades segmentation quality. These are useful experiments, but they leave several proposed components coupled together.

First, the integrity-preservation objective combines prediction consistency and skeleton-based structural overlap. Removing the entire objective does not reveal whether the gains come primarily from the consistency term, the skeleton term, or their interaction. A controlled comparison of topology correction alone, correction with consistency, correction with skeleton preservation, and correction with both would make this contribution clearer.

Second, the curriculum changes both the frequency of persistent homology computation and the strength of the preservation objective. Its effects on optimization quality and computational cost are therefore difficult to separate. Comparisons against dense recomputation and fixed sparse recomputation, with preservation weights controlled, would help establish what the schedule contributes.

Finally, the whole-view baseline is already strong. On ATM, clDice increases from 86.27 for the patch-based Dice+CE baseline to 90.48 for HVModel, before reaching 92.27 with TopoSculpt. The HVModel comparison therefore usefully demonstrates the additional benefit of refinement. However, comparisons with alternative topology objectives under the same whole-view setting would better establish the advantage of the proposed objective over existing alternatives.

Overall, although Appendix G provides theoretical motivation for the design choices, the experiments establish the effectiveness of the complete framework more clearly than the separate empirical contributions of the preservation components and curriculum choices.

Major weakness 2: The computational efficiency and practical resource requirements are insufficiently characterized.

The method uses a lightweight network, but parameter count alone does not establish memory efficiency for 3D segmentation. Whole-region processing can require substantial activation memory, particularly when refinement performs backward passes through the network. The relevant cost depends on input dimensions and resolution as well as model size.

The refinement procedure also introduces substantial additional computation relative to a standard forward pass. The reported configuration performs 90 optimization updates per test image, together with skeleton-related operations and periodic persistent homology computations. Although the proposed schedule reduces the number of persistent homology computations, the remaining optimization updates still incur computation.

The manuscript does not provide enough measured evidence to assess this tradeoff. Reporting the experimental hardware does not establish the minimum hardware requirement, and reducing persistent homology evaluations does not by itself quantify end-to-end acceleration.

Minor weakness 1: The empirical scope of the topology prior is limited.

The explicitly described settings emphasize a target of one connected component: this is the intended airway topology, and the selected CoW cases satisfy this condition. This is a reasonable and useful setting because many anatomical structures are expected to be connected. I therefore regard it as a limitation of demonstrated scope rather than a major validity concern. Nevertheless, the authors should state the exact prior used for every dataset, particularly CTCA, and qualify generality claims accordingly.

Minor weakness 2: Limited characterization of incorrect reconnections.

The improvements in clDice, branch detection, tree-length detection, and surface accuracy, together with the qualitative examples in Appendix C, provide meaningful evidence that refinement improves segmentation quality beyond reducing connected-component errors. Nevertheless, these results do not fully characterize whether refinement occasionally introduces anatomically incorrect connections. Different branching arrangements can share the same connected-component count, and a short false bridge may have a limited effect on aggregate metrics. A targeted analysis of unsuccessful reconnections or representative failure cases would strengthen the anatomical interpretation of the results. I regard this as a remaining evaluation limitation rather than evidence that the reported improvements are merely topological or anatomically invalid.

Questions for authors

Please refer to the Weaknesses sections. In short,

  1. Could you provide more ablations to identify each design's contribution?
  2. What are the end-to-end runtime and peak GPU/CPU memory per volume for the baseline and complete refinement procedure, and is there a trade-off between accuracy and the refinement steps?
  3. More responses about the two minor weaknesses are also welcome.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.