acceptodds
Withdrawn from ICLR 2027

HyperGCN-Vessel: Hybrid Geometric Representation Learning for Topology-Preserving Vascular Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate vascular segmentation requires not only high voxel-wise accuracy but also the preservation of the underlying vessel topology. Although recent deep learning models achieve impressive Dice scores, they frequently produce fragmented predictions on thin and distal vessels, limiting their utility in downstream clinical analysis. We attribute this limitation to the mismatch between conventional Euclidean feature representations and the hierarchical organization of vascular trees. To address this challenge, we propose **HyperGCN-Vessel**, a hybrid Euclidean–hyperbolic framework that combines efficient local feature extraction with geometry-aware representation learning in hyperbolic space. Specifically, the proposed framework introduces three key components: (1) an *early hybrid geometric encoder* that progressively embeds vascular representations into the Poincaré ball to better capture hierarchical structures; (2) *Topo-Gate*, an uncertainty-aware differentiable topology-preserving module that recovers disconnected vessel branches without explicit skeletonization; and (3) *HyperConv*, an efficient linear-complexity tangent-space convolution operator that enables scalable and numerically stable hyperbolic feature learning for high-resolution 3D volumes. Extensive experiments on the TubeTK dataset demonstrate that HyperGCN-Vessel substantially improves vascular connectivity while maintaining competitive segmentation accuracy, reducing the Betti number error by 65% compared with nnU-Net and achieving inference efficiency comparable to conventional Euclidean architectures.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Vascular Segmentation,Hyperbolic Representation Learning,Topology-Preserving Segmentation,Medical Image Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dwT2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposed HyperGCN-Vessel, a hybrid Euclidean–hyperbolic framework for topology-preserving vascular segmentation, motivated by the difficulty of SOTA methods in maintaining the connectivity of thin and distal vessels. The method combines an early transition from Euclidean features to the Poincaré ball with HyperConv, an efficient tangent-space convolution operator, and Topo-Gate, an uncertainty-aware differentiable regulariser designed to reduce vessel disconnections without explicit skeletonisation.

Critical strengths

Strong engineering work.

Critical weaknesses
  1. The novelty is limited The paper introduces multiple concepts, including a hybrid Euclidean–hyperbolic architecture, HyperConv, Topo-Gate, Flow-Init, HCA, and ALM, most components appear to be simple adaptations or combinations of existing geometric learning, attention, and topology-aware techniques. The main contribution is therefore their integration for vessel segmentation rather than a fundamentally new learning mechanism with strong theoretical justification. There is no new key insights neither new analytical support.

  2. The ablation study in Table 2 does not independently validate each component. Several rows change multiple factors simultaneously, making it difficult to attribute improvements to individual modules. For example, Euclidean backbone + Euclidean TG achieves a lower β0\beta_0 error than HyperGCN segmentation-only (63.10 vs. 64.70), so the results do not clearly demonstrate that hyperbolic representations outperform a simpler Euclidean topology regularizer. Moreover, removing HCA increases the β0\beta_0 error from 48.79 to 211.40 and the β1\beta_1 error from 7.40 to 100.30. This suggests either that HCA is the dominant component or that the training is unstable, neither of which is adequately discussed.

  3. Table 1 is disconnected from the main scope of the paper. Table 1 reports results on two 2D retinal vessel datasets, DRIVE and CHASE_DB1, whereas the main paper focuses on 3D MRA vessel segmentation using Conv3D. The manuscript does not explain why these datasets are included, how the 3D architecture is adapted to 2D, what training and evaluation protocols are used, or whether the baselines are retrained. The results are also not discussed anywhere in the main text, making the table appear to be an experimental remnant that was not properly integrated into the paper.

  4. Most of the citations are in the wrong format (\citet vs \citep)

  5. The statistical analysis of significance is not included in all the empirical studies.

Questions for authors

What's the typical value of the curvature c after training?

Reviewer_gcs8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

HyperGCN-Vessel combines a Euclidean segmentation network with Poincare feature representations, tangent-space convolutions, and a gated penalty on neighboring vessel features. The intended benefit is better preservation of thin vascular branches. The paper reports segmentation and connectivity results on brain MRA datasets, with additional retinal vessel results.

Critical strengths

Evaluating connectivity alongside voxel overlap is appropriate for this problem. The reported reduction in connected-component error on TubeTK is substantial enough to warrant investigation, rather than being a marginal improvement in Dice alone.

Critical weaknesses
  • The claimed relationship between Topo-Gate and repairing disconnected branches does not follow from the defined loss. Its edges connect foreground voxels that are already spatial neighbors. How does this construction produce a corrective signal across missing branch segments? Making feature embeddings closer does not itself reconnect the predicted mask. Indeed, identical foreground embeddings and zero auxiliary risk make this regularizer zero even when the foreground has several disconnected components. The cited tree embedding argument does not establish the asserted reduction in Betti error. Topo-Gate could still be a useful smoothness regularizer, but that requires a different justification and a comparison against an ordinary local feature penalty. As written, the central topology-preserving interpretation is unsupported.

  • The experiments do not establish that negative curvature is the source of the gains. HyperConv applies a standard Euclidean convolution and activation between logarithmic and exponential maps. Across successive blocks, matching maps cancel in the absence of clipping or additional manifold operations, so the wrapping alone does not demonstrate a richer hierarchical model. The full system also adds preprocessing, channel attention, clDice, and ALM. Although a Euclidean Topo-Gate variant is included, it does not retain the full combination used by the proposed model. A Euclidean counterpart with the same operators, attention, losses, and training schedule is needed to distinguish a geometric contribution from architectural and regularization changes. Without that control, the conclusion that shallow Euclidean features inherently destroy vascular topology is much stronger than the evidence.

Questions for authors

See weaknesses.

Reviewer_7XX6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HyperGCN-Vessel to address fragmented predictions in vessel segmentation. The method combines Euclidean and hyperbolic feature representations with components such as Topo-Gate to improve connectivity. Results are reported on 3D brain vessel and 2D retinal vessel datasets. On TubeTK, the reported β0 error decreases from 138.41 to 48.79 compared with nnU-Net, alongside improvements in Dice and clDice, suggesting potential practical value.

Critical strengths

S1: The paper addresses a practically relevant problem.

Disconnections in small vessels can affect downstream structural analysis even when overall segmentation accuracy is high. Reporting Dice, clDice, and topological errors provides a broader assessment than overlap metrics alone.

S2: The main experiments report improved connectivity at a reasonably modest computational cost.

In Table 4, the TubeTK β0 error decreases by approximately 64.7% relative to nnU-Net, while Dice increases from 63.52 to 69.45. Table 5 reports an inference time of 0.18 seconds per crop, compared with 0.15 seconds for nnU-Net. These results support further investigation of the approach.

Critical weaknesses

W1: Experimental settings and evaluation protocols are insufficiently documented.

Table 1 reports 2D experiments on DRIVE and CHASE_DB1, but Section 5.1 does not describe their data splits, training settings, or adaptation of the 3D network to 2D inputs. Section 5.2 does not discuss these results. The main 3D experiments also need clarification regarding vessel annotation sources, the TubeTK training/test split, and whether each metric uses raw predictions or largest connected component (LCC) outputs. These details affect reproducibility and comparison fairness.

W2: The definition and mechanistic explanation of Topo-Gate require clarification.

Section 4.2 initially describes Topo-Gate as using predictive entropy, but subsequently defines its output through a separate auxiliary branch. How this output is trained and why it represents uncertainty should be explained. The ablations support the empirical effectiveness of Topo-Gate, but do not by themselves establish that reducing neighboring feature distances directly reduces β0. The authors should clarify the scope of this interpretation and align the claims with the available evidence.

Questions for authors

Q1: What data splits, training settings, and 2D adaptation are used for Table 1?

Q2: Please specify the 3D annotation sources, the composition of the 692 IXI subjects, the TubeTK split, and the post-processing and significance-testing settings for each metric.

Q3: Clarify the two baselines. “Baseline (All Off)” in Table 2 and nnU-Net in Table 4 should not be assumed to be the same model. Please specify the former’s architecture and explain their differences.

Q4: Discuss the metric trade-offs associated with ALM. Section 5.4 explains its role in training stability. However, Table 2 shows that adding ALM reduces β0 error from 61.20 to 48.79 while Dice decreases from 70.62 to 69.45 and HD95 increases from 17.82 to 19.34. A brief discussion would be helpful.

Q5: Check the highlighting in Table 2. The clDice value of 65.79 exceeds the boldfaced 63.91.

Q6: Include the author AI disclosure. The provided PDF does not appear to contain the AI use statement required by ICLR 2027.

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

Related papers

Discussion (0)

Sign in to comment.