acceptodds
Under review as a conference paper at ICLR 2027

TopoMatch: Topology-Constrained Semi-Balanced Transport for Open-World 3D Vascular Graph Parsing

Abstract

Anatomical vessel labeling is more structured than closed-set branch classification: expected vessels may be absent, one vessel may be fragmented into multiple observations, and unseen anatomy may coexist with extraction artifacts. We introduce TopoMatch, an open-world framework that formulates 3D vascular parsing as structured correspondence between observed branches and an anatomical label space. A reversal-invariant curve tokenizer and an incidence-aware encoder retain ordered geometry and physical branch–junction structure. Topology-aware semi-balanced matching preserves branch-wise assignment, softly constrains the mass of known classes, and leaves separate unknown-anatomy and artifact slots unconstrained in total mass. A learned gate combines local and structured predictions before conditional topology-aware decoding. On CoW-VTP, TopoMatch achieves 94.72% Macro-F1 and 42.0% exact-graph accuracy. Capacity-matched controls remain close on clean recognition, whereas controlled experiments locate the clearest benefit of relaxed target mass in missing-anatomy regimes. Strict open-set evaluation further shows that separate free slots improve the distinction between unknown anatomy and extraction artifacts. Corruption-aware training improves robustness with little change in clean performance. Task-specific retraining yields consistent branch-level gains across graph-construction pipelines, larger vascular graphs, and coronary anatomy, though whole-graph recovery remains challenging on TopBrain. These results support open-world structured matching without a fixed one-to-one anatomical template.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geometric Deep Learning, Vascular Graph Parsing, Semi-Balanced Optimal Transport, Open-World Recognition, Topology-Aware Matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pnx8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper addresses anatomical vessel labelling a very specific task on vascular representations which is essentially a sub-class of graph labelling. The method employs specific tokenisation, encoding and matching strategies. The method is evaluated on three dataset, where one (TopBrain) is structurally essentially an extension of the TopCow dataset (more vessels labeled in the same type of images).

Critical strengths
  • The empirical results over the presented baselines are strong.
Critical weaknesses
  • In my opinion the work is not a fit for the general machine learning audience at ICLR. I would recommend to submit to a more specialised conference or journal.

  • I find the paper hard to follow, especially the introduction is unclear and also Figure 2 is very crowded and small font and hard to grasp. Overall it is very choppy and unclear why things are done in the way they are.

  • All benchmarks assume the existence of an ideal graph derived from ground truth labels, which is hardly going to be the case in the real world, where vessel segmentation models often contain spurious segments, disconnected vessels, and more.

    • Likewise, there is seemingly (unless I missed it) no consideration of real world anatomical variants, some of which are actually present in the training data for TopCoW/TopBrain at the very least.
  • There is no qualitative analysis of model outputs beyond a trajectory visualization in the supplementary materials, which again contributes to the feel that this work is very disconnected from its purported real world applications.

  • No comparison with models trained on fine-grained vessel segmentation, which would provide a more realistic comparison against how the task is done in practice

Questions for authors
9 comments
AuthorsOpenReview ↗

Rebuttals (1/5)

Q1: In my opinion the work is not a fit for the general machine learning audience at ICLR. I would recommend to submit to a more specialised conference or journal.

Response. The broader relevance of TopoMatch lies in a structured prediction problem: assigning a variable set of interconnected observations to a shared semantic vocabulary when categories may be absent, multiple observations may share a label, and some observations fall outside the vocabulary or arise from observation errors. Vascular parsing provides a concrete instance of this problem.

TopoMatch addresses these conditions through an assignment formulation that combines local evidence, sample-specific class availability, and relations between observations. Each observed branch retains its assignment mass, while known-class mass is softly guided by predicted visibility. Separate rejection categories distinguish valid out-of-vocabulary structures from extraction artifacts without prescribing their total mass. Compatibility at physical junctions couples the label assignments while permitting exceptions to the learned anatomical prior.

These design choices concern broader machine learning questions in structured prediction, partial matching, and open-world recognition: how to accommodate incomplete observations, how to represent different rejection causes, and how to incorporate relational priors without enforcing a fixed template. This formulation provides the methodological connection to a general machine learning audience. We will make these questions and their connection to the proposed design more explicit in the introduction and discussion.

Q2: I find the paper hard to follow, especially the introduction is unclear and also Figure 2 is very crowded and small font and hard to grasp. Overall it is very choppy and unclear why things are done in the way they are.

Response.

We will revise the introduction, framework illustration, and methodological exposition to address these four concerns.

  • For the clarification of introduction, we will organize the introduction around the central problem: assigning an observed vascular graph to an anatomical vocabulary when classes may be absent, branches may be fragmented, and observations may include unknown anatomy or extraction artifacts. These challenges will be introduced before the architectural components. We will then explain the limitations of fixed class-mass assumptions and undifferentiated rejection, and connect each proposed design to the difficulty it addresses.

  • For the improving of Figure 2, we will redraw Figure 2 with larger fonts, shorter labels, and clearer arrows. The main diagram will follow three stages: curve and incidence-based encoding, structured assignment, and prediction fusion with conditional decoding. Dense equations and secondary implementation details will be moved out of the main flow into the method description. A separate matching-detail panel will show initialization, junction-based feedback, and assignment refinement. Training-only supervision will be placed in a clearly marked dashed panel. This revised illustration will be included in the revised manuscript.

  • For improving continuity between sections, we will align the introduction, framework diagram, and method subsections with the same processing sequence. Each method subsection will begin by stating its purpose, inputs, and outputs, followed by an explanation of how its output is used by the next stage. We will also use consistent terminology for transport mass, class probabilities, local predictions, and structured assignments, and separate the training objectives from the inference procedure.

For explaining the motivation of each design, we will make the connection between task requirements and method choices explicit. Reversal-invariant encoding addresses arbitrary curve traversal direction. Incidence aggregation represents the relationships among multiple branches sharing a physical junction. Relaxed known-class mass accommodates absent anatomical categories without enforcing fixed class occupancy. Two rejection outputs represent unknown anatomy and extraction artifacts separately, while junction compatibility coordinates neighboring label assignments.

We will also explain the roles of the final prediction stages: learned fusion balances local and structured evidence, validation-based calibration adjusts prediction confidence, and conditional decoding invokes structural refinement when incompatible known-label assignments occur at junctions. These explanations will accompany the corresponding modules, so that their motivation is clear before their mathematical details are introduced.

AuthorsOpenReview ↗

Rebuttals (2/5)

Q3: All benchmarks assume the existence of an ideal graph derived from ground truth labels, which is hardly going to be the case in the real world, where vessel segmentation models often contain spurious segments, disconnected vessels, and more. Likewise, there is seemingly no consideration of real world anatomical variants, some of which are actually present in the training data for TopCoW/TopBrain at the very least.

Response.

  • For performance on graphs containing real segmentation and extraction errors, we agree that the original benchmark inputs do not fully represent the errors encountered in automatically extracted vascular graphs. CoW-VTP uses expert-corrected centerlines, TopCoW-mask and TopBrain use graphs skeletonized from reference-mask foregrounds, and ImageCAS-X uses released centerline polylines. Constructing these graphs without semantic labels does not make them equivalent to graphs extracted from segmentation predictions.

To address this concern, we evaluated graphs extracted from actual segmentation predictions for 26 CTA cases, using frozen TopoMatch models without target-domain fine-tuning. We compare TopoMatch with label voting on the same extracted graphs, keeping the graph input fixed (Table R14). TopoMatch has a paired Dice difference of -2.922 percentage points relative to label voting (95% CI [-3.876,-2.024]), with lower scores in 25 of 26 patients.

Table R14. Anatomical labeling on graphs extracted from segmentation predictions for 26 CTA cases. Dice is evaluated over the full reference volume and averaged over reference-present classes, so upstream missed vessels remain part of the evaluation.

MethodMean Dice (%)
Label voting on the same predicted graph77.122
TopoMatch on the same predicted graph74.201

We additionally report patient-level comparisons, per-class Dice, upstream missed-vessel diagnostics, and qualitative examples (Figure R2, Figure R3, and Figure R4). The full-volume evaluation includes the consequences of extraction failures rather than scoring only successfully extracted branches. It tests performance in the presence of real prediction and extraction errors, although it does not isolate the contribution of each error type. Missing vessels also cannot be recovered by relabeling branches already present in the graph.

Figure R2. Patient-level results on 26 CTA cases using actual segmentation-derived graphs. Mean full-volume reference-present-class Dice is 77.122% for label voting and 74.201% for TopoMatch on the same extracted graphs. The paired difference is -2.922 percentage points (95% CI [-3.876,-2.024]). Direct voxel labeling is shown as an additional upstream reference. Reported parser scores average the metrics from five frozen models.

Figure R3. Per-class Dice and upstream missed-vessel diagnostics on the same predicted-input cohort. Evaluation covers the full reference volume, including vessels missed by the upstream pipeline. Relabeling the supplied graph cannot reconstruct these absent structures.

  • For the performance on naturally occurring anatomical variants, we also provide an annotation-based subgroup analysis in the CoW-VTP/TopCoW cohort, covering fetal-type PCA, fenestration, third-A2, missing A1, and missing Pcom. These are naturally occurring configurations identified in the dataset, rather than variations generated by synthetic class removal. We compare TopoMatch with the capacity-matched GPS-style control and the no-topology ablation using branch accuracy, strict ground-truth junction exactness, and exact-graph accuracy (Figure R5). Table R15 reports the whole-graph labeling results and subgroup sizes.

Table R15. Exact-graph accuracy (%) in annotated anatomical-variant groups. A graph is exact when all scored branch labels are correct. Groups overlap, and the fenestration subgroup contains only four patients.

VariantPatients / graphsTopoMatchMatched controlNo topology
Fetal PCA35 / 6533.8532.3130.77
Fenestration4 / 714.2928.570.00
Third-A216 / 3122.5829.0329.03
Missing A18 / 1618.756.250.00
Missing Pcom90 / 17447.1344.8339.08
AuthorsOpenReview ↗

Rebuttals (3/5)

TopoMatch has higher exact-graph accuracy than the matched control for fetal-type PCA, missing A1, and missing Pcom, but lower accuracy for fenestration and third-A2. Compared with the no-topology ablation, its exact-graph accuracy is higher in four of the five groups. Across these variant-containing graphs, TopoMatch achieves 94.13–96.48% strict junction exactness. However, the matched control has higher branch-accuracy and strict junction-exactness point estimates in every displayed subgroup. The analysis therefore documents performance on actual anatomical variants while retaining the mixed outcomes.

Q4: There is no qualitative analysis of model outputs beyond a trajectory visualization in the supplementary materials, which again contributes to the feel that this work is very disconnected from its purported real world applications.

Response.

To make the model’s improvements and remaining errors easier to interpret, we have expanded these comparisons with reference annotations, baseline predictions, and branch-level error maps. We additionally provide qualitative examples from actual segmentation predictions to show how the method performs on realistic inputs.

1. Qualitative analysis of model outputs.

The expanded TopBrain examples compare ground truth, TopoMatch, and TaG-Net on the same input graph using fixed x–z and y–z views (Figure R14–Figure R17). The anatomical-label panels show which category is assigned to each branch. Separate error panels mark incorrect labels in red, correct labels in blue, and unresolved reference labels in gray. These panels allow readers to inspect both the predicted anatomy and the spatial distribution of labeling errors.

We include a case near the median TopoMatch error rate, the highest-error-rate case, and the cases with the largest branch-accuracy advantage and disadvantage relative to TaG-Net. In the favorable comparison, TopoMatch makes 113 errors versus 136 for TaG-Net; in the unfavorable comparison, it makes 152 versus 104 errors. Showing both comparisons makes the improvements and failures visible, while the error counts quantify their extent.

The accompanying confusion matrix provides context for these examples (Figure R6). In particular, same-side M2/M3 confusions account for 1,539 of 5,112 incorrect branch labels (30.11%), identifying a recurrent labeling difficulty across the evaluated graphs. The qualitative panels show where errors occur, and the confusion matrix summarizes which anatomical categories are frequently confused. We will further add local enlargements, arrows, and explicit reference/predicted label annotations to selected discrepancies in the revised manuscript.

2. Qualitative analysis on realistic predicted inputs.

To address the connection to practical applications, we additionally compare reference annotations, direct voxel labeling, and TopoMatch outputs on CTA cases whose graphs are extracted from actual segmentation predictions (Figure R4). These examples complement the TopBrain visualizations, which use reference-derived graphs.

The displayed cases cover the worst, lower-median, and best mean Dice differences relative to direct voxel labeling. For the displayed fixed fold-0 model, case 0016 shows a deterioration from 85.4% Dice for direct voxel labeling to 72.8% for TopoMatch, whereas case 0033 shows an improvement from 77.4% to 78.4%. The side-by-side reference and prediction panels allow inspection of anatomical label disagreements and missing vessel regions. They illustrate that downstream parsing can improve some predictions but can also introduce additional labeling errors, while vessels absent from the extracted graph cannot be recovered through relabeling.

Case selection uses mean Dice differences across five frozen parsers, while all displayed TopoMatch outputs use the same fold-0 parser. The qualitative comparator is direct voxel labeling; the principal quantitative comparison additionally uses label voting on the same extracted graph. Together, these visualizations connect the reported results to concrete model outputs under both reference-derived and actual predicted-input conditions.

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

This paper presents TopoMatch, an open-world framework for parsing 3D vascular graphs by matching observed vessel branches to an anatomical label space. The method combines a reversal-invariant curve tokenizer, an incidence-aware encoder, topology-aware semi-balanced matching, and conditional topology-aware decoding to incorporate geometric and structural information while accommodating missing anatomy, unknown vessels, and extraction artifacts. Experiments on CoW-VTP demonstrate promising branch-labeling performance, while controlled evaluations investigate the benefits of relaxed target mass and separate unknown-anatomy and artifact slots. Additional experiments across different graph-construction pipelines, larger vascular graphs, and coronary anatomy suggest that the method can generalize to diverse vascular parsing settings, although whole-graph recovery remains challenging.

Critical strengths
  1. The paper addresses a meaningful problem in medical image analysis by explicitly considering missing anatomy, fragmented branches, unseen vessels, and extraction artifacts rather than assuming a fixed one-to-one correspondence between observed branches and anatomical labels.
  2. The combination of geometric encoding, topology-aware matching, and separate representations for unknown anatomy and artifacts is reasonable, and the experiments include several relevant evaluations of robustness and generalization.
Critical weaknesses
  1. The novelty appears moderate. Topology-aware graph representation, optimal-transport-based matching, and semi-balanced assignment are established concepts; the paper should more clearly articulate the fundamental methodological advances beyond integrating these techniques into a vascular parsing framework.
  2. The experimental results suggest that some proposed components provide limited benefits under clean conditions, as the capacity-matched controls remain competitive in recognition performance. The paper should more clearly quantify the improvements attributable to each component and establish whether the gains are sufficiently substantial across different evaluation settings.
  3. The paper's presentation has considerable room for improvement. The overall framework diagram should better illustrate the data flow and interactions among the main components, while result figures and tables should present comparisons, ablation results, and cross-dataset evaluations more clearly.
  4. The technical writing and discussion require further polishing. Several methodological concepts and experimental findings need more precise explanations, particularly the distinction between unknown anatomy and extraction artifacts, the role of semi-balanced matching, and the implications of the remaining limitations in whole-graph recovery.
Questions for authors
  1. Novelty: What are the fundamental methodological contributions of TopoMatch beyond existing topology-aware graph learning and optimal-transport-based matching approaches? Could the authors clarify which components introduce genuinely new mechanisms rather than adapting established techniques?
  2. Component effectiveness: Given that capacity-matched controls perform similarly on clean recognition, how much improvement can be attributed specifically to topology-aware semi-balanced matching? Could the authors provide more detailed quantitative comparisons and ablations?
  3. Open-world recognition: How robustly can the method distinguish unknown anatomical structures from extraction artifacts when their geometric and topological characteristics overlap? How sensitive is this distinction to the composition of the training data?
  4. Generalization and limitations: Why does whole-graph recovery remain challenging on TopBrain despite the reported branch-level gains? What are the primary failure modes, and how do they affect the practical applicability of the proposed framework?
  5. Presentation and reproducibility: Could the authors improve the framework illustration, result figures and tables, and technical descriptions to make the proposed matching mechanism, experimental comparisons, and implementation details easier to understand and reproduce?
9 comments
AuthorsOpenReview ↗

Rebuttals (1/6)

Q1: What methodological advances does TopoMatch introduce beyond combining established techniques?

Response.

Thanks for the suggestion. Our methodological contribution is a structured graph-to-vocabulary matching formulation for incomplete vascular anatomy. The central design treats observed branches and anatomical classes asymmetrically: every branch must receive an assignment, while known-class mass is softly guided by patient-specific visibility and the unknown-anatomy and artifact categories have no prescribed target totals. This accommodates absent classes and multiple fragments sharing a label, while retaining distinct rejection semantics.

The formulation further couples label assignments through anatomical compatibility at shared physical junctions. Topology therefore influences the matching decisions directly, in addition to informing the learned branch representations.

The contribution lies in specifying these assignment, class-mass, rejection, and junction relationships within one structured matching problem. The graph encoders and transport updates used to solve it build on established methods.

Q2: How much does TopoMatch improve over the capacity-matched control, what does each component contribute, and are the gains substantial across evaluation settings?

Response.

We have expanded the quantitative comparisons to distinguish the improvement over the capacity-matched control, the effects of the tested component ablations, and their dependence on the evaluation setting.

  • For the improvement over the capacity-matched control, on clean CoW-VTP, TopoMatch improves Macro-F1 by 0.07% and exact-graph accuracy by 1.6% over the capacity-matched GraphGPS control, which uses the same conditional decoder (Table R9). We agree that the clean recognition difference is small. We additionally compare the two methods across 25 combinations of class removal and connectivity-preserving branch subdivision (Figure R1). TopoMatch has higher decoded Macro-F1 point estimates in all 25 conditions, with differences ranging from 0.06 to 1.33 percentage points. This comparison extends the evidence beyond clean recognition while controlling model capacity and the decoder.

Table R9. Clean CoW-VTP comparison with the capacity-matched GraphGPS control. Both methods use the same conditional decoder.

MethodMacro-F1 (%)Exact graph (%)
Matched GraphGPS with the same decoder94.6540.4
TopoMatch94.7242.0
Difference (%)+0.07+1.6
  • For the component ablations, Table R10 summarizes the available matching, topology, and decoding ablations using exact-graph accuracy, which requires all scored branch labels to be correct. The largest structural difference is associated with removing the topology pathways: exact-graph accuracy decreases from 42.0% to 35.2%. The transport-mode ablation decreases it to 39.2%, whereas removing final decoding decreases it to 41.2%. The corresponding fold-mean Macro-F1 differences for the transport-mode and no-topology comparisons are only 0.068 and 0.198 percentage points, respectively. Thus, the observed benefit is more pronounced in whole-graph correctness than in average branch recognition.

Table R10. Structural outcomes of the tested ablations. The final column reports the full model’s exact-graph accuracy minus that of each variant, in percentage points. These differences are separate comparisons and are not additive.

ConfigurationExact graph (%)Full minus variant (pp)
Full TopoMatch42.0–
Transport-mode ablation39.2+2.8
Without topology pathways35.2+6.8
Without final decoding41.2+0.8

To further examine the assignment design, we compare four combinations of known-class target treatment and rejection-mass treatment in Table R11. With four classes removed, changing known-class targets from nominal hard to soft improves Macro-F1 by 0.116 percentage points under fixed rejection mass and by 4.381 points under free rejection mass. Conversely, freeing rejection mass improves Macro-F1 by 0.347 points with soft known-class targets but decreases it by 3.918 points with nominal hard targets. These results show that the two design choices interact.

Table R11. Decoded Macro-F1 in the separately trained four-configuration comparison. The four-class-removal condition has no branch subdivision. These configurations are distinct from the main-model comparison in Table R9. Nominal hard targets are not enforced exactly in this experiment.

Known target / rejection massClean (%)Four classes removed (%)
Nominal hard / fixed95.14287.117
Nominal hard / free95.02083.199
Soft / fixed94.95687.233
Soft / free94.41087.580
AuthorsOpenReview ↗

Rebuttal (2/6)

  • For the gains across evaluation settings, the improvement over matched GraphGPS depends on the evaluation condition. On clean CoW-VTP, TopoMatch improves Macro-F1 by 0.07 percentage points and exact-graph accuracy by 1.6 points. These gains are modest and do not establish a substantial clean-recognition advantage.

Under class removal and connectivity-preserving branch subdivision, the Macro-F1 improvement ranges from 0.06 to 1.33 percentage points across 25 conditions (Figure R1). The larger gains exceed the clean-data difference, indicating that the method is more beneficial under some incomplete-anatomy conditions. However, the improvement remains small in other conditions. All 25 point estimates are positive, while only four pointwise 95% confidence intervals exclude zero, without correction for multiple comparisons. These results support an incremental robustness benefit, with greater practical relevance in selected missing-anatomy conditions.

For the separate matching-design ablation, Figure R10 shows that the benefit also depends on the evaluation condition. Compared with nominal hard/fixed matching, the soft/free configuration scores 0.732 percentage points lower on clean data but 0.463 points higher when four classes are removed without subdivision. Thus, jointly relaxing the mass specifications provides a modest benefit in this missing-class setting, rather than a consistent improvement across conditions. Figure R11 further reports the factor effects, their interactions, and confidence intervals.

Figure R10. Decoded Macro-F1 for the four combinations of known-class target treatment and rejection-mass treatment across class removal and connectivity-preserving branch subdivision. The comparison shows how the relative performance of the configurations changes with the evaluation condition.

Figure R11. Paired effects of known-class target treatment and rejection-mass treatment, in percentage points. The plotted main effects use hard-minus-soft and fixed-minus-free differences, each averaged over the other factor. The interaction measures how one factor’s effect depends on the other. Asterisks indicate pointwise 95% confidence intervals excluding zero, without correction for multiple comparisons.

AuthorsOpenReview ↗

Rebuttals (3/6)

Q3: How reliably can the method distinguish unknown anatomy from extraction artifacts when their features overlap, and how sensitive is this distinction to training-data composition?

Response.

  • The distinction concerns the origin of a branch: unknown anatomy is a genuine vessel outside the known label vocabulary, whereas an artifact is an erroneous structure introduced during extraction. It cannot therefore be defined simply by whether a branch has an unusual shape or connectivity. TopoMatch learns this distinction through separate supervision for the two rejection categories, using branch geometry and graph context. Similar local shapes may still be distinguished when their surrounding connection patterns differ. The category-specific results indicate uneven recognition across anatomical classes. Unknown-specific AP reaches 61.416% for BA but only 5.193% for 3rd-A2, while the artifact-specific results in Figure R12 remain high on the evaluated synthetic corruptions. Thus, some unknown anatomical structures remain difficult to recognize even when synthetic artifacts are readily detected.

We additionally compare single- and dual-slot designs in Table R12. The single-slot model combines both causes into Other, whereas the dual-slot model learns separate outputs. Both are evaluated on the same combined-other task to assess the effect of this design choice on overall rejection and open-set recognition. Dual slots improve both metrics for BA, Acom, and 3rd-A2, but not for R-PCA or R-Pcom. Importantly, 3rd-A2’s high Other AP of 96.704% coexists with its low unknown-specific AP: successful rejection does not necessarily identify the correct rejection cause.

Table R12. Corrected single- versus dual-slot comparison. Other AP combines held-out anatomy and synthetic artifacts as positives, with known vessels as negatives. OSCR evaluates known-class recognition together with rejection. These metrics assess overall rejection rather than confusion between the two rejection causes.

Held-out classOther AP (%) — SingleOther AP (%) — DualOSCR (%) — SingleOSCR (%) — Dual
BA89.36090.67593.22093.701
R-PCA81.57680.67887.93086.756
R-Pcom93.76093.68094.09993.522
Acom98.91399.09196.85996.921
3rd-A296.22096.70495.96096.517
  • For the sensitivity to training-data composition, we evaluated four combinations of nominal pseudo-unknown holdout and corruption target probabilities, each set to 0.1 or 0.4, while keeping the test set fixed. For BA, increasing the holdout target probability from 0.1 to 0.4 improves unknown-specific AP by 11.348 percentage points, averaged over the corruption settings (95% CI [8.918,13.264]). At the same time, OSCR decreases by 0.612 points (95% CI [-0.906,-0.304]). This shows that increasing pseudo-unknown supervision can improve unknown detection while introducing a trade-off in overall open-set recognition. For 3rd-A2, unknown-specific AP remains 4.607–5.951% across the four settings. The tested changes in training composition therefore do not resolve its weak recognition. The response to training composition is consequently different for BA and 3rd-A2.

Figure R12 presents the four recognition metrics across training settings. Figure R13 quantifies the effects of holdout, corruption, and their interaction. For BA, the effect of corruption on unknown-specific AP depends on the holdout setting, showing that the training factors interact.

Figure R12. Training-composition sensitivity for BA and 3rd-A2 with a fixed test set. Panels show Other AP, unknown-specific AP, artifact-specific AP, and OSCR across four combinations of holdout and corruption target probabilities. Historical center references are descriptive and excluded from factorial contrasts. Error bars indicate pointwise 95% patient-bootstrap confidence intervals; 3rd-A2 unknown-AP intervals condition on resamples with positive support in every fold.

Figure R13. Paired effects of training composition. Main effects compare target probabilities of 0.4 and 0.1, averaged over the other factor; interactions show how one factor’s effect depends on the other. For BA, increasing holdout probability improves unknown AP by 11.348 percentage points while reducing OSCR by 0.612 points. Pointwise 95% patient-bootstrap intervals are conditional on the fitted single-seed models and are not corrected for multiple comparisons.

Reviewer_gCH1OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a topology-aware framework for anatomical labeling of 3D vascular graphs.

An important feature is its open-world formulation, which allows missing anatomical classes, fragmented branches, unknown anatomy, and extraction artifacts. The method also introduces adaptive prediction fusion and topology-aware decoding.

The authors evaluate it on four vascular graph benchmarks, covering cerebral and coronary anatomy. The results show improvements in exact-graph accuracy and robustness to missing anatomy.

Critical strengths

x. The combination of geometric features, physical junction information, and structured matching is well motivated.

x. The experimental evaluation is relatively comprehensive. I like the capacity-matched comparisons, robustness experiments, and analysis of individual components.

x. The authors are reasonably transparent about the limitations, including the small performance difference against the strongest matched baseline and the difficulties of recognizing unknown anatomy.

Critical weaknesses

x. The contribution of semi-balanced transport over a strong graph model remains limited

The main results are encouraging. However, Table 3 shows that a capacity-matched GraphGPS with the same decoder achieves 94.65% Macro-F1 and 40.4% exact-graph accuracy, compared with 94.72% and 42.0% for TopoMatch. The difference is relatively small. I would suggest extending the evaluation to the GraphGPS baseline. This would more clearly demonstrate whether semi-balanced transport is necessary.

x. The open-world evaluation is interesting, but its practical effectiveness remains uncertain

Distinguishing unknown anatomical structures from extraction artifacts is a useful contribution. However, the open-world experiments rely on leave-one-class-out evaluation and synthetically generated artifacts. The results also show substantial variation in unknown-anatomy detection. For example, the unknown-class AUPRC ranges from 61.42% for BA to only 5.19% for 3rd-A2. This suggests that the method may struggle with some anatomically relevant unknown structures. Moreover, the near-perfect artifact detection is demonstrated on synthetic corruptions, which may not reflect the complexity of errors produced by real vessel segmentation and centerline extraction algorithms.

It would be useful to evaluate naturally occurring anatomical variants and real extraction errors, or at least provide a more detailed analysis of the failure cases. The class prevalence and corresponding chance-level AUPRC would also help interpret these results.

x. The topology evaluation does not fully establish anatomical correctness

An important part of the proposed method is its use of an anatomical compatibility matrix learned from the training data. However, the topology consistency metric is also measured against this training-derived matrix.

As acknowledged by the authors, this metric is not independent of the proposed anatomical prior. A prediction may be consistent with the learned template while still being anatomically incorrect.

I would therefore encourage the authors to give greater emphasis to topology evaluation based on independent ground-truth annotations. In particular, it would be interesting to examine whether the method can preserve rare but anatomically valid vascular configurations. The junction-exact metric is useful in this respect, but a more detailed analysis of anatomical variants would strengthen the central claim.

x. Generalization to unseen graph construction and more complex anatomy remains unclear

The evaluation includes several datasets and different vascular anatomies, which is a strength. However, all methods are retrained separately for each benchmark. The results on TopBrain reveal a limitation. Although TopoMatch achieves the highest branch-level Macro-F1, all methods obtain zero exact-graph accuracy, and TopoMatch does not achieve the best junction-exact accuracy or topology consistency. I suggest providing a more detailed failure analysis on TopBrain and considering intermediate structural metrics, such as exact labeling of individual vascular subtrees.

Questions for authors

x. The inference time is approximately 137 ms per graph, compared with 15 ms for TaG-Net. The authors could report the computational cost of the matched GraphGPS control and how runtime scales with graph size.

x. The paper could provide a clearer distinction between soft topology compatibility and strict topological constraints, as the final decoder allows violations through slack variables.

7 comments
AuthorsOpenReview ↗

Rebuttals (1/5)

Q1: The contribution of semi-balanced transport over a strong graph model remains limited.

Response.

Thanks for the reviewer’s suggestion. We extend the comparison between TopoMatch and the capacity-matched baseline to 25 conditions combining zero to four missing classes with five levels of connectivity-preserving branch subdivision. TopoMatch and this baseline use the same conditional decoder and matched model capacity, allowing us to evaluate the additional benefit over a strong graph model under these conditions.

As shown in Figure R1, TopoMatch achieves higher decoded Macro-F1 than the capacity-matched baseline across all 25 conditions, with gains of 0.06%–1.33% percentage points. With three missing classes, the improvement reaches 1.04%–1.33% across the subdivision levels. These results show that the advantage over the strong matched baseline becomes more pronounced under the tested missing-class conditions. Exact-graph differences remain mixed under severe class removal.

Figure R1. TopoMatch versus the capacity-matched baseline under class removal and connectivity-preserving branch subdivision. Both methods use the same conditional decoder. All 25 decoded Macro-F1 point differences favor TopoMatch, with the largest gain observed when three classes are absent.

This extension examines the conditions that motivate semi-balanced transport: missing classes change the observed anatomical inventory, making prescribed target class masses less appropriate. Branch subdivision separately changes graph granularity while preserving connectivity. By controlling for model capacity and decoder choice, the comparison provides additional evidence of TopoMatch’s benefit over a strong graph baseline when the anatomical inventory is incomplete.

Q2: The open-world evaluation is interesting, but its practical effectiveness remains uncertain.

Response.

Thanks for the reviewer’s suggestions. We have added an evaluation on graphs extracted from actual CTA segmentations and report class prevalence alongside unknown-anatomy AUPRC to clarify the practical effectiveness of the method.

  • For the real extracted graphs, we evaluated 26 CTA cases using graphs extracted from actual segmentation predictions, comparing TopoMatch with direct label voting on the same predicted graphs. The evaluation uses full-volume Dice over reference-present anatomical classes, including vessels missed by the predicted foreground. TopoMatch is applied without target-domain fine-tuning. As shown in Table R1, the paired TopoMatch-minus-voting difference is -2.922% (95% CI [-3.876,-2.024]), with negative differences in 25 of 26 patients. Thus, TopoMatch does not improve anatomical labeling in this frozen-transfer setting. We have included this negative result and revised the discussion to make the limitation on real predicted inputs explicit. To clarify this, we also visualize several results. Figure R2 presents the patient-wise comparisons, Figure R3 shows class-wise performance and missed foreground, and Figure R4 provides qualitative examples. These analyses also illustrate a practical limitation: relabeling cannot recover vessels absent from the extracted input.

Table R1. Anatomical labeling on real segmentation predictions from 26 CTA cases. Both methods use the same extracted graphs.

MethodMean Dice (%)
Label voting on the same predicted graph77.122
TopoMatch on the same predicted graph74.201
  • For the class prevalence and chance-level reference, we now report unknown-specific AUPRC together with the positive-class prevalence from the same five-fold evaluation (Table R2). Prevalence provides a no-information precision reference for interpreting AUPRC. BA and R-PCA achieve AUPRC values of 61.416% and 51.169%, compared with fold-mean prevalence references of 6.107% and 10.611%, respectively, indicating an above-reference ranking signal. However, 3rd-A2 reaches only 5.193% AUPRC despite exceeding its 0.819% prevalence reference. We therefore describe unknown recognition as class-dependent and explicitly acknowledge that detection of some anatomically relevant unknown structures remains to be improved.
AuthorsOpenReview ↗

Rebuttals (2/5)

Table R2. Unknown-class prevalence and mean-fold unknown-specific AUPRC. Prevalence is calculated over evaluated branch-condition observations, with known branches and synthetic artifacts as negatives. The prevalence columns provide the chance-level precision references.

Held-out classFold-mean prevalence (%)Pooled prevalence (%)Unknown AUPRC (%)
BA6.1076.11261.416
R-PCA10.61110.61151.169
R-Pcom1.8401.83836.709
Acom0.4220.42211.857
3rd-A20.8190.8195.193

Figure R2. Patient-wise anatomical labeling results on all 26 CTA cases using actual segmentation predictions. The primary comparison is TopoMatch versus label voting on the same predicted graph; direct voxel labels are also shown. Error bars indicate the range across models, not confidence intervals.

Figure R3. Class-wise full-volume Dice and reference voxels missing from the predicted foreground. The missed-foreground analysis identifies input omissions that graph relabeling cannot recover.

Figure R4. Qualitative comparisons on real predicted inputs, showing reference anatomical labels, direct segmentation labels, and TopoMatch predictions for three selected cases. The examples illustrate variation in labeling outcomes; the complete patient-wise comparison is provided in Figure R2.

Q3: The topology evaluation does not fully establish anatomical correctness.

Response.

We have expanded the ground-truth-based structural evaluation and added comparisons across five annotated anatomical-variant groups. We address the three aspects as below.

  • For topology evaluation against independent ground truth, we agree that consistency with the training-derived compatibility matrix measures agreement with the learned prior and does not independently establish anatomical correctness. We therefore emphasize branch accuracy, ground-truth junction exactness, and exact-graph accuracy, which compare predicted labels directly with the reference annotations. Table R3 reports the overall ground-truth junction-multiset results. TopoMatch achieves higher junction-multiset exactness than TaG-Net on CoW-VTP but lower exactness on TopBrain, indicating that its structural labeling advantage depends on the dataset.

Table R3. Ground-truth junction-multiset exactness (%). The reference is the annotated incident-label multiset, not the learned compatibility matrix.

DatasetTopoMatchTaG-Net
CoW-VTP96.4093.07
TopBrain70.3972.29
  • For rare but anatomically valid configurations, we evaluate five annotation-defined subgroups: fetal-type PCA, fenestration, third-A2, missing A1, and missing Pcom. These include rare configurations such as fenestration. We compare TopoMatch with the capacity-matched GPS-style control using the same decoder and with the no-topology ablation. Table R4 reports exact-graph accuracy. TopoMatch achieves higher exact-graph accuracy than the matched control for fetal-type PCA, missing A1, and missing Pcom, but lower accuracy for fenestration and third-A2. Compared with the no-topology ablation, TopoMatch achieves higher exact-graph accuracy in four of the five groups, with third-A2 as the exception.

Table R4. Exact-graph accuracy (%) in annotated anatomical-variant groups. A graph is exact only when all scored branch labels are correct. Groups are not mutually exclusive.

VariantPatients graphsTopoMatchMatched controlNo topology
Fetal PCA35 / 6533.8532.3130.77
Fenestration4 / 714.2928.570.00
Third-A216 / 3122.5829.0329.03
Missing A18 / 1618.756.250.00
Missing Pcom90 / 17447.1344.8339.08
  • For the detailed junction-exact analysis, we evaluate strict ground-truth junction exactness within each anatomical-variant group: every evaluated incident branch must receive its correct anatomical label. This is stricter than the overall multiset metric above, which requires the correct collection of incident labels but permits their permutation among branches. As shown in Table R5, TopoMatch achieves 94.13–96.48% strict junction exactness across the five groups. The matched control has higher point estimates in every group, while differences relative to the no-topology ablation are mixed. These scores assess junction-label correctness throughout variant-containing graphs. High junction-level accuracy can coexist with lower exact-graph accuracy because the latter requires every scored branch label in a graph to be correct.
AuthorsOpenReview ↗

Rebuttals (3/5)

Table R5. Strict ground-truth junction exactness (%) in the same anatomical-variant groups. Each evaluated incident branch must have the correct reference label.

VariantTopoMatchMatched controlNo topology
Fetal PCA95.7096.1195.88
Fenestration94.3496.2392.89
Third-A295.0395.9195.13
Missing A194.1394.4293.18
Missing Pcom96.4896.8996.28

Figure R5 presents branch accuracy, exact-graph accuracy, and strict junction exactness together with their uncertainty. These ground-truth-based comparisons show improvements over the no-topology ablation in several variant groups, while advantages over the matched control depend on the metric and anatomical subgroup.

Figure R5. Ground-truth-based evaluation across annotated anatomical-variant groups, comparing TopoMatch, the capacity-matched GPS-style control, and the no-topology ablation. Panels show branch accuracy, exact-graph accuracy, and strict junction exactness. Error bars indicate patient-level bootstrap 95% confidence intervals. Groups overlap; patient and graph counts are provided in Table R4.

Q4: Generalization to unseen graph construction pipelines and more complex anatomy remains unclear.

Response.

We have clarified the scope of the cross-benchmark results and expanded the TopBrain evaluation with detailed error analysis and intermediate structural metrics.

  • For the generalization across graph-construction pipelines, the original cross-benchmark experiments use separate training. The TopBrain analysis uses out-of-fold predictions on held-out cases, but the corresponding models are trained on other TopBrain graphs constructed using the same pipeline and anatomical vocabulary. These results therefore evaluate performance on held-out cases with more complex anatomy, rather than transfer to an unseen pipeline without target-domain training. To examine transfer without target-domain fine-tuning, we additionally evaluated frozen models on 26 CTA cases using graphs extracted from actual segmentation predictions, as reported in our response to Question 2 (Table R1). TopoMatch achieves 74.201% Dice compared with 77.122% for label voting on the same predicted graphs, a paired difference of -2.922 percentage points. This experiment does not establish a relabeling advantage under pipeline transfer, and we have limited the generalization claim accordingly.

  • For the detailed failure analysis on TopBrain, we provide a 36-class confusion matrix, case-level error diagnostics, and enlarged comparisons with TaG-Net (Figure R6, Figure R7, and Figure R14–Figure R17). TopoMatch correctly labels 78.20% of resolved branches, but 5,112 of 23,445 branches remain incorrect (21.80%). Among these errors, 1,539 are confusions between same-side M2/M3 categories, accounting for 30.11% of all errors. Thus, zero exact-graph accuracy reflects substantial residual errors, including systematic category confusions.

The case comparisons include near-median, high-error, favorable, and unfavorable outcomes (Table R6). Even the case with the largest relative advantage over TaG-Net contains 113 TopoMatch errors, compared with 136 for TaG-Net. Conversely, the largest relative disadvantage contains 152 versus 104 errors. These examples show that the aggregate branch-level gain coexists with substantial absolute errors and variable case-level performance. Discrete decoding leaves the labels unchanged in all 50 graphs, so it does not repair these errors in the evaluated setting.

Table R6. TopBrain case comparisons on the same observed graphs. Errors count incorrectly labeled resolved branches. The examples include both favorable and unfavorable outcomes.

Case selectionResolved branchesTopoMatch errorsTaG-Net errors
Near-median error rate (topcow_ct_013)3316970
Highest TopoMatch error rate (topcow_ct_001)384206203
Largest relative advantage (topcow_mr_016)349113136
Largest relative disadvantage (topcow_ct_004)480152104

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.