- 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
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.
| Method | Mean Dice (%) |
|---|---|
| Label voting on the same predicted graph | 77.122 |
| TopoMatch on the same predicted graph | 74.201 |
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.
| Variant | Patients / graphs | TopoMatch | Matched control | No topology |
|---|---|---|---|---|
| Fetal PCA | 35 / 65 | 33.85 | 32.31 | 30.77 |
| Fenestration | 4 / 7 | 14.29 | 28.57 | 0.00 |
| Third-A2 | 16 / 31 | 22.58 | 29.03 | 29.03 |
| Missing A1 | 8 / 16 | 18.75 | 6.25 | 0.00 |
| Missing Pcom | 90 / 174 | 47.13 | 44.83 | 39.08 |
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.
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.