acceptodds
Under review as a conference paper at ICLR 2027

DGP-HPCL: Dual-Prior-Guided Hypergraph Prototype Contrastive Learning for Brain Tumor Segmentation

Abstract

Geometric supervision can improve segmentation outputs without specifying which intermediate features should participate in relational learning. We propose dual geometric prior-guided hypergraph prototype contrastive learning (DGP-HPCL) for multimodal brain tumor segmentation. A Gaussian center prior and an interior distance prior, both derived from training labels, jointly select decoder locations. Feature similarity then defines Soft-KNN hyperedges and their aggregated representations. Within each subject, the low- and high-center-response tails define two prototypes and the embeddings supervised by the contrastive objective. Each endpoint embedding is contrasted with both prototypes, favoring its own geometric group over the opposite group. Middle-ranked nodes remain available for aggregation without receiving binary prototype labels. Auxiliary prior prediction and cross-gated bidirectional consistency complement this relational objective during training. On BraTS 2019 five-fold validation, DGP-HPCL achieves 82.38% DSC, 73.54% surface Dice, and 6.91 HD95, the best means among the evaluated models. On the BraTS 2021 evaluation split, it achieves 88.87% DSC, 84.52% surface Dice, and 5.14 HD95. Direct cross-year evaluation and controlled design comparisons further support geometry-guided sampling and endpoint supervision. The framework regularizes a shared decoder during training and requires neither labels nor hypergraph construction at inference.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
brain tumor segmentation, multimodal MRI, geometric priors, hypergraph learning, prototype contrastive learning

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_pF5sOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DGP-HPCL for multimodal brain tumor segmentation. The method uses label-derived center and interior-distance priors to select decoder features, applies Soft-KNN aggregation, and performs prototype contrast between high- and low-center-response groups. Auxiliary prior prediction and bidirectional structural consistency provide additional training supervision. Experiments are conducted on BraTS 2019 and 2021.

Critical strengths
  1. Using geometric information to guide feature selection and auxiliary representation learning is a reasonable direction.

  2. The auxiliary branches are used only during training, without requiring labels or hypergraph construction at inference.

  3. The paper includes comparisons of sampling, aggregation, and prototype supervision strategies, and achieves good results on public benchmark datasets.

Critical weaknesses
  1. The writing focuses on implementation, with limited discussion of design motivations.

  2. For the endpoint hyperedge–prototype contrast, the priors are derived from the whole-tumor mask (Y>0Y>0), without distinguishing tissue classes. Why should separating high- and low-center-response features improve tissue segmentation if the same tissue class may appear in both groups?

  3. The proposed hypergraph aggregation appears to be standard similarity-weighted KNN pooling. What higher-order interactions does it capture beyond conventional neighborhood aggregation? Since the neighbors are already highly similar to the anchor, please clarify what the benefit of this aggregation step is.

  4. The method introduces several loss weights. In particular, the sensitivity analysis does not cover lambda_c, lambda_d, and lambda_b.

Questions for authors
  1. Why should center-response-based grouping improve tissue discrimination despite its possible mismatch with semantic classes?

  2. What distinguishes the proposed hypergraph operation from standard KNN attention pooling?

4 comments
AuthorsOpenReview ↗

Response to Reviewer pF5s (Part 1 / 3)

We sincerely appreciate Reviewer pF5s for the detailed and constructive feedback. We are pleased that you recognize geometry-guided feature selection and auxiliary representation learning as a reasonable research direction.

Your concerns about design motivation and the relationship between geometric endpoint groups and tissue semantics are particularly valuable. We have expanded the corresponding explanations in Appendix F.7 and provide additional experimental evidence below.

Q1: The writing focuses on implementation, with limited discussion of design motivations.

A1: Thank you for this suggestion. The motivation of DGP-HPCL is to incorporate geometric information into intermediate decoder representation learning, complementing the dense tissue segmentation objective.

Accurate brain tumor segmentation requires both tissue discrimination and representations that capture irregular lesion geometry.

Our framework addresses two connected questions.

1) Where should relational learning occur? The Gaussian center prior captures global position relative to the whole-tumor centroid, while the interior-distance prior captures local foreground depth. Their fusion guides decoder feature selection.

2) Which contextual representations should receive direct supervision? Soft-KNN constructs contextual hyperedge embeddings from the selected features. The two center-response ranking tails provide geometric prototype targets, while intermediate candidates contribute relational context.

The resulting design connects geometric feature selection, relational aggregation, and endpoint-specific contrastive supervision.

Q2: For the endpoint hyperedge–prototype contrast, the priors are derived from the whole-tumor mask ((Y>0)), without distinguishing tissue classes. Why should separating high- and low-center-response features improve tissue segmentation if the same tissue class may appear in both groups?

A2: We appreciate this important conceptual question. The central idea is that geometric organization and tissue identity provide complementary supervision signals.

Our geometric priors are derived from the whole-tumor foreground. The center-response ranking characterizes relative geometric position within each subject.

Meanwhile, the dense segmentation objective supervises individual tissue labels. Both objectives update the shared decoder.

A tissue class may occur in either endpoint group because these groups encode relative geometry rather than semantic tissue categories.

The dense segmentation loss encourages tissue discrimination, while the prototype objective encourages contextual features to preserve geometric organization.

This joint representation learning provides a structural signal that complements tissue-specific supervision.

Q3: Why should center-response-based grouping improve tissue discrimination despite its possible mismatch with semantic classes?

A3: The benefit comes from learning semantic identity and geometric context together.

With 256 selected candidates and an endpoint fraction of 0.20, DGP-HPCL uses 51 low-response and 51 high-response anchors for direct prototype supervision. The remaining 154 candidates contribute to relational aggregation.

The endpoint groups provide geometrically separated targets, while the dense segmentation loss maintains the tissue-specific learning objective.

Experimental Evidence. The original Table 4 compares pairwise contrast, all-node hard prototypes, fixed-threshold endpoints, and quantile endpoint supervision.

The proposed quantile endpoint strategy achieves the best mean DSC (82.38%), SDC (73.54%), and HD95 (6.91).

We additionally conducted a matched-budget random-half experiment (Appendix F.2). With the same 102 supervised anchors, endpoint selection achieves 82.38% DSC, 73.54% SDC, and 6.91 HD95, compared with 82.22%, 73.40%, and 7.23 for random-half selection.

The regional evaluation (Appendix F.3) also shows particularly strong improvements in tumor-core and whole-tumor Surface Dice.

Together, these results support geometry-guided endpoint contrast as a complementary structural learning signal for the shared segmentation decoder.

We thank the reviewer for encouraging a clearer explanation of the motivation and mechanism.

AuthorsOpenReview ↗

Response to Reviewer pF5s (Part 2 / 3)

Q4: The proposed hypergraph aggregation appears to be standard similarity-weighted KNN pooling. What higher-order interactions does it capture beyond conventional neighborhood aggregation? Since the neighbors are already highly similar to the anchor, please clarify what the benefit of this aggregation step is.

A4: Thank you for this insightful question. Soft-KNN uses similarity-weighted neighborhood aggregation, which is closely related to conventional KNN attention pooling.

The central contribution of DGP-HPCL is how geometry-guided selection, multi-node contextual aggregation, and endpoint prototype contrast are coordinated.

The mechanism contains three stages.

1) Geometry-guided candidate selection. The Gaussian center and interior-distance priors select 256 decoder locations, determining which intermediate features participate in relational learning.

2) Feature-relational aggregation. For each selected feature, cosine similarity identifies its eight most similar candidate features. Temperature-scaled softmax weights, using a temperature of 0.10, determine their contributions to the aggregated embedding.

Each hyperedge embedding jointly summarizes information from multiple feature-related decoder locations.

This addresses why aggregation is useful even when neighbors are similar: the contrastive objective operates on a neighborhood-contextualized embedding rather than an isolated node representation.

Feature similarity determines the aggregation neighborhood, while the geometric priors determine the candidate pool.

3) Endpoint prototype supervision. The aggregated representations are ranked according to their center-prior responses. The two ranking tails construct subject-specific prototypes and receive direct contrastive supervision.

Middle-ranked candidates continue contributing information through relational aggregation.

The framework therefore combines geometry-guided sampling, contextual multi-node representations, and selective geometric supervision.

Q5: What distinguishes the proposed hypergraph operation from standard KNN attention pooling?

A5: The similarity-weighted pooling operation is related to standard KNN attention. The methodological distinction is its integration into a geometry-conditioned representation-learning objective, where dual priors select candidates and endpoint-specific prototypes supervise their contextual embeddings.

The multi-node aspect is represented by each hyperedge embedding aggregating several selected features. Our experimental comparisons evaluate the practical value of this contextualization relative to other relation mechanisms.

Experimental Evidence. The original Table 4 and Appendix F.7, Table 22 compare four relational mechanisms under fixed geometric candidate selection.

R3-Table1: Relation-mechanism comparison on BraTS 2019.

Relation mechanismDSC ↑SDC ↑HD95 ↓
Identity node81.2571.927.84
Pairwise KNN graph81.5272.127.43
Hard-KNN hyperedge81.9972.997.58
Soft-KNN hyperedge82.3873.546.91

Results Analysis.

1) Compared with identity-node representations, Soft-KNN improves DSC by 1.13 percentage points (pp) and SDC by 1.62 pp, while reducing HD95 by 0.93.

2) Compared with the pairwise KNN graph, DSC improves by 0.86 pp, SDC by 1.42 pp, and HD95 decreases by 0.52.

3) Compared with hard-KNN hyperedges, DSC improves by 0.39 pp, SDC by 0.55 pp, and HD95 decreases by 0.67.

These comparisons demonstrate the empirical value of similarity-weighted multi-node contextualization within our geometric prototype-learning framework.

We have expanded Appendix F.7 to explain this relationship more precisely and clarify the contribution of the full geometry-guided relational representation-learning objective.

AuthorsOpenReview ↗

Response to Reviewer pF5s (Part 3 / 3)

Q6: The method introduces several loss weights. In particular, the sensitivity analysis does not cover lambda_c, lambda_d, and lambda_b.

A6: Thank you for this important suggestion. We conducted additional sensitivity experiments for the three geometric supervision weights and incorporated the results into Appendix F.5, Table 19.

The original Appendix E already examines the candidate budget, neighborhood size, contrastive temperature, endpoint fraction, prior-fusion coefficient, and HPCL loss weight.

The new analysis specifically examines center-prior supervision, interior-distance supervision, and bidirectional structural consistency (BSC).

Implementation. We use a fixed BraTS 2019 fold-1 setting.

The default weights are 0.30 for center supervision, 0.30 for distance supervision, and 0.20 for BSC.

Each weight is independently halved or doubled while the remaining configuration is held fixed.

New Experiments.

R3-Table2: Geometric loss-weight sensitivity on BraTS 2019 fold 1.

Center weightDistance weightBSC weightDSC ↑SDC ↑HD95 ↓
0.300.300.2082.9574.816.34
0.150.300.2082.8874.827.14
0.600.300.2083.0474.426.92
0.300.150.2082.8774.406.86
0.300.600.2082.7974.156.90
0.300.300.1083.1574.356.66
0.300.300.4083.1274.706.47

Results Analysis. DSC ranges from 82.79% to 83.15%, spanning 0.36 percentage points (pp), while SDC ranges from 74.15% to 74.82%, spanning 0.67 pp.

The established default configuration achieves the lowest HD95 (6.34) and remains within 0.20 pp of the highest measured DSC and 0.01 pp of the highest SDC.

These results demonstrate stable overlap and surface agreement under the tested local changes and support the selected default as a balanced configuration.

Further Evidence. The revised manuscript now includes several complementary experiments addressing your methodological concerns.

1) Independent loss ablations (Appendix F.1). Removing BSC decreases BraTS 2021 SDC by 0.69 pp, while removing HPCL decreases it by 1.10 pp.

2) Matched-budget endpoint comparison (Appendix F.2). With the same 102 supervised anchors, endpoint selection achieves better five-fold mean DSC, SDC, and HD95 than random-half sampling.

3) Relational aggregation comparisons (Appendix F.7). Soft-KNN achieves the best means among identity-node, pairwise KNN graph, hard-KNN hyperedge, and similarity-weighted hyperedge alternatives.

Together, these experiments support the central contribution of DGP-HPCL: dual geometric priors guide decoder feature selection, Soft-KNN constructs contextual representations, and endpoint prototype contrast provides targeted structural supervision.

We sincerely appreciate the reviewer's constructive questions regarding design motivation, geometric and semantic supervision, relational aggregation, and loss-weight sensitivity.

The expanded explanations and new experiments substantially strengthen the technical justification and empirical foundation of DGP-HPCL.

We respectfully hope you would consider revisiting your rating in light of these additional results and clarifications. Thank you again for your time and thoughtful feedback. We would be happy to address any remaining questions.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper!

Reviewer_1KdcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DGP-HPCL introduces a training regularizer for multimodal brain tumor segmentation. Label-derived Gaussian center and interior-distance priors select decoder features, which are aggregated through similarity-weighted neighborhoods. Hyperedge embeddings from the two tails of the center-response ranking receive within-subject prototype contrastive supervision. Auxiliary prior prediction and cross-gated consistency complete the objective. Experiments include BraTS 2019 five-fold validation, a BraTS 2021 evaluation split, cross-year evaluation, and component and design comparisons. Inference retains only the dense segmentation path.

Critical strengths

The method makes a useful distinction between where relational features are sampled and which aggregated representations receive supervision. Retaining middle-ranked candidates for aggregation while restricting direct prototype targets to the endpoints is a concrete, testable design. The auxiliary branches operate during training, so deployment does not require geometric targets or hypergraph construction.

The experiments go beyond a full-model comparison. Table 4 separately examines sampling, aggregation, and prototype choices. The strongest empirical evidence is the surface Dice result: DGP-HPCL leads all five BraTS 2019 folds and improves the mean over SuperLightNet by 1.31 percentage points. The BraTS 2021 evaluation also improves surface Dice and HD95 over that backbone, although its Dice gain is modest. Together, these results support the framework as a useful segmentation regularizer.

Critical weaknesses

The endpoint comparison does not fully separate geometric selection from sparse supervision. Section 3.4 directly supervises 51 anchors at each endpoint, whereas the all-node comparator supervises every selected anchor. Consequently, the comparison changes both the ranking range and the populations used for prototype construction and contrastive supervision. Table 4 supports the endpoint configuration empirically, but leaves open whether its advantage specifically requires the ranking tails. A useful control would sample 51 anchors from each half of the ranking for both prototype construction and direct supervision, keeping the aggregation pool and training settings fixed. This would clarify an important design choice.

The geometric sampling rule may underrepresent disconnected components. The center prior uses a single foreground centroid, is unmasked, and has a fusion weight of 0.7 before global top-256 selection. For disconnected lesions, the centroid can lie outside the foreground, potentially concentrating candidates away from smaller components. Dense segmentation supervision still covers the full volume; the concern is where the additional relational regularization acts. Section 5 acknowledges multi-component lesions as future work, but the aggregate results do not establish the benefit in this setting. Selected-node coverage and component-level segmentation comparisons on existing disconnected-lesion cases would help define the method’s scope.

Questions for authors

Does endpoint supervision retain its advantage over a comparator using 51 randomly selected anchors from each ranking half for both prototypes and direct supervision?

In existing cases with disconnected lesions, how well does the selected candidate set cover the components, and how does component-level performance compare with SuperLightNet?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1Kdc (Part 1 / 2)

We sincerely appreciate Reviewer 1Kdc for the thoughtful and insightful feedback. We particularly appreciate your recognition of the distinction between geometric candidate selection and endpoint-specific supervision, as well as the consistent Surface Dice improvements.

Your suggestion to evaluate endpoint supervision against random-half sampling under the same supervision budget is particularly valuable. We have completed the requested experiment and incorporated the results into Appendix F.2.

Q1: Does endpoint supervision retain its advantage over a comparator using 51 randomly selected anchors from each ranking half for both prototypes and direct supervision?

A1: Thank you for proposing this precise experimental control. The original all-node comparison changes both the number and geometric positions of supervised features. We therefore designed a new experiment to isolate geometric endpoint selection from the supervision budget.

Implementation. We follow the BraTS 2019 five-fold protocol and retain the same 256 dual-prior-selected candidates and Soft-KNN aggregation.

Candidates are ranked according to their center-prior responses.

For the random-half configuration, we select 51 anchors randomly from each ranking half. The same 102 anchors are used both for prototype construction and direct contrastive supervision.

The proposed endpoint strategy instead selects 51 anchors from each ranking tail, likewise using 102 supervised anchors.

Both configurations therefore use the same candidate pool, relational aggregation mechanism, supervision count, and prototype-construction budget. The controlled difference is the geometric placement of the supervised anchors.

New Experiments. We evaluated the random-half configuration across all five BraTS 2019 folds.

R2-Table1: Matched-budget random-half comparison (Appendix F.2, Table 12).

ConfigurationDSC ↑SDC ↑HD95 ↓
Random-half Fold 182.8074.756.40
Random-half Fold 282.0973.869.54
Random-half Fold 381.9973.557.67
Random-half Fold 486.4178.285.62
Random-half Fold 577.8266.576.93
Random-half mean82.2273.407.23
Endpoint mean82.3873.546.91

Results Analysis. Endpoint selection achieves better five-fold means across all three metrics:

1) DSC: 82.38% versus 82.22%, improving by 0.16 percentage points.

2) SDC: 73.54% versus 73.40%, improving by 0.14 points.

3) HD95: 6.91 versus 7.23, decreasing by 0.32.

The improvements are modest but favor the endpoint strategy across all three aggregate metrics under a matched supervision budget.

The two ranking tails provide geometrically separated prototype targets. Meanwhile, middle-ranked candidates continue contributing contextual information through Soft-KNN aggregation.

These results complement the original Table 4, where quantile endpoint supervision outperforms pairwise contrast, all-node hard prototypes, and fixed-threshold endpoints.

The combined evidence supports the role of geometric endpoint positioning beyond simply reducing the number of directly supervised anchors.

We sincerely appreciate this excellent suggestion. The additional experiment provides more direct evidence for one of the central design choices in DGP-HPCL.

AuthorsOpenReview ↗

Response to Reviewer 1Kdc (Part 2 / 2)

Q2: In existing cases with disconnected lesions, how well does the selected candidate set cover the components, and how does component-level performance compare with SuperLightNet?

A2: Thank you for this insightful question. We conducted a connected-component analysis to examine where the additional relational supervision acts and how the final segmentation behaves across disconnected lesions.

The complete results are reported in Appendix F.4, Tables 17–18.

Implementation. We use the original 251-case BraTS 2021 test split. Whole-tumor components are defined using 26-connectivity within the 128-cubed evaluation field of view, with original component identities retained after cropping.

Among the 251 test cases, 137 contain multiple visible original foreground components.

We analyze 137 largest components, 589 smaller components, and 95 smaller components with at least 50 visible voxels, distributed across 53 cases.

New Experiments: Candidate Coverage.

A reference component is covered if at least one selected candidate or directly supervised endpoint anchor lies within it.

R2-Table2: Coverage of disconnected tumor components.

Component groupCandidate coverageEndpoint coverage
Largest135/137135/137
All smaller0/5890/589
Smaller ≥ 50 voxels0/950/95

Results Analysis. The selected candidates cover the largest component in 135 of 137 cases (98.54%).

The global Top-K strategy concentrates its relational anchors on the dominant tumor component. This behavior is consistent with the unmasked Gaussian center prior and the global centroid reference.

The segmentation backbone still receives dense voxel-level supervision throughout the volume. We therefore additionally evaluated the segmentation of individual components.

New Experiments: Component-Level Segmentation.

Predicted components are matched one-to-one with reference components using intersection-over-union. Detection requires a matched IoU of at least 0.1.

R2-Table3: Case-weighted component results (percentages).

ComponentDGP-HPCL DiceSuperLightNet DiceDGP-HPCL DetectionSuperLightNet Detection
Largest91.3091.27100.00100.00
All smaller6.427.9010.4712.74
Smaller ≥ 5024.3823.1036.3234.37

R2-Table4: Component-weighted results (percentages).

ComponentDGP-HPCL DiceSuperLightNet DiceDGP-HPCL DetectionSuperLightNet Detection
Largest91.3091.27100.00100.00
All smaller4.064.716.797.47
Smaller ≥ 5023.0823.1036.8433.68

Results Analysis.

1) Largest components. DGP-HPCL achieves 91.30% Dice, compared with 91.27% for SuperLightNet. Both methods achieve 100% detection.

2) Smaller components with at least 50 voxels. DGP-HPCL improves case-weighted Dice by 1.28 percentage points and detection by 1.95 points. Under component weighting, detection improves from 33.68% to 36.84%.

3) All smaller components. This group includes 494 components below 50 voxels. SuperLightNet achieves higher aggregate Dice and detection in this group, identifying very small disconnected lesions as a limitation of the global sampling strategy.

These measurements clarify the scope of DGP-HPCL: relational sampling emphasizes dominant tumor structures, while dense segmentation supervision covers the complete volume.

Component-aware sampling is a natural extension for distributing auxiliary relational supervision across disconnected lesions.

We sincerely appreciate the reviewer's constructive suggestions and positive assessment. The matched-budget experiment and connected-component analysis directly address the two principal concerns.

We respectfully hope these results will further strengthen your favorable assessment of DGP-HPCL. Thank you again for your time and valuable feedback. We would be happy to discuss any further questions.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily. If there are any additional points or feedback you'd like us to consider, please let us know. Your insights are invaluable to us, and we’re eager to address any remaining issues to improve our work.

Thank you for your time and effort in reviewing our paper!

Reviewer_QCBZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DGP-HPCL, a training-time regularization framework for multimodal brain tumor segmentation that combines dual geometric priors, Soft-KNN hypergraph aggregation, and endpoint-based prototype contrastive learning. A Gaussian center prior and an interior distance prior guide the selection of decoder features, while hyperedges capture feature similarities among selected locations. Rather than assigning binary targets to all sampled nodes, the method supervises only the low- and high-center-response groups, leaving intermediate nodes available for feature aggregation. Experiments on BraTS 2019 and BraTS 2021 report improvements in Dice, surface Dice, and Hausdorff distance over the evaluated baselines, with no additional hypergraph computation required at inference.

Critical strengths
  • Relatively lightweight training-time addition with promising segmentation results: The framework retains the original dense prediction pathway at inference, avoiding the need to construct hypergraphs for new scans. On BraTS 2019, it improves the SuperLightNet baseline from 81.77% to 82.38% DSC and from 72.23% to 73.54% surface Dice. The cross-year evaluation and boundary-sensitive metrics also provide useful evidence beyond overlap-based evaluation alone.
Critical weaknesses
  • Methodological novelty and the specific contribution of each component need stronger justification: The framework combines established ideas from geometric supervision, hypergraph learning, and prototype contrastive learning. Although the ablations support the selected design choices, it remains unclear how much of the improvement comes specifically from the proposed endpoint-based contrastive objective rather than from the additional geometric supervision and training losses. In particular, Table 3 does not include an ablation that removes only the bidirectional structural consistency (BSC) loss, making its individual contribution difficult to assess.

  • Empirical improvements are relatively modest, and the evaluation does not establish their robustness. On BraTS 2021, the DSC improvement over nnU-Net is only 0.09 percentage points, while the gain over SuperLightNet is 0.23 points. The hyperparameter sensitivity analysis uses only one validation fold, and the paper does not report repeated-seed experiments. Furthermore, the cross-year evaluation explicitly acknowledges that overlapping cases between BraTS releases require auditing. These limitations make it difficult to determine whether the reported gains are statistically reliable and reflect genuine generalization rather than experimental variability or dataset overlap.

Questions for authors
  • Given the relatively small DSC improvements on BraTS 2021, have the authors evaluated the method using multiple independent training seeds and statistical significance tests? How consistent are the gains across individual tumor regions (ET, TC, and WT), rather than only the average across these regions?

  • Have the authors audited patient-level overlap between the BraTS 2019 and BraTS 2021 datasets? If overlapping cases exist, how would excluding them affect the cross-year generalization results?

  • How were the main hyperparameters (KK, kk, τ\tau, ρ\rho, α\alpha, and the loss weights) selected? Were the BraTS 2019 fold-1 sensitivity results used to choose the final configuration, and was that configuration fixed before evaluation on BraTS 2021?

  • The method introduces hypergraph construction and several auxiliary losses during training. What are the additional training time and peak GPU memory requirements compared with SuperLightNet? Can the authors also report measured inference latency to substantiate the claim that the method adds no inference-time overhead?

7 comments
AuthorsOpenReview ↗

Response to Reviewer QCBZ (Part 1/4)

We sincerely appreciate Reviewer QCBZ for the detailed and constructive feedback. We appreciate your recognition of the Surface Dice improvements and the fact that our training-time regularization does not require hypergraph construction at inference.

Following your suggestions, we have conducted additional controlled experiments and expanded the analysis in Appendix F of the revised manuscript. Below, we provide detailed responses to your concerns.

Q1: Methodological novelty and the specific contribution of each component need stronger justification: The framework combines established ideas from geometric supervision, hypergraph learning, and prototype contrastive learning. Although the ablations support the selected design choices, it remains unclear how much of the improvement comes specifically from the proposed endpoint-based contrastive objective rather than from the additional geometric supervision and training losses. In particular, Table 3 does not include an ablation that removes only the bidirectional structural consistency (BSC) loss, making its individual contribution difficult to assess.

A1: Thank you for this important question. The central motivation of DGP-HPCL is to extend geometric supervision from segmentation outputs to intermediate decoder representation learning.

Our contribution lies in a coordinated design addressing where relational features are selected and which contextual representations receive direct contrastive supervision.

1) Dual-prior-guided selection. The Gaussian center prior captures global position relative to the whole-tumor centroid, while the interior-distance prior captures local foreground depth. Their fusion determines the decoder locations participating in relational learning.

2) Soft-KNN aggregation. Feature similarity constructs multi-node neighborhoods and contextual hyperedge embeddings from the selected features.

3) Endpoint prototype contrast. The low- and high-center-response ranking tails define subject-specific prototype targets, while intermediate candidates contribute relational context without direct endpoint supervision.

The novelty lies in integrating geometric candidate selection, relational aggregation, and endpoint-specific prototype supervision within a shared segmentation decoder.

New Experiments. To clarify the independent contributions of the training objectives, we conducted separate BSC and HPCL removal experiments on BraTS 2021 (Appendix F.1).

R1-Table1: Independent BSC and HPCL ablations.

ConfigurationDSC ↑SDC ↑HD95 ↓
Full DGP-HPCL88.8784.525.14
w/o BSC88.3683.836.64
w/o HPCL88.0983.425.91

Results Analysis. Removing BSC reduces DSC by 0.51 percentage points (pp) and SDC by 0.69 pp, while increasing HD95 by 1.50.

Removing HPCL reduces DSC by 0.78 pp and SDC by 1.10 pp, while increasing HD95 by 0.77.

The full-model row is the original manuscript reference, and the removal rows are their respective final evaluations.

These experiments demonstrate the complementary contributions of geometric consistency and hypergraph prototype contrast.

The original Table 4 further shows that dual-prior sampling improves DSC/SDC by 0.53/0.68 pp over random Top-K, Soft-KNN improves them by 1.13/1.62 pp over identity-node representations, and endpoint prototypes improve them by 0.37/0.60 pp over all-node hard prototypes.

Together, the original mechanism comparisons and new isolated ablations provide a clearer empirical justification for each component of DGP-HPCL.

AuthorsOpenReview ↗

Response to Reviewer QCBZ (Part 2 / 4)

Q2: Empirical improvements are relatively modest, and the evaluation does not establish their robustness. On BraTS 2021, the DSC improvement over nnU-Net is only 0.09 percentage points, while the gain over SuperLightNet is 0.23 points. The hyperparameter sensitivity analysis uses only one validation fold, and the paper does not report repeated-seed experiments. Furthermore, the cross-year evaluation explicitly acknowledges that overlapping cases between BraTS releases require auditing. These limitations make it difficult to determine whether the reported gains are statistically reliable and reflect genuine generalization rather than experimental variability or dataset overlap.

A2: Thank you for this careful assessment. We agree that the DSC margins should be examined alongside the structural metrics and additional evaluation evidence.

DGP-HPCL aims to improve geometric representation learning and boundary-sensitive segmentation quality. We therefore consider DSC, Surface Dice (SDC), and HD95 together.

R1-Table2: Original BraTS 2021 evaluation results.

MethodDSC ↑SDC ↑HD95 ↓
nnU-Net88.7883.866.30
SuperLightNet88.6484.125.84
DGP-HPCL88.8784.525.14

Results Analysis. Compared with nnU-Net, DSC improves by 0.09 percentage points (pp), SDC by 0.66 pp, and HD95 decreases by 1.16.

Compared with SuperLightNet, DSC improves by 0.23 pp, SDC by 0.40 pp, and HD95 decreases by 0.70 (approximately 12%).

These results show that the structural metrics provide particularly relevant evidence for the proposed geometric learning objective.

On BraTS 2019, DGP-HPCL improves mean DSC by 0.61 pp, SDC by 1.31 pp, and HD95 by 0.89 over SuperLightNet. Importantly, it achieves the highest SDC across all five validation folds.

Our experiments follow the official SuperLightNet experimental configuration and supplied data partitions. BraTS 2019 uses five folds over 335 cases, while BraTS 2021 uses fixed 875 training / 125 validation / 251 test case lists.

We have additionally conducted independent loss ablations and an overlap-filtered cross-year comparison, described in Appendices F.1 and F.3.

Q3: Given the relatively small DSC improvements on BraTS 2021, have the authors evaluated the method using multiple independent training seeds and statistical significance tests? How consistent are the gains across individual tumor regions (ET, TC, and WT), rather than only the average across these regions?

A3: We strictly followed the official SuperLightNet experimental settings: five-fold cross-validation on BraTS 2019 and a fixed training-validation-test split on BraTS 2021, using the same random seed for training. Independent multi-seed experiments and formal statistical significance tests were not conducted.

New Experiments. To investigate individual tumor regions, we added complete ET/TC/WT measurements on the 184-case overlap-filtered cross-year target set (Appendix F.3, Table 15).

R1-Table3: Regional comparison with SuperLightNet.

MetricRegionSuperLightNetDGP-HPCL
DSC ↑ET83.7583.63
TC86.3686.80
WT88.1688.46
SDC ↑ET85.1485.08
TC79.7480.86
WT75.9277.72
HD95 ↓ET6.846.22
TC8.608.15
WT12.2611.87

Results Analysis. TC SDC improves by 1.12 pp and WT SDC by 1.80 pp, with corresponding DSC gains of 0.44 and 0.30 pp.

HD95 improves in all three regions. ET DSC and SDC remain close to SuperLightNet.

The results reveal especially strong improvements in tumor-core and whole-tumor surface delineation, aligning with the structural learning objective of DGP-HPCL.

AuthorsOpenReview ↗

Response to Reviewer QCBZ (Part 3 / 4)

Q4: Have the authors audited patient-level overlap between the BraTS 2019 and BraTS 2021 datasets? If overlapping cases exist, how would excluding them affect the cross-year generalization results?

A4: Thank you for raising this important concern. We conducted a systematic cross-year overlap audit and repeated the evaluation after conservatively excluding all flagged cases. The complete results are provided in Appendix F.3, Tables 13–16.

Implementation. We audited the union of 335 BraTS 2019 source cases against the BraTS 2021 data lists.

Cross-year identifier mapping identified 239 candidates in the BraTS 2021 training list, 29 in validation, and 67 in the original 251-case test list.

Among the 67 flagged test cases, four-modality image-content verification identified 21 exact duplicates. The remaining 46 cases had mapping-based overlap evidence without exact image equality.

We conservatively excluded all 67 flagged test cases, retaining 184 target cases.

New Experiments. We re-evaluated eight methods, each with its five BraTS 2019 source-fold checkpoints, on the same 184 filtered cases. This produced 40 separate evaluations, without target-domain fine-tuning or ensemble predictions.

R1-Table4: Overlap-filtered cross-year evaluation (Appendix F.3, Table 14).

MethodDSC ↑SDC ↑HD95 ↓
nnU-Net86.0779.509.52
TransBTS75.5959.1812.09
Swin UNETR86.0077.659.11
MedNeXt83.8474.3710.46
SegMamba84.5076.3010.04
SuperLightNet86.0980.279.23
ACM-UNet85.6478.2411.17
DGP-HPCL86.3081.228.75

Results Analysis. DGP-HPCL achieves the best overall DSC, SDC, and HD95 among all eight methods after overlap filtering.

Compared with SuperLightNet, DSC improves by 0.21 percentage points (pp), SDC by 0.95 pp, and HD95 decreases by 0.48.

These results strengthen the cross-year generalization evidence under a conservative, common target-set evaluation. Appendix F.3 also reports all five source-fold measurements and complete regional results.

Q5: How were the main hyperparameters (K, k, τ, ρ, α, and the loss weights) selected? Were the BraTS 2019 fold-1 sensitivity results used to choose the final configuration, and was that configuration fixed before evaluation on BraTS 2021?

A5: The backbone and data protocol follow the official SuperLightNet experimental setting. The main geometric and relational parameters were examined using the BraTS 2019 development setting, including fold-1 sensitivity experiments.

The configuration uses 256 candidate nodes, 8 neighbors, a contrastive temperature of 0.10, an endpoint fraction of 0.20, and a prior-fusion coefficient of 0.70.

The center, distance, and BSC loss weights are 0.30, 0.30, and 0.20, respectively. The HPCL coefficient is linearly warmed up to 0.05 over 30 epochs.

The complete configuration was fixed before evaluation on BraTS 2021 and remained unchanged.

The original Appendix E evaluates the candidate budget, neighborhood size, temperature, endpoint fraction, fusion coefficient, and HPCL weight.

We further added geometric loss-weight sensitivity experiments in Appendix F.5, Table 19.

Across seven tested configurations, DSC ranges from 82.79% to 83.15% and SDC from 74.15% to 74.82%, spanning only 0.36 and 0.67 pp. The established default achieves the lowest HD95 (6.34) while remaining competitive in DSC and SDC.

These additional measurements support the default as a balanced configuration under the examined local variations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

For clarity and simplicity, we refer to Reviewers QCBZ, 1Kdc, and pF5s as R1, R2, and R3, respectively, in the following response.

We sincerely thank all reviewers for their thoughtful and constructive feedback. We appreciate their recognition of the strengths and potential of our work.

We have carefully addressed the reviewers' concerns and incorporated additional experiments, methodological clarifications, and relevant updates to strengthen our manuscript. Below, we summarize the core contributions of DGP-HPCL and the revisions made during rebuttal.


Core Contributions of Our Work

  1. Dual-Prior-Guided Feature Selection: We combine a Gaussian center prior and an interior-distance prior to guide decoder feature selection, integrating global positional information and local foreground geometry into relational representation learning.

  2. Hypergraph Prototype Contrastive Learning: We construct contextual hyperedge embeddings through Soft-KNN aggregation and introduce endpoint-specific prototype contrast. The low- and high-center-response groups define subject-specific prototypes, while intermediate candidates contribute relational context without receiving direct endpoint labels.

  3. Bidirectional Structural Consistency: We introduce cross-gated consistency between auxiliary center- and distance-prior predictions, providing complementary geometric supervision for the shared decoder and improving structural segmentation quality.


Updates of Experimental Results During Rebuttal

  1. Appendix F.1: Added independent BSC and HPCL ablation experiments to quantify their individual contributions to segmentation performance (R1).

  2. Appendix F.2: Conducted a matched-budget random-half comparison with the same 102 directly supervised anchors, further validating the geometric endpoint selection strategy (R2).

  3. Appendix F.3: Completed a cross-year patient-overlap audit and re-evaluated eight methods on 184 filtered BraTS 2021 cases. Added complete ET/TC/WT regional comparisons to examine segmentation performance across tumor regions (R1).

  4. Appendix F.4: Added candidate-coverage and component-level segmentation analyses on 137 cases with disconnected lesions, examining the performance and limitations of global geometric sampling (R2).

  5. Appendix F.5: Conducted additional sensitivity experiments for the center, distance, and BSC loss weights to evaluate the stability of the selected configuration (R1, R3).

  6. Appendix F.6: Added measured training time, peak GPU memory, parameter counts, and inference latency comparisons with SuperLightNet, clarifying the computational costs of the additional training objectives and the deployed segmentation pathway (R1).


Updates of In-Depth Discussions During Rebuttal

  1. Main Text, Section 4.4: Added an explicit reference to the independent BSC and HPCL ablations in Appendix F.1, strengthening the explanation of their individual contributions (R1).

  2. Main Text, Section 5: Expanded the discussion of independent loss ablations, matched-budget endpoint supervision, and disconnected lesions. Clarified the limitations of global geometric sampling and identified component-aware sampling as a future direction (R1, R2).

  3. Appendix F.7: Expanded the discussion of dual geometric priors, the complementary roles of geometric endpoint supervision and tissue-specific segmentation learning, and the relationship between Soft-KNN hypergraph aggregation and conventional neighborhood aggregation (R3).


We believe these additional experiments and clarifications comprehensively address the reviewers' concerns and further strengthen the methodological clarity and experimental evidence of DGP-HPCL. All revisions are highlighted in blue-colored text in the revised manuscript for ease of reference, including the updates to main-text Sections 4.4 and 5 and Appendix F.1–F.7.

We look forward to the reviewers' favorable consideration and remain grateful for their valuable feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.