acceptodds
Under review as a conference paper at ICLR 2027

Reconstruction First: What Drives Reliable Coronary Vessel Recovery?

Abstract

Coronary vessel segmentation is structurally fragile: a few missed pixels can disconnect thin distal branches even when overall regional overlap is high. We ask whether a substantial part of this fragility can be addressed by better preserving and reconstructing fine local vessel evidence, and what additional benefit structure-aware modeling provides once reconstruction is strengthened. To disentangle these contributions, we develop and systematically decompose a reconstruction-first framework centered on a Context-Guided Vessel Reconstruction Decoder (CVRD), which preserves high-resolution local evidence, integrates multi-level contextual features to guide its interpretation, and progressively restores full-resolution vessel predictions. In controlled component ablations, weakening local detail extraction or replacing staged recovery produces the largest drops in both IoU and clDice among the tested reconstruction components. Across XACV, DCA1, and XCAD, the streamlined model yields consistent results over multiple seeds. Structural auxiliary objectives provide small but repeatable gains, but matched non-structural supervision and inference-time controls do not support a geometry-specific explanation for these improvements. Together, these results support reconstruction quality as a key contributor to coronary vessel recovery, while the evaluated structural supervision provides a smaller residual benefit once reconstruction is strong.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
coronary vessel segmentation, reconstruction-first learning, thin-structure segmentation, multi-scale feature reconstruction, structure-aware supervision, X-ray coronary angiography

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ve9oOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a pipeline for robust blood vessel segmentation in X-ray angiography. To this end, the paper proposes training a decoder to reconstruct blood vessel segments while jointly minimizing multiple objective functions, including vessel prediction, region prediction, centreline, and structural membership. The experiments are performed on three different benchmarks—XACV, DCA1, and XCAD—and the results are compared with those of multiple existing works. The experimental results demonstrate superior performance compared to existing methods across multiple metrics.Segmentation of blood vessels in X-ray angiography has significant importance in clinical practice. Making such a pipeline robust and reliable would be crucial for its adoption in clinical practice. The experimental results look competitive compared to those of existing methods.

Critical strengths

1.Segmentation of blood vessels in X-ray angiography has significant importance in clinical practice. Making such a pipeline robust and reliable would be crucial for its adoption in clinical practice.

  1. The experimental results look competitive compared to those of existing methods.
Critical weaknesses
  1. While reading the paper, I observed that its design choices are not well justified. As the objective function consists of multiple components, including objective functions for vessel prediction, region prediction, etc., it is important that the significance of each component is motivated and explained properly. Moreover, the hypothesis also needs to be supported by rigorous experiments, which this paper lacks.

  2. I also observed that the related works are quite narrow. Additionally, the references are not properly cited in the table summarizing the performance comparison (please see Table 4).

  3. The pipeline is evaluated on only a single clinical problem. Therefore, it is also hard to tell how the proposed method would generalize to other domains. As ICLR is a venue of broader interest in AI and machine learning, I am unsure how large the audience at ICLR interested in this paper will be.

  4. The diagrams are also unclear; in particular, the fonts are not clear. Moreover, mapping important components in the figures to the equations in the paper would enhance the readability of the paper.

  5. There are no ablation studies for the individual objective functions. Without such studies, it is difficult to understand the role of each component in the pipeline, and it would also be hard to draw any conclusions.

Questions for authors
  1. How does each component of the objective function impact overall performance?
  2. Does this pipeline also generalize to datasets other than X-ray angiography datasets?
Reviewer_NTVqOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how reconstruction quality and structural auxiliary supervision affect coronary vessel segmentation. It introduces CVRD, which combines local detail extraction, contextual fusion, and staged recovery. Experiments cover three datasets. Component ablations support the reconstruction design. Matched continuation controls show small auxiliary gains in IoU, but mixed connectivity results. A repeated-occupancy control finds no clear geometry-specific advantage in the tested setting.

Critical strengths
  1. Robustness is good: The model demonstrated excellent stability across three datasets and three random seeds. Coupled with the results from external benchmark tests, it indeed proves that this model is competitive in the segmentation task.
  2. Experimental design is targeted: The control experiment for auxiliary supervision was designed quite ingeniously, enabling a good separation of the actual contribution of this module.
Critical weaknesses
  1. The description of data division is ambiguous: I have some doubts about the way the data set was divided. The article did not clearly state whether the test set truly divided according to patients (patient-level) or independently according to sequences to strictly avoid data leakage. I hope the author can clearly add this point.
  2. Insufficient verification of specific comparative experiments: Regarding the comparison between geometric labels and ordinary vascular labels, it seems that only tests have been conducted on the validation set of the XACV dataset and a single random seed. The lack of cross-validation on independent test sets or other datasets makes the generalization ability of this part of the conclusion seem a bit weak.
Questions for authors
  1. In the ablation experiments, what is the specific baseline structure used to replace the LPDE and FSR modules? After the replacement, how much difference is there in the number of parameters and computational cost of the model compared to the original modules?
  2. In the introduction section, can the author further elaborate on what new insights or different ideas the paper provides in the design of the "reconstruction process" compared to existing high-resolution vessel segmentation methods?
  3. In the Experimental Section, please clearly state whether the training set, validation set, and test set are strictly isolated by patient ID or video sequence?
  4. In the Experimental Section, Have all the baseline models been adequately trained to converge? Will different training settings affect the performance comparisons shown in Table 4?
Reviewer_XJQUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses an important problem in vessel segmentation. However, the proposed method offers limited novelty and is not sufficiently differentiated from existing segmentation approaches. Moreover, the experiments do not adequately demonstrate improvements in vessel continuity, which is central to the paper's motivation. The small test sets further limit the reliability and generalizability of the reported results, making the conclusions less convincing.

Critical strengths
  1. The motivation is reasonable, particularly the focus on preserving fine vessel details and improving structural continuity.
  2. The paper provides ablation studies to analyze the contributions of different components. It also includes controlled experiments to assess the benefits of structure-aware supervision, helping clarify which factors affect segmentation performance.
Critical weaknesses
  1. The paper assumes that reconstruction quality is a key factor in reliable vessel recovery. However, this hypothesis is mainly evaluated through ablation studies of individual CVRD components. These experiments show only the contributions of these components within the proposed architecture, but they do not sufficiently establish that reconstruction matters more than structure-aware learning.
  2. The proposed method mainly combines existing segmentation techniques. The authors should clarify the key technical contributions and explain how the proposed components fundamentally differ from established segmentation modules.
  3. The paper emphasizes reliable vessel recovery, particularly the continuity of thin vessels. However, the current evaluation does not directly demonstrate improvements in vessel connectivity. The paper needs additional metrics that explicitly measure vessel continuity, such as branch disconnection and connectivity errors, to support its main claims.
  4. The proposed method uses a pretrained CLIP encoder, while the baseline models differ in their architectures and training configurations. It is therefore difficult to determine whether the improvements come from the proposed decoder or the stronger encoder. The authors should compare the proposed decoder against standard segmentation decoders using the same CLIP encoder to ensure a fair evaluation.
  5. The test sets are relatively small, which limits the reliability of the reported improvements. Moreover, the experiments are restricted to coronary angiography datasets, with no cross-dataset evaluation. Additional experiments are needed to demonstrate the proposed method's robustness and generalizability.
Questions for authors
  1. The paper assumes that reconstruction quality is a key factor in reliable vessel recovery. However, the current ablation studies mainly show the contributions of individual CVRD components. How do these results support the claim that reconstruction is more important than structure-aware learning?
  2. What are the key technical contributions of the proposed method? How are the proposed modules fundamentally different from existing segmentation techniques?
  3. The paper focuses on improving vessel continuity. Although clDice and other structural metrics are reported, the experiments do not clearly show whether vessel disconnections are actually reduced. Could the authors provide more direct evaluations of vessel connectivity, such as branch disconnection errors or branch recovery rates?
  4. The proposed method uses a pretrained CLIP encoder, while the baselines use different architectures and training configurations. How can the authors ensure a fair comparison? Could they compare CVRD with standard segmentation decoders using the same CLIP encoder?
  5. The test sets are relatively small, making it difficult to assess the reliability and generalizability of the results. Could the authors provide additional experiments, particularly cross-dataset evaluations, to support their conclusions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.