List of Papers Browse by Subject Areas Author List
Abstract
Vascular graph extraction converts medical images into structured representations of vessel connectivity, enabling quantitative analysis of branching topology. However, training image-to-graph models requires expert-annotated vascular graphs, creating a substantial annotation bottleneck. Recent work leverages large non-medical datasets for structural pretraining, yet the structural alignment between source and vascular domains remains underexplored. We introduce a nature-to-nature pretraining paradigm that uses botanical branching networks as a biologically aligned source domain for vascular graph extraction. Leveraging large-scale annotated grapevine skeleton graphs and geometry-preserving graph representations, our approach promotes relational priors consistent with vascular growth and connectivity.
Across heterogeneous 2D and 3D benchmarks, botanical pretraining consistently outperforms infrastructure-based pretraining under identical training settings. In addition, biologically aligned supervision achieves comparable performance with substantially fewer annotated vascular volumes. Multi-domain training further enhances robustness to previously unseen vascular datasets. These findings demonstrate that structural congruence between source and target domains provides a stronger inductive bias for vascular graph extraction than scale alone.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2313_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/erc-caravel/Vascular-Graph-Extraction.git
Link to the Dataset(s)
See Paper
BibTex
@InProceedings{ScaAle_Structural_MICCAI2026,
author = { Scavone, Alessandro AND Borrego, Lucia AND Marcianò, Vincenzo AND Mata, Christian AND Giraldo, Beatriz AND Munuera, Josep AND Zuluaga, Maria A.},
title = { { Structural Congruence Matters: Biological Pretraining for Vascular Graph Extraction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper studies cross-domain pretraining for vascular graph extraction from medical images. Building on an existing RelationFormer-based framework, it proposes a biologically aligned “nature-to-nature” pretraining strategy that uses botanical branching graphs instead of road networks as the structural source domain, together with graph preprocessing choices such as border-aware edge clipping and degree-2 node retention.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1: The paper studies a relevant problem. Reducing the need for expert-annotated vascular graphs is an important problem for image-to-graph learning in medical imaging. 2: The main hypothesis is at least clear and testable. The paper explicitly asks whether a biologically aligned source domain (plants) transfers better than road networks for vascular graph extraction, and evaluates this under matched source size in part of Table 1.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1: The methodological novelty is limited. The paper does not introduce a new image-to-graph architecture or learning objective. It largely reuses RelationFormer and the cross-domain / cross-dimensional training framework of Berger et al. , while replacing the pretraining source domain and adding preprocessing choices such as border-aware clipping and degree-2 node retention. This is a relatively small step beyond prior work. 2: The central claim is stronger than the evidence provided. The paper argues that “structural congruence matters more than scale alone,” but the evidence is only a comparison between roads and plants within one inherited framework. This does not establish that structural congruence is the main causal factor, since the reported gains are entangled with representation and preprocessing changes, especially the retention of degree-2 nodes. 3: Important baselines are missing. The experiments compare mainly against no pretraining and road-based pretraining. There is no comparison to stronger medically relevant pretraining strategies, despite the paper explicitly discussing medical foundation models and “purely semantic medical pretraining. ” As written, the evaluation is insufficient to support broad claims about the superiority of the proposed pretraining paradigm. 4: No ablation study. The experimental design is not sufficiently clean to isolate the main contribution of different components.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a relevant problem, namely reducing the need for expert-annotated vascular graphs for image-to-graph learning. However, I recommend rejection because the methodological novelty is limited: the submission largely builds on existing RelationFormer-based cross-domain training and mainly changes the pretraining source domain and graph preprocessing choices, rather than introducing a substantively new model or learning formulation. In addition, the experimental evidence does not fully support the strength of the paper’s claims. Important baselines are missing, particularly stronger medically relevant pretraining alternatives, and the current design does not cleanly isolate the contribution of structural source alignment from that of graph representation/preprocessing choices such as degree-2 node retention. While the paper includes experiments, the overall contribution remains incremental and falls below the expected bar for MICCAI.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
(1) The authors introduce a “nature-to-nature” pretraining framework that utilizes botanical branching networks (e.g., grapevine skeletons) as a biologically aligned source domain for vascular graph extraction.
(2) The work demonstrates that structural alignment between the source and target domains—specifically shared principles like hierarchical branching and flow optimization—provides a stronger inductive bias than data scale alone.
(3) It leverages a 2D-to-3D rendering technique to allow structural knowledge from 2D botanical graphs to be transferred to 3D volumetric vascular targets.
(4) The method shows superior sample efficiency, achieving competitive results with significantly less annotated data, and improves zero-shot generalization through multi-domain training.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
In my view, this is a truly surprising and inspiring paper. There is a certain sense of beauty in the experimental findings, particularly in how they reveal the deep connections between disparate biological systems.The conceptual innovation here is profound: it demonstrates that biological transport systems provide far superior priors for vascular extraction than human-made infrastructure like road networks. This suggests that nature-made structures—including our own vascular systems—share a latent generative logic with the botanical world. By aligning our models with these shared principles of growth and flow efficiency, the authors have found a way to bridge the gap between scale and structural truth.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Even though I enjoy reading this paper, I would like to respectfully note some weaknesses.
The evaluation seems primarily restricted to the RelationFormer architecture. It remains unclear if these biological priors offer similar gains for other image-to-graph frameworks or traditional segmentation-skeletonization pipelines.
While grapevines are used effectively, the study only explores one type of botanical data. Testing on other biological networks (e.g., leaf veins or fungal hyphae) would further validate the “structural congruence” hypothesis.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(6) Strong Accept — must be accepted due to excellence
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Interesting findings with a sense of natural beauty.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper demonstrates that pretraining on biologically aligned structural priors improves the accuracy and efficiency of 2D and 3D vascular segmentation. The authors have properly explained and experimentally validated, how the structural congruence between the source and target domains for cross domain transfer matters a lot and that the plant transportation system governed by similar mechanisms(hierarchical branching, flow optimization, and space-filling organization) are more aligned with human vascular system then roads or other man-made structures which are governed by organization and city planning.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has the following the strengths:
-
The paper discusses a highly novel perspective on the problem formulation of image to graph extraction for vascular networks. The author have provided in depth intuitive explanation, how the structural congruence between human vascular system and the plant transport networks are driven by similar mechanisms(hierarchical branching, flow optimization, and space-filling organization).
-
The authors have experimentally demonstrated that how this structural congruence between source domain used for pretraining and target domain could lead to better results even within limited data regimes under pre-training, fine-tuning and zero-shot transfer setups.
-
The findings of this paper align with broader trends in the medical deep learning community, which suggest that pretraining on high-quality, structurally aligned data is often more effective than pretraining on large-scale datasets from highly different domains. These results further highlight the importance of domain relevance in transfer learning. The findings of this paper could also open up opportunities for future research in this direction and encourage further discussion on domain-aligned pretraining strategies.
-
The preprocessing strategies adopted, effectively preserve topological consistency and geometric detail. In particular, border-aware edge clipping avoids boundary artifacts, while degree-2 node retention maintains curvature information critical for accurate modeling. Together, these choices provide more reliable and informative supervision for downstream learning.
-
The authors have provided code, enabling the proposed approach to be readily adopted for developing more accurate and efficient single-stage image-to-graph extraction pipelines, facilitating its application in zero-shot settings for vascular graph extraction.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper has the following major weaknesses:
- While the paper proposes a highly interesting pre-training and pre-processing strategies, the contributions in terms of overall architecture, loss functions etc. seem limited. The authors have largely used the architecture and hyperparameter choices of Berger et. al and RelationFormer.
Shit, S. et al. (2022). Relationformer: A Unified Framework for Image-to-Graph Generation. In: Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T. (eds) Computer Vision – ECCV 2022.ECCV 2022.Lecture Notes in Computer Science, vol 13697.Springer, Cham. https://doi.org/10.1007/978-3-031-19836-6_24
Berger, Alexander H., et al. “Cross-domain and cross-dimension learning for image-to-graph transformers.” Proceedings of the Winter Conference on Applications of Computer Vision. 2025.
-
The authors shall summarize their contributions properly so that it is easy for the audience to follow through and understand the scope of the paper and what value it adds to the existing literature.
-
“We build upon the RelationFormer architecture [22] within the cross-domain and cross-dimensional training framework of Berger et al. [3]. The model consists of a convolutional backbone followed by a transformer encoder–decoder that jointly predicts node coordinates and pairwise connectivity in a single stage. We adopt the full training strategy of [3], including their hybrid loss formulation, regularized edge sampling, and adversarial domain adaptation modules. All architectural components and optimization settings follow the original implementation.”
This seems like a very simplified and shallow version for the methodology, even if the authors adopted their architecture and hyperparameters from other setups, it would be helpful for the readers if the authors can briefly discuss about the approach in their paper as well.
-
“Nodes are matched using fixed-radius tolerance, and edges are considered correct when they connect correctly matched node pairs.” Kindly add more details on this like the amount of tolerance used as well as the providing the formulae for some less commonly known metrics like SMD would be helpful
-
Figure 3 seem a little bit confusing and overcluttered. The authors may consider using another representation(like bar graph) which may be more easy to interpret for these results
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
-
The authors could also add a few visual examples to illustrate the structural similarities and congruence between the human vessels and the grapevine using both raw images and extracted graphs to further strengthen their claim and make it easy for readers to relate with the same
-
The authors can consider trying to make the overall methodology figure(Figure 1) by including more detailed representation of architecture as well as stage-wise pre-processing for the audience to follow through it better
-
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a novel formulation of the problem and the findings of the paper could be useful for the overall medical community, motivating further research on the effects of domain congruence for large scale pre-training and transfer learning. Also, the proposed method could be directly adopted for many downstream use cases which rely on vascular graph extraction.
However, there are certain concerns(see weaknesses section), the authors shall try justifying the same in rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank the meta-reviewer and all reviewers for their careful evaluation and constructive feedback. We are particularly encouraged that multiple reviewers recognized the relevance of investigating structural congruence for vascular graph extraction and appreciated the controlled experimental design, especially the scale-matched comparisons and sample-efficiency analysis.
Our intention in this work was not to introduce a new image-to-graph architecture, but to isolate the role of biologically aligned structural priors under controlled conditions. For this reason, we intentionally retained the RelationFormer/Berger framework, losses, and optimization settings across all experiments. We believe this strengthens the comparison by avoiding confounding architectural modifications.
Regarding the concern that structural congruence is entangled with preprocessing, Table 1 explicitly controls for these factors. Roads are evaluated at both 99.2k and 25.9k source samples, and degree-2 retention is ablated for both roads and plants. Thus, roads+2-degree vs. plants+2-degree isolates the effect of source domain under matched graph representation settings. Under these controlled conditions, plant-based pretraining consistently outperforms infrastructure-based pretraining in both 2D and 3D.
We also acknowledge the suggestion regarding broader medical pretraining baselines. Existing medical foundation models such as vesselFM target dense segmentation rather than image-to-graph prediction and do not provide graph-level supervision compatible with RelationFormer-style graph extraction. We therefore compare against the most directly relevant prior work, Berger et al.’s road-based structural pretraining framework.
We agree that extending the analysis to additional biological transport systems and other graph extraction paradigms beyond RelationFormer would be valuable future work. Due to space constraints and the goal of maintaining a controlled study, we focus this MICCAI submission on the current setting, while considering broader evaluations for a future journal extension. In the camera-ready version, we have further clarified the methodology and improved the figures to make the experimental design and contributions easier to follow.
We thank the reviewers again for their insightful comments.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
After careful consideration, I find the core idea – that plant branching networks transfer better than road networks for vascular graph extraction – genuinely novel and thought-provoking. Honestly, I think the paper is good for MICCAI. The paper is well written and the findings are interesting enough to warrant publication.
That said, the paper has clear limitations. The methodological contribution is incremental, as the architecture and training framework are largely borrowed from prior work. The experimental evidence does not fully support the claims, with no ablation study to isolate the contribution of structural alignment from preprocessing changes. Missing baselines further weaken the evaluation. However, the conceptual insight is strong enough that I lean toward acceptance rather than rejection. If the authors integrate the comments given by the reviewers, the paper would be significantly stronger.
