List of Papers Browse by Subject Areas Author List
Abstract
Vessel segmentation is crucial for analyzing brain vasculature and understanding cerebral functions and disease mechanisms. Current deep-learning models for segmenting blood vessels within brain images are supervised and depend on extensive labeled data, which requires expert annotation and is both time-consuming and resource-intensive. To address these challenges, we propose Vessel-Dictionary Selection Net (V-DiSNet), a one-shot active learning (OSAL) framework specifically designed for vessels that can be used to select a small, representative set of informative and diverse samples for expert annotation and training, given an unlabeled dataset in a single iteration. The selection process involves sampling from a latent space designed by leveraging the recurrent properties of brain vessel patterns. Specifically, we combine dictionary learning with k-means clustering to learn a latent representation integrating fundamental basis elements representing recurrent vessel features such as shape, connectivity, and structures. We experimentally demonstrate the effectiveness of our method on three publicly available 3D Magnetic Resonance Angiography datasets, showing that V-DisNet consistently outperforms random sampling and other state-of-the-art OSAL methods in terms of standard vessel segmentation metrics. Our code is available at github.com/i-vesseg/V-DiSNet.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2457_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/i-vesseg/V-DiSNet
Link to the Dataset(s)
IXI Dataset: https://brain-development.org/ixi-dataset/
OASIS-3 Dataset: https://sites.wustl.edu/oasisbrains/home/oasis-3/
SMILE-UHURA Dataset: https://www.synapse.org/Synapse:syn47164761/wiki/620033
CAS Dataset: https://codalab.lisn.upsaclay.fr/competitions/9804
BibTex
@InProceedings{FalDan_Oneshot_MICCAI2025,
author = { Falcetta, Daniele and Chaptoukaev, Hava and Galati, Francesco and Zuluaga, Maria A.},
title = { { One-shot active learning for vessel segmentation } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15965},
month = {September},
page = {478 -- 488}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper introduces a pipeline for one-shot active learning applied to vessel segmentation. The proposed approach begins by applying compressed sensing techniques to a labeled dataset to learn a dictionary. Subsequently, K-means clustering is performed on the resulting sparse latent vectors. The authors then employ supervised contrastive learning to train a Siamese network, using the same labeled dataset, where images serve as inputs and clustering labels determine the positive and negative pairs. Finally, the trained Siamese network is used to process an unlabeled dataset, from which samples are selected using K-means clustering and farthest point sampling (FPS).
- 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 experimental setup is well-detailed. The experiments are comprehensive and clearly presented, with multiple experiments conducted using different random seeds to ensure robustness. Figures 3 and 4 are intuitive and effectively illustrate the benefits of the dictionary learning and contrastive learning components.
- 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 proposed method consists of several components, but it remains unclear how each contributes individually to the final performance. For instance, it is not evident whether supervised contrastive learning is essential, or if a self-supervised variant could achieve similar results. The current ablation study does not explore this possibility. While the paper emphasizes a one-shot active learning framework, neither the proposed method nor the baselines are inherently restricted to a single labeling round. The paper does not explore whether using multiple rounds with the same total labeling budget could lead to different performance.
- Please rate the clarity and organization of this paper
Satisfactory
- 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/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.
(4) Weak Accept — could be accepted, 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 solid approach but its impact is limited by the absence of key ablation studies.
- Reviewer confidence
Not confident (1)
- [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
Authors introduce a vessel-specific one-shot active learning framework (V-DiSNet) that uses dictionary learning to identify and cluster recurring anatomical patterns, enabling a single-pass selection of highly informative training patches for brain vessel segmentation.
- 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.
-
Novel and Interpretable Representation for Vascular Structures The authors propose a dictionary-learning approach to detect and cluster recurrent vessel patterns. This representation is both novel and interpretable, setting it apart from existing one-shot active learning methods.
-
Fair Validation The method is evaluated on three public datasets—CAS, OASIS-3, and SMILE-UHURA—and demonstrates relatively good performance than random selection, achieving segmentation results with only 30% labeled data on the CAS dataset. Each experiment is repeated 8 times with different random seeds to ensure robustness. Ablation studies
-
Clear Presentation The pipeline is well-structured, and the figures effectively support the methodology. Each component of the framework is clearly described and easy to follow.
-
- 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.
-
Patch-Based Segmentation Limits Global Context: The proposed method relies entirely on patch-level sampling, but the paper does not address how it preserves global vessel continuity or context. This omission raises concerns about the method’s ability to capture long-range vessel structures.
-
Limited Dictionary Dimension (2D vs. 3D): The method processes 2D patches, which simplifies computation and enables sparse dictionary encoding. However, recent advances in brain vessel segmentation increasingly employ 3D models to better capture structural continuity in volumetric data [1–3]. Acknowledging this limitation and discussing the potential extension to 3D would strengthen the paper.
-
Incomplete Baseline Comparisons for OASIS-3 and SMILE-UHURA: In Fig. 2, the proposed method is only compared with random sampling. Yet, as Table 1 shows, random sampling consistently underperforms relative to other baselines. Including stronger baselines (e.g., RA, CA, AET) for these datasets would provide a more convincing evaluation.
[1] Xia, Likun, et al. “3D vessel-like structure segmentation in medical images by an edge-reinforced network.” Medical Image Analysis 82 (2022): 102581. [2] Chen, Huai, et al. “3D vessel segmentation with limited guidance of 2D structure-agnostic vessel annotations.” IEEE Journal of Biomedical and Health Informatics (2024). [3] Deng, Zhiwei, et al. “Shape-aware 3D small vessel segmentation with local contrast guided attention.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2023.
-
- 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 does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
In Section 3, Segmentation Performances, the authors claim that V-DiSNet consistently outperforms competitors across all fractions. However, in the third row (5% labeled) of Table 1, the proposed method shows no superior performance compared to other SOTAs.
- 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 — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The novelty of the proposed OSAL framework with its dictionary-based latent space, along with the clear structure of the paper and the authors’ effort to interpret the latent space, contributed to my overall positive score.
- 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
This paper introduces an automated Active Learning (AL) framework specialized for vessel segmentation. The authors implement their framework using dictionary learning to identify key components of the vascular tree and use the concept of dictionary learning to learn the sparse representation of the vascular tree leading to increased efficiency of the AL technique to minimize the external labeling effort.
- 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 authors give a great background to the subject and the reader can easily see their motivations and how this technique could lead to a streamlined implementation of active learning. The figure describing the pipeline of the paper was much appreciated since this is a relatively novel technique utilizing many different technologies. The reader can appreciate the concrete improvements in the one shot active learning framework with the takeaway of only needing to annotate 30% of the data for the optimal performance. The interpretability aspect using dictionary learning is also appreciated since many deep learning techniques might not be interpretable.
- 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 only principle weakness in the dictionary learning methodology is that the patch sizes used are usually in the 2 dimensional space while the medical images are in 3 dimensions. However, the reader finds that the authors did implement the model correctly despite the inherent limitations of the dictionary learning framework for 3 dimensional analysis.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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/2025/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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The reader feels that the novelty in incorporating dictionary learning for active learning shows great theoretical promise as there is evidence for sparse representations useful for vessel analysis with earlier works on Kwitt, R., Pace, D., Niethammer, M., Aylward, S. (2013). Studying Cerebral Vasculature Using Structure Proximity and Graph Kernels MICCAI 2013 https://doi.org/10.1007/978-3-642-40763-5_66. This work is an innovative approach as dictionary learning and sparsity approaches are interpretable since there is a sparse representation in this case, the fractal nature of vessel structures leads itself to a sparse analysis since the topology has some sort of scale invariance that can be used to help the active learning.
- 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 all reviewers (R1,R2,R3) for their constructive feedback.
We are pleased that V-DiSNet was well-received, recognizing our novel approach to vessel segmentation using dictionary learning, comprehensive experimental validation, and model interpretability. Our primary contribution lies in introducing a novel framework that leverages dictionary learning for one-shot active learning in vessel segmentation. The interpretability aspect of our method, which all reviewers appreciated, offers potential benefits beyond performance metrics alone—it provides insights into the underlying anatomical patterns that guide the selection process.
Below, we provide answers to the reviewer’s comments:
-On ablation studies and component contributions (R1): We agree that evaluating individual components would strengthen our work. While we couldn’t include this due to space constraints, we acknowledge the limitation. We plan to analyze each component’s contribution in an extended version of this work, particularly comparing supervised vs. self-supervised contrastive learning as suggested.
-On multi-round vs. one-shot AL comparison: R1 raises an important point about comparing one-shot vs. classical multi-round AL with the same labeling budget. Our focus was primarily on minimizing annotations with a single-pass approach, but this comparison would be valuable and should be included in an extended (journal) version.
-On 2D vs. 3D analysis limitations (R2, R3): We acknowledge the inherent limitations of our 2D patch-based approach for 3D vessel structures. We plan to extend V-DiSNet to operate directly on 3D patches to capture volumetric vessel continuity better while maintaining the benefits of our dictionary learning approach. This extension would involve adapting our dictionary learning framework to capture 3D vessel patterns and modifying the latent space construction accordingly. However, this may introduce significant computational requirements, as 3D dictionary learning typically requires larger memory footprints and longer training times.
-On global vessel connectivity (R3): We acknowledge the concern about preserving global vessel continuity in a patch-based approach. While our current method uses clDice to verify vessel continuity and overlapping patches at inference to partially address boundary issues, incorporating global context remains challenging. Future work will explore multi-scale, multi-view, and 3D-patch approaches to maintaining vessel topology across patches.
-On incomplete baseline comparisons: R3 correctly notes that we only compared against random sampling for OASIS-3 and SMILE-UHURA datasets. This was primarily due to computational constraints for the comprehensive 8-seed experiments. An extended version of this work plans to extend benchmarks to more datasets and methods.
-On the claim about 5% labeled data (R3): We thank the reviewer for this careful observation. We will correct our claim to more accurately reflect that while V-DiSNet generally outperforms competitors across most fractions, performance at 5% labeled data shows comparable results to CA rather than clear superiority.
-We also appreciate R2’s suggestion to cite Kwitt et al. (2013), which represents relevant prior work on vessel structure analysis using proximity and graph kernels. We will incorporate this reference in the final version to properly acknowledge the theoretical foundations that complement our approach.
Looking forward, we are excited about the possibility of discussing V-DiSNet further with the MICCAI community. Thank you again for your valuable feedback, which will enrich the extensions of our work and help guide our ongoing research. Addressing the points raised by the reviewers will not only improve the performance of our method but also contribute to the broader understanding of efficient vessel segmentation approaches.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
The manuscript introduces V-DiSNet, a one-shot active-learning pipeline that leverages dictionary learning and contrastive Siamese training for segmentation.
The paper has received two weak accept and one accept. Reviewers appreciate the clear experimental design, robust multi-seed validation on three public datasets, and the interpretability afforded by dictionary-based representations. The consensus about the merits of this paper leads to a decision of provisional accept.
Reviewers also mentioned several recommendations to be incorporated in the next revision by (1) adapting the ablation study to isolate the impact of supervised versus self-supervised contrastive learning, (2) evaluating whether multiple sampling rounds with the same labeling budget yield further gains, and (3) discussing the 2D patch limitation versus potential 3D extensions, and (4) considering stronger baselines. Some of the points should be relatively straightforward to address and may further articulate the added value of this paper.