List of Papers Browse by Subject Areas Author List
Abstract
Accurate guidewire segmentation is essential for safe navigation in fluoroscopy-guided interventions but remains challenging in Digital Subtraction Angiography (DSA) images. Beyond pixel-level difficulties, guidewires are tightly constrained by vascular anatomy, a dependency often ignored by existing methods that segment them independently. To address this problem, we propose RDE (Role-Disentangled Experts), a framework combining parameter-efficient domain specialization, role-disentangled decoding, and conservative residual fusion. We instantiate RDE on the Segment Anything Model (SAM), resulting in RDE-SAM. RDE-SAM injects LoRA into the image encoder and role-disentangled experts with residual routing into the decoder. We further incorporate a lightweight anatomy-constrained refinement that uses the predicted vessel mask to define a tolerance band and repair small guidewire discontinuities, improving structural consistency. On a clinically collected DSA dataset of over 3,000 images with dual pixel-level annotations, RDE-SAM consistently outperforms strong baselines, demonstrating the benefit of explicitly modeling anatomical constraints for thin-structure instrument segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2147_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/singer502/RDE-Seg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaYin_RDESeg_MICCAI2026,
author = { Zhang, Yining AND Long, Chengjiang AND Xu, Wenju AND Zhang, Tingbao AND Feng, Yu AND Zhao, Jianhui},
title = { { RDE-Seg: Role-Disentangled Experts with Residual Routing and Anatomy Constraints for DSA Guidewire Segmentation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes RDE-Seg, an architecture-agnostic framework for anatomy-constrained instrument segmentation, instantiated on SAM for guidewire segmentation in DSA images. The key idea is to explicitly model the dependency between guidewires and vessels through role-disentangled expert decoders and a residual routing mechanism. In addition, a lightweight anatomy-constrained refinement module is introduced to enforce spatial feasibility and improve structural continuity. The method is evaluated on a newly collected dual-annotated DSA dataset.
- 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．Well-motivated problem formulation. The paper addresses the lack of explicit anatomical constraints in guidewire segmentation and frames it as structured prediction. 2．Simple yet effective architecture design. The expert decoders with residual routing provide a practical way to introduce specialization while keeping training stable. 3．Clinical relevance and reasonable validation. The method shows consistent improvements, especially in structural metrics (e. g. , HD95), on a clinically meaningful DSA dataset.
- 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.Limited expressiveness in routing and robustness concerns. The expert-routing design relies on coarse image-level gating, while the anatomy refinement depends on vessel predictions and may propagate errors; endpoint-bridging is under-specified and risks hallucinated connections. 2.Insufficient ablation and analysis. The anatomy-constrained refinement module lacks quantitative ablation and failure case analysis. 3.Unclear data splitting and reproducibility. Patient-level split, implementation details, and some hyperparameters are not clearly specified, affecting reproducibility.
- 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.
- 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.
(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?
This paper addresses a clinically relevant problem and presents a technically sound method with improved structural performance. However, the novelty is somewhat limited, and key aspects such as the refinement module and evaluation protocol require further clarification. Strengthening these points would improve the paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal clarifies several implementation details and partially addresses concerns regarding the training strategy and anatomy-constrained refinement. In particular, the authors provided additional explanations about the progressive optimization schedule and the role of residual routing, which improves the overall readability of the method.
I also appreciate the clarification regarding the dataset construction and the additional discussion on structural consistency improvements. The response helps better justify the motivation behind anatomy-constrained guidewire segmentation.
However, some concerns remain only partially resolved. The experimental scope is still relatively limited to a single collected cohort, and the rebuttal does not fully address the question of cross-site or cross-device generalization. In addition, although the authors discussed the refinement module in more detail, the paper would still benefit from stronger quantitative ablations and failure-case analysis, especially regarding endpoint bridging and robustness to vessel prediction errors.
I also believe comparisons with methods explicitly modeling vessel-instrument relationships would further strengthen the empirical validation.
Overall, the rebuttal improves the clarity of the submission and addresses part of the reproducibility concerns, but several methodological and experimental limitations remain. Nevertheless, I still find the paper technically sound and clinically relevant, and I keep my original recommendation.
Review #2
- Please describe the contribution of the paper
This paper addresses the problem of guidewire segmentation in DSA images as an anatomically constrained structured prediction task and proposes the RDE (Role-Disentangled Experts) framework. This framework organically integrates three modules: parameter-efficient domain adaptation (LoRA), role-disentangled expert decoders, and conservative residual routing. It is instantiated as RDE-SAM based on SAM, explicitly modeling the anatomical dependencies between the guidewire and vessels. By combining joint segmentation with tolerance band constrained refinement, it improves the spatial合理性 and topological continuity of guidewire segmentation. Furthermore, the authors construct a dual-annotated dataset comprising 3,069 clinical DSA images with pixel-level annotations for both guidewire and vessels, providing a valuable benchmark resource for research on anatomical constraint modeling.
- 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) Existing methods generally treat the guidewire as an independent fine-structure segmentation target. This paper redefines it as an anatomically constrained structured prediction problem. This shift in perspective has strong clinical justification, with the 99.7% guidewire-vessel overlap rate in the dataset robustly supporting the modeling motivation, and the problem definition maintains high consistency with the methodological design. (2) Expert predictions are injected into the global branch in a residual manner (with α gradually increasing in Equation 3), ensuring that performance in the worst case is not inferior to the baseline. Ablation studies clearly show that introducing the Router significantly reduces HD95 from 8.32 to 2.98.This design achieves effective expert refinement while maintaining training stability, distinguishing it from the hard routing approach of traditional MoE. (3) Generalization experiments validate the effectiveness of RDE on two distinct backbone networks, UNet and SAM, demonstrating that the core ideas of role disentanglement and conservative residual fusion do not depend on a specific foundation model and possess strong generality and transferability. (4) The three-stage progressive training strategy effectively avoids optimization instability issues caused by increased model capacity. The design logic from Phase I to Phase III is clear, complementing the residual fusion mechanism, and demonstrates strong engineering practicality.
- 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 data source involves only about 20 patients, representing a relatively small patient cohort. The generalizability to different angiographic sites (e.g., intracranial arteries, peripheral vessels) and different acquisition devices has not been validated, limiting the universality of the conclusions. (2) The paper lacks comparisons with other methods that explicitly model vessel-instrument relationships (e.g., the cited multi-task segmentation methods). Existing baselines all treat the guidewire as an independent target, making it difficult to fully assess the incremental value of the core contribution—anatomical constraint modeling—relative to similar approaches. (3) In the anatomical constraint refinement module, the dilation radius r of the tolerance band and the endpoint bridging distance threshold dmax are key hyperparameters. The paper does not provide a sensitivity analysis, leaving the robustness of these parameters under varying vessel morphologies and image qualities in doubt. (4) The average inference time per image is 176.23 ms (RTX 4090, batch size 1), which still lags behind the requirements for real-time interventional procedures. The paper does not sufficiently discuss this issue nor provide optimization directions for real-time performance.
- 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.
- 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.
(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?
(1) Redefining guidewire segmentation as an anatomically constrained structured prediction problem constitutes a substantial contribution. The 99.7% guidewire-vessel overlap rate provides strong data support, and this modeling perspective is consistently reflected throughout all aspects of the method design. The problem definition and technical solution are logically coherent. (2) The ablation study results are compelling, with a clear hierarchical contribution of each component: HD95 decreases from the SAM+LoRA baseline of 9.25 to 8.32 with the expert branch, and further to 2.98 with residual routing. Compared to the strong baseline H-SAM, the HD95 improvement exceeds 56%, demonstrating that anatomical constraint modeling brings substantial structural accuracy improvements rather than marginal gains. (3) The consistent effectiveness on both UNet and SAM architectures, along with the construction of the dual-annotated clinical dataset, elevates the contribution of this paper beyond single model design, providing independent reference value for subsequent research in the field. (4) Despite limitations such as limited data scale and lack of some hyperparameter analyses, these issues do not undermine the validity of the core contributions. The overall design of the method is complete, and the experimental validation is sufficient.
- 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
1) Proposing RDE, a general and architecture-agnostic framework for anatomy-constrained instrument segmentation, integrating LoRA-based domain specialization, role-disentangled decoding, and conservative residual routing. 2) Modeling the anatomical dependency between guidewires and vessels through joint segmentation and constraint-aware refinement, improving structural continuity and spatial plausibility. 3) Constructing a dual-annotation DSA dataset to validate anatomy-constrained modeling and demonstrate the effectiveness of RDE-Seg.
- 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.
First, it is the first work to explicitly introduce vascular anatomical constraints into the guidewire segmentation task, overcoming the limitation of existing methods that treat guidewires as independent targets while ignoring their strong dependency on vascular anatomy. Second, the paper proposes the Role-Disentangled Experts (RDE) framework, which combines parameter-efficient domain specialization, role-disentangled decoding, and conservative residual fusion. By instantiating RDE on the Segment Anything Model (SAM), the resulting RDE-SAM injects LoRA into the image encoder and introduces role-disentangled experts with residual routing into the decoder, achieving efficient and specialized segmentation. Third, the paper designs a lightweight anatomy-constrained refinement module that uses the predicted vessel mask to define a tolerance band and repairs small discontinuities in the guidewire, significantly improving the continuity and structural consistency of thin-structure segmentation. Finally, on a clinically collected DSA dataset comprising over 3,000 images with dual pixel-level annotations, RDE-SAM consistently outperforms multiple strong baselines, fully demonstrating the effectiveness and clinical value of explicitly modeling anatomical constraints for thin-structure instrument segmentation.
- 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) Formatting issues. In Table 1 on Page 8, there is a horizontal line in the entries for “SAMed (Preprint 2023)” and “FGA-Net (CMIG 2025)”, and it is unclear what this line represents. Additionally, the text in Fig. 1 is too densely packed, which impairs readability. 2) Unclear model details. In Section 2.3, \mathbf{L}{global}, \mathbf{L}{wire}, and \mathbf{L}{vessel} should be generated by the global expert, guidewire expert, and vessel expert, respectively. The architecture of these three expert networks is not specified. Are the network structures of the global expert and the structure-specific experts (guidewire and vessel) different? 3) Insufficient experimental data. The paper only evaluates the proposed method on a single self-constructed DSA dataset, which limits the persuasiveness of its generalizability and robustness. It is recommended to conduct experiments on additional datasets. 4) Incomplete experimental design. The paper lacks hyperparameter analysis for the dilation rate r and the maximum distance d{max} introduced in Section 2.4, and does not specify the appropriate values for these hyperparameters. Furthermore, although a progressive conservative training strategy is described in Section 2.5, no experiments are provided to demonstrate whether this strategy actually brings performance gains to the model.
- 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.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The manuscript requires major revisions. Acceptance may be considered only if the authors can adequately address the above concerns.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank the Reviewers for their constructive feedback and recognition of our work. We address the main concerns below. Shared question: Sensitivity of r and d_max We set r=8 to match the average guidewire width (6.23 px), and d_max=20 to bridge short local gaps. Performance is stable across r in [6, 10] and d_max in [15, 25]. Undersized r over-prunes true pixels; oversized r weakens anatomical constraints; undersized d_max fails to bridge gaps; oversized d_max risks hallucinated connections.
R1 Q1: Limited expressiveness of routing Our early exploration showed pixel-level routing yielded marginal Dice/HD95 differences but suffered from overfitting on thin, noisy guidewires. Thus, we adopted image-level gating for stable optimization. Q1: Error propagation from vessel-mask prediction We perturbed the predicted vessel masks with morphological erosion/dilation at radii of 2, 4, 6 px. The refinement remained stable, consistent with our design using a dilated vascular tolerance band rather than a hard vessel-mask constraint. The anatomical prior is supported by 99.7% GT guidewire-vessel overlap reported in our dataset analysis. Q1: Hallucinated connections Each image in our dataset contains exactly a single guidewire, avoiding misconnections. Bridging is also tightly constrained: a candidate pair is bridged only if their distance is below d_max and the path lies within the vascular tolerance band. Q2: Ablation With the refinement included, Dice reaches 84.2% and HD95 drops to 2.36 px. We will add them in the revision. Q3: Data splitting and reproducibility We use a patient-level split to prevent data leakage as stated in Sec. 3.1.We will add the key hyperparameters in the revision (e.g., AdamW optimizer, lr=0.005, LoRA rank=4, r=8, d_max=20). The complete source code, pre-trained weights will be released upon acceptance. Q4: Ethics Check As stated in Sec. 3.1, this study was conducted under the formal approval of the Medical Ethics Committee.
R2 Q1: Limited data source The current dataset is intended as a first validation of the anatomy-constrained formulation rather than a definitive test of external generalization. We will expand the patient cohort and clinical scenarios in future work. Q2: Comparison The cited YOLOP-UNet performs multi-task segmentation but does not enforce spatial constraints between tasks. Moreover, the paper reports that its guidewire predictions are nearly blank, making a direct quantitative comparison unreliable. Our ablation already provides a similar comparison: the “+Experts” setting acts as a standard multi-task baseline, and adding the residual router isolates the value of anatomy-aware modeling, reducing HD95 from 8.32 to 2.98 px. Q4: Real-time deployment Our work targets DSA (1-4 FPS), so 5.7 FPS exceeds the clinical update rate. The remaining bottleneck is the SAM encoder; RDE-SAM can be replaced with a lighter backbone such as RDE-UNet for real-time deployment, as demonstrated in Table 3. R4 Q1: Formatting issues The horizontal lines in Table 1 separate baseline categories. We will clarify it in the caption and reformat Fig. 1 for clarity. Q2: Model details The experts are not independent full decoders. They share the Two-Way Transformer and mask upscaling. The global expert uses the original SAM hypernetwork for L_global, while the guidewire and vessel experts are additional lightweight hypernetwork heads applied to the shared upscaled embedding for L_wire and L_vessel. We will clarify them in Sec. 2.3.Q3: Public datasets To the best of our knowledge, no public DSA datasets provide paired pixel-level guidewire-vessel annotations, motivating our in-house dataset collection. We will expand the dataset scale and clinical scenarios in future work. Q4: Progressive training We compared progressive against joint training. Under joint training, the severe class imbalance induced expert collapse and unstable gate saturation. Thus, we use progressive training for stable optimization.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
This paper tackles a clinically relevant guidewire segmentation problem in DSA and proposes to model it as an anatomically constrained structured prediction task rather than as an isolated thin-structure segmentation problem. Reviewers found this reformulation well motivated and appreciated the proposed role-disentangled expert design with residual routing, as well as the anatomy-constrained refinement module. The method was viewed as practically meaningful, and the reported improvements, particularly in structural metrics such as HD95, are promising. The validation on two different backbones was also seen as a positive indication of generality.
At the same time, the reviewers identified several important issues that should be addressed in the rebuttal. The main concerns relate to experimental scope and methodological completeness. In particular, the study is based on a relatively small, single-dataset cohort, which limits confidence in generalizability across sites, devices, and vascular territories. Reviewers also noted missing comparisons with methods that more explicitly model vessel-instrument relationships, making it harder to isolate the incremental value of the proposed anatomical constraint modeling. In addition, key components such as the anatomy-constrained refinement and the progressive training strategy would benefit from more complete ablation and sensitivity analyses, especially for critical hyperparameters such as the tolerance-band radius and endpoint-bridging threshold.
Further concerns include clarity and reproducibility. Some model details remain under-specified, including the architecture of the experts, the exact data split protocol, and several implementation choices. There were also questions regarding robustness of the routing/refinement design and the practical suitability of the reported inference time for real-time intervention settings.
Overall, the paper presents a promising and clinically meaningful idea, but the current reviews indicate that important questions remain regarding generalization, experimental support, and reproducibility. I encourage the authors to use the rebuttal to address these points in detail.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
This paper reformulates guidewire segmentation in DSA as an anatomically constrained structured prediction task and proposes a role-disentangled expert design with residual routing and anatomy-aware refinement. The clinical motivation is strong, the problem reformulation is well justified by the 99.7% guidewire-vessel overlap observed in the data, and the reported structural improvements (HD95 reduction exceeding 56% over a strong baseline) are compelling.
After rebuttal, the authors have substantially addressed the main concerns. Sensitivity analyses for the critical hyperparameters r and d_max were provided, showing stable performance across reasonable ranges. The ablation study was completed to quantify the contribution of the refinement module. Implementation details, including the expert architecture and progressive training rationale, were clarified. The concerns regarding the single-dataset scope and lack of public benchmarks are acknowledged as limitations, which is acceptable for a first validation of a novel formulation. The consistent gains across UNet and SAM backbones provide meaningful evidence of generality.
Reviewer #2 explicitly endorses the contribution. Reviewer #1 confirms technical soundness and clinical relevance, and Reviewer #4’s remaining formatting and detail requests are minor. I recommend accept. For the camera-ready version, the authors must: (1) include the hyperparameter sensitivity analyses and the completed ablation with refinement in the main text; (2) clearly describe the shared expert architecture as clarified in the rebuttal; (3) correct the formatting issues in Table 1 and improve the readability of Fig. 1; and (4) explicitly discuss the single-dataset limitation and the plan for future multi-center validation.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers generally found the paper clinically relevant and technically sound, with a well-motivated formulation of guidewire segmentation as an anatomy-constrained structured prediction problem. The rebuttal clarified several important points, including the expert architecture, patient-level split, progressive training strategy, refinement hyperparameters, and robustness of the anatomy-constrained refinement. I therefore recommend acceptance.
Some limitations remain, especially the relatively small single-cohort dataset, the lack of external validation, and limited comparison with other vessel-instrument relationship modeling methods. I encourage the authors to discuss these limitations clearly and strengthen the ablation/failure-case analysis in the camera-ready version. Great that the authors promise to release the code and pretrained weights for the benefit of the research community.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Based on the outcome of the rebuttal, one reviewer assigned an “Accept” rating, while the other two maintained their ratings of “Accept” and “Weak Reject.” Taking into account the collective reviewer feedback as well as the technical review of the paper, this work demonstrates significant practical relevance, reasonable technical design, and overall empirical strength. Consequently, the overall assessment is positive, and the final decision is “Accept.”