List of Papers Browse by Subject Areas Author List
Abstract
Vessel segmentation in digital subtraction angiography (DSA) is crucial for intraoperative visualization and quantitative assessment during endovascular interventions. However, accurate segmentation remains challenging due to complex thin vascular branches and strong background interference from residual anatomical structures after subtraction. Although recent methods incorporate temporal modeling using CNNs, recurrent networks, or transformers, they often treat temporal information implicitly and fail to explicitly exploit the contrast flow dynamics in DSA sequences. In this work, we present a spatiotemporal DSA vessel segmentation framework that combines a forward phase-difference prior with Mamba-based temporal sequence modeling. The Mamba-based temporal encoder models long-range temporal dependencies and contrast flow dynamics with linear computational complexity. To capture contrast-driven intensity changes, we introduce a forward phase-difference projection that enhances contrast-responsive vascular structures while suppressing temporally stable background tissues. A minimum intensity projection (MinIP) is further used as a complementary static spatial prior to provide coarse vessel structural cues. These priors are integrated with temporal features through a lightweight domain-aware fusion module before decoding. Extensive experiments on the DIAS and DSCA datasets demonstrate that the proposed method consistently outperforms state-of-the-art approaches, particularly in preserving vascular connectivity and segmenting thin distal vessels. Code is available at https://github.com/kkk123lll/FPD-SegDSA.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1936_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/kkk123lll/FPD-SegDSA
Link to the Dataset(s)
DIAS: https://zenodo.org/records/11637181
DSCA: https://zenodo.org/records/11255024
BibTex
@InProceedings{LiuKun_Temporal_MICCAI2026,
author = { Liu, Kun AND He, Ziyang AND Zheng, Bin AND Zhao, Wenyi AND Zhu, Mengke AND Xu, Weijin AND Liu, Wentao AND He, Zijun AND Chen, Hanlin AND Lu, Bofeng AND Liang, Zhiyuan AND Yang, Huihua},
title = { { Temporal Phase-Difference Guided Spatiotemporal Learning for DSA Vessel 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
This paper studies vessel segmentation in DSA sequences and proposes a spatiotemporal framework that combines: (1) a Mamba-based temporal encoder for modeling contrast-agent flow dynamics; (2) a forward phase-difference MaxIP prior (FPD-MaxIP) that explicitly encodes intensity changes between early and late phases; (3) a MinIP-based static prior, which is fused with temporal features through a lightweight domain-aware fusion module. The method is evaluated on the DIAS and DSCA datasets, where improved performance is reported over existing methods, particularly in preserving thin distal vessels and vascular connectivity. A key strength of the paper is that it attempts to explicitly encode a clinically intuitive prior rather than relying solely on implicit temporal modeling. This idea is well motivated for the DSA setting, and the comparison between FPD-MaxIP and MinIP in Fig. 1 provides intuitive visual support for this design. However, the experimental validation is still not sufficient to fully support several of the paper’s stronger claims. In particular, the claimed “cross-dataset generalization” is not rigorously demonstrated. In addition, the manuscript suffers from noticeable issues in writing and presentation, including multiple formatting and editing errors.
- 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 core motivation is reasonable from both clinical and technical perspectives. The paper correctly points out that, in DSA, vessel visibility is largely driven by the temporal propagation of the contrast agent, while background tissues and bony structures usually remain relatively stable after subtraction. Based on this observation, introducing an explicit phase-difference prior is a domain-motivated design choice rather than a simple generic architectural modification.
2.The overall architecture is clearly organized. The authors decouple temporal dynamics modeling from spatial prior guidance: temporal features are extracted by the TFE blocks and pooled into 2D representations, while MinIP and FPD-MaxIP are introduced as static and dynamic priors, respectively, and fused at later stages. This decoupled design makes the framework more modular and also facilitates analysis of the contribution of each component.
3.The quantitative results are competitive. According to the reported results, the proposed method outperforms the listed baselines on both DIAS and DSCA, achieving the best performance across metrics such as DSC, IoU, AUC, and clDice, while maintaining a moderate model size and computational cost in terms of parameters and FLOPs.
- 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 experimental results are not sufficient to support the claim of “cross-dataset generalization. ” The introduction and contribution statements explicitly mention improved “cross-dataset generalization. ” However, from the experimental section, the authors only conduct standard in-domain training and testing on DIAS and DSCA separately. A truly cross-dataset setup is missing. For example, the authors should train on DIAS and test on DSCA (and ideally vice versa) to properly support this claim.
2.The fairness of the baseline comparison remains insufficiently clarified. The paper does not clearly state whether all baseline methods were retrained under the same preprocessing pipeline, patch size, and data split settings.
3.The discussion of failure cases and clinical applicability is limited. The paper emphasizes improved preservation of thin distal vessels and better background suppression, but it does not include a dedicated analysis of failure cases. For example, under what conditions does FPD-MaxIP fail? Could it amplify noise, motion artifacts, venous contamination, or abnormal contrast propagation patterns? Since this prior relies on a fixed early-to-late phase difference, its robustness to irregular temporal dynamics and variation in acquisition protocols should be discussed more explicitly. For instance, in clinical practice, patients may differ substantially in blood flow velocity, heart rate, and contrast injection rate, meaning that the arrival and washout phases of the contrast agent are not fixed.
4.Typos and formatting issues. a) In the last paragraph of the Introduction, there appears to be hidden Chinese text in the sentence around “We develop … sting methods. ” b) The caption of Fig. 2 contains two items labeled (d).
- 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?
Please refer to the summary and the weakness sections.
- 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.
I thank the authors for preparing the rebuttal. Since my concerns have been addressed well, I am happy to recommend an acceptance. The authors are encouraged to incorporate the suggest revisions into the camera-ready paper.
Review #2
- Please describe the contribution of the paper
This paper introduces an advanced framework for vessel segmentation using DSA sequences. The proposed method integrates a SSM for effective temporal modeling and computes both the phase-difference map and the MinIP map as spatial priors to enhance contrast-induced features. Comprehensive experiments on the DIAS and DSCA datasets demonstrate the effectiveness and potential of the proposed segmentation framework.
- 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.Temporal modeling is crucial for analyzing sequential medical images. This paper introduces the SSM block to capture temporal variations across consecutive frames, tailored specifically for vessel segmentation. And the corresponding experimental results effectively demonstrate its utility.
2.The proposed forward phase-difference prior looks novel. Visual comparisons in Fig. 1 highlight that the FPD-MaxIP provides distinct and complementary information compared to the traditional MinIP prior, reinforcing the segmentation process through improved contrast-induced feature representation.
3.The authors perform extensive experiments, including comparisons with multiple approaches and ablation studies, validating the proposed model design and highlight its advantages.
- 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.
Method: 1.TFE block: Why is there a reshape operation in every layer of the TFE block (as shown in Fig. 2(a))? The TFE module integrates both SSM-based temporal modeling and temporal attention mechanisms. What are the distinct purposes of these two steps, and why are both necessary? 2.FPD-MaxIP: Does the simultaneous use of MinIP and MaxIP-based difference priors risk introducing conflicting features for the model to learn? When computing the forward phase-difference prior, why is T/2 chosen to distinguish between the early stage and the late stage? Would it be possible to adopt a more adaptive approach to select this parameter? Will using a difference-based method to compute features potentially make the features more sensitive to respiratory motion?
Experiments: 1.Considering temporal attention is integrated in the TFE module, is it possible to visualize the attention maps to provide interpretability for the temporal modeling process? 2.Based on table. 1, TemSAM is trained on a subset of the DIAS dataset and achieves strong generalization ability on the DSCA dataset, while the author’s method requires training and testing on the same data distribution within the DSCA dataset, the experimental results show that TemSAM and the author’s method has different advantages across various metrics on both datasets. Can the authors provide a more detailed discussion on the key advantages of the proposed method compared to SAM-based methods. Does the proposed method also exhibit generalization ability?
- 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.
(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 authors present a detailed modeling approach for vessel segmentation, incorporating SSM-based temporal modeling and the FPD-MaxIP prior to enhance contrast-induced features. However, some design details in the methodology require to be clarified, and the advantages of the proposed method over SAM-based approaches could be more emphasized.
- 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 authors have responded to my comments and committed to clarifying several points in the method and experiment sections in the final revision.
Review #3
- Please describe the contribution of the paper
This paper introduces FPD-MaxIP as spatial prior for DSA vessel segmentation that enhances contrast-reponsive vascular structures and suppresses temporally stable background tissues. Meanwhile, a Mamba-based framework is proposed to efficiently exploit the contrast flow dynamics with the guidance of FPD-MaxIP.
- 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 conceptual design of FPD-MaxIP is highly novel. The experimental evaluation is comprehensive and thoroughly validates the proposed method.
- 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.Since FPD-MaxIP is explicitly designed to suppress persistent tissue and bony artifacts, the evaluation metric “Precision” should be introduced in the experiments to better reflect the model’s ability to reduce false positives. 2.The calculation of the AUC metric needs to be double-checked for potential errors. In the original papers of VSS-Net and TemSAM, the reported AUC scores for almost all compared methods are > 0.97.However, in Table 1 of this manuscript, the AUC values for the compared methods are surprisingly < 0.9.Please clarify this discrepancy. 3.Although the ablation study in Table 2 demonstrates that jointly introducing MinIP and FPD-MaxIP yields the best performance, the specific synergistic mechanism between these two priors in the DSA vessel segmentation task requires further explanation. As shown in Fig. 1, FPD-MaxIP effectively suppresses background structures and enhances vessel details; what, then, are the specific limitations or shortcomings of FPD-MaxIP compared to MinIP that necessitate their combination?
- 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 introduces a spatiotemporal DSA vessel segmentation framework that combines a spatial prior (FPD-MaxIP) with Mamba-based temporal sequence modeling. The methodological design is novel, and the experimental evaluation is comprehensive. However, some specific details require further clarification and verification.
- Reviewer confidence
Very confident (4)
- [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 proposed methodology is novel and interesting, and the baseline experiments are comprehensive. However, the authors’ rebuttal failed to provide the requested AUC and precision metrics, offering descriptions that are too speculative rather than concrete data. To ensure the soundness of the evaluation, the authors must include the correct and complete quantitative results for AUC and precision in the final manuscript. While I recommend acceptance based on the technical novelty and existing experiments, this approval is strictly conditional upon the inclusion of these missing metrics in the camera-ready version.
Author Feedback
We thank the reviewers for their constructive comments. We address the major concerns by grouping related comments. To support reproducibility, we will release the code. a. Generalization claim and typos issues (R1). We appreciate the reviewer’s comment regarding the scope of our claim. To avoid overstatement, we clarify that the submitted experiments demonstrate consistent performance on DIAS and DSCA under dataset-specific training/testing, not strict cross-dataset transfer. This does not affect the core contribution, namely FPD-MaxIP and its integration with temporal modeling and projection priors. We will revise “cross-dataset generalization” to “consistent in-domain performance across two DSA benchmarks”. Besides, we will also correct the reported typos and formatting issues. b. AUC interpretation (R4). After checking the submitted evaluation pipeline, we found that AUC was computed from thresholded prediction maps. Thus, it should not be interpreted as standard AUC computed from continuous probability scores, explaining the difference from prior VSS-Net and TemSAM reports. Since our main conclusions rely on Dice, IoU, clDice, and qualitative results, we agree that the current AUC presentation may be misleading. We will recompute AUC from probability scores or remove this column. c1.TFE design and T/2 split (R2). SSM captures long-range contrast propagation over [B,THW,C], while temporal attention predicts weights [B,T,1,1,1] and reweights [B,T,H,W,C] by broadcasting, emphasizing informative phases. Without SSM, temporal dependencies may be weak; without attention, uniform pooling may dilute transient vessel responses. The T/2 split is an empirical setting that provides broad early-late phase contrast in our experiments. Compared with raw MaxIP or local frame differences, FPD-MaxIP better suppresses stable bony or soft-tissue residuals, although severe motion may still introduce residual edges. We will clarify this limitation and explore adaptive phase selection in the future. c2.Complementarity and failure discussion of MinIP and FPD-MaxIP (R1,2,4). MinIP preserves global vessel topology, while FPD-MaxIP suppresses temporally persistent background responses through phase-difference computation. Their combination couples structural support with dynamic contrast enhancement, explaining the gains in Table 2.However, FPD-MaxIP may become less reliable when the fixed early-late split mismatches patient-specific contrast dynamics or acquisition conditions, such as delayed/uneven filling or severe motion. MinIP can partially mitigate this issue by providing stable vessel support and topology cues, reducing local discontinuity. Future work will explore adaptive phase selection to improve robustness. d. False-positive suppression (R4). We agree that Precision is important for false-positive suppression. Although not listed, IoU penalizes false positives in its denominator and shows a consistent trend. Precision can also be derived from DSC and Sen as (DSC×Sen)/(2Sen−DSC) under the same threshold and aggregation setting. These results support FPD-MaxIP’s false-positive suppression ability. We will add Precision if revision is permitted. e. Baseline fairness and TemSAM (R1,2,4). Except for TemSAM, all compared baselines were retrained using the same official splits and protocol, including preprocessing and 64×64 patch training. TemSAM is reported separately as an external SAM-based reference, using results from the original paper due to reproducibility constraints under our matched protocol. Our method leverages DSA-specific phase-difference and projection priors to suppress stable background responses and preserve vessel connectivity, without large foundation models. The results support consistent performance on DIAS and DSCA under dataset-specific training/testing, while strict cross-dataset generalization requires further validation. We will clarify these settings in the revision.
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.
Scores are 3/4/4, with the most confident reviewer (R4, confidence 4) recommending weak accept on the strength of the FPD-MaxIP design and comprehensive evaluation. The weak reject (R1) is driven primarily by the unsupported cross-dataset claim and missing failure-case analysis rather than by disagreement with the core idea. The paper is borderline but the concerns are largely addressable, so the paper is invited for rebuttal. Authors should prioritize (a) a true cross-dataset transfer experiment (e.g., DIAS→DSCA) or a retraction/softening of the generalization claim, (b) reconciling the AUC discrepancy with the original VSS-Net and TemSAM papers, (c) clarifying the SSM-vs-temporal-attention rationale, the T/2 split, and the MinIP/FPD-MaxIP synergy mechanism, (d) reporting Precision and a brief failure-case discussion, and (e) confirming that all baselines were retrained under identical preprocessing and splits.
- 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.
Reviewers agreed that the FPD-MaxIP prior is conceptually novel and clinically well-motivated, the decoupling of temporal dynamics modeling from spatial prior guidance is a clean architectural choice, and the quantitative results are competitive across two benchmarks. They raised several concerns: (1) the claimed cross-dataset generalization is not supported by actual cross-dataset training/testing experiments; (2) reported AUC values (<0.9) are inconsistent with prior VSS-Net and TemSAM reports (>0.97), suggesting a metric-computation discrepancy; (3) Precision is missing, despite the method’s focus on false-positive suppression; (4) the TFE block design (SSM + temporal attention + reshape operations) is under-justified; (5) the T/2 phase split is empirical, with no adaptive alternative; (6) the synergy between MinIP and FPD-MaxIP needs clearer explanation; (7) failure cases for FPD-MaxIP (motion artifacts, irregular contrast dynamics, patient-specific variation) are not discussed; (8) baseline fairness (matched preprocessing, splits) and TemSAM’s reuse from original paper need clarification; and (9) typos and formatting issues are present.After rebuttal, the concerns were largely resolved or addressed — the authors agreed to soften the cross-dataset claim, acknowledged the AUC issue and committed to recomputing or removing it, clarified TFE design, explained MinIP/FPD-MaxIP complementarity and limitations, and confirmed baseline retraining under matched protocols. All three reviewers raise their recommendations to Accept (one explicitly conditional on including the corrected AUC and Precision metrics in the camera-ready).
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 proposed FPD-MaxIP prior is a clinically motivated and technically meaningful design, as it explicitly captures contrast-induced temporal changes rather than relying only on implicit temporal modeling. Its integration with Mamba-based temporal encoding and MinIP-based structural guidance is also well supported by the ablation studies.
The rebuttal satisfactorily addressed the main concerns raised during review, including the scope of the generalization claim, the role of temporal modeling and attention, the complementarity between MinIP and FPD-MaxIP, baseline fairness, and potential failure cases. All post-rebuttal reviewer recommendations are positive, and I agree that the paper is above the acceptance threshold.
For the camera-ready version, the authors should carefully revise the manuscript to avoid overstating cross-dataset generalization, correct the AUC computation or remove the misleading AUC column, clarify the baseline training settings, and fix the remaining formatting/typo issues. These revisions will make the contribution and evaluation more precise and convincing.
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.
The author’s response addresses the reviewers’ concerns.