Abstract

Digital Subtraction Angiography (DSA) is the clinical gold standard for vascular imaging but is frequently degraded by motion artifacts due to patient physiological movement. While deep learning offers a potential alternative to traditional subtraction, existing models often fail to handle cases with severe motion artifacts and frequently yield incomplete vessel reconstructions. To address these challenges, we propose a Prior-guided Hierarchical Vector Quantized Variational Autoencoder (PH-VQVAE). First, our model features a dual-branch encoder that extracts features from both the pre-contrast mask and contrast images at multiple scales. We introduce a window-based cross-attention mechanism at each level to efficiently capture local contrast differences and fuse anatomical priors, enabling precise “soft subtraction” robust to local misalignments. Second, we abandon the continuous latent representations used in conventional generative models in favor of a hierarchical discrete modeling approach. By mapping multi-scale features to learned discrete codebooks, our method effectively filters out background noise and motion inconsistencies, which lack compact representations in the discrete space, while preserving the topological integrity of vessels. Experimental results demonstrate that our approach significantly reduces motion artifacts and achieves superior reconstruction accuracy compared to the state-of-the-art generative models, offering a promising tool for safer and more efficient interventional imaging.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1534_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiuYun_PriorGuided_MICCAI2026,
        author = { Liu, Yunbi AND Yang, Zheyun AND Tang, Enqi AND Li, Shiyu AND Tian, Yifei AND Suo, Shiteng AND Liu, Qingshan},
        title = { { Prior-Guided Hierarchical Vector Quantized-VAE for Digital Subtraction Angiography Generation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16888},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors present a method to remove the pre-bolus background signals from raw DSA images with robustness against motion artifact. The method is a modified VAE which integrates with hierarchical, limited-span attention and vector quantisation to suppress background noise while preserving signals from fine vessels.

  • 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 is well written, with a clear structure and a smooth, logical flow. The motivation is effectively presented, and the design is described and explained in a clear and coherent manner. The results provide convincing evidence that the proposed method achieves higher reconstruction accuracy compared to the referenced approaches.

  • 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.

    While the paper is generally well articulated, the clarity of several arguments could be improved. For example, (1) in the Introduction, the limitations of existing methods are presented as factual statements but are not supported by references. These sentences should be revised to clarify whether they reflect established findings or the authors’ own interpretations, and appropriate citations should be added where necessary. (2) Similarly, in the Methods section, the statement “Continuous latent spaces often encode stochastic motion artifacts and high-frequency noise indistinguishably from the vascular signal” should be supported with relevant references.

    How to justify the use of SSIM and RMAE as quantitative metrics, when the referenced “ground truth” image contain motion artifacts? The authors should perhaps dicuss.

  • 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.

    (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?

    Overall, this work is well presented and the performance of the proposed method surpasses existing methods according to the evidence.

  • 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 addressed the key concerns. Therefore, I recommend acceptance.



Review #2

  • Please describe the contribution of the paper

    1) Hierarchical Discrete Modeling (VQ-VAE) The authors pivot away from standard continuous latent representations toward a Hierarchical Vector Quantized-Variational Autoencoder (PH-VQVAE) 2) Prior-Guided Dual-Branch Architecture Instead of treating the DSA task as a single-image translation, the authors utilize a dual-branch encoder to process the pre-contrast mask and the contrast image simultaneously. 3) Multi-Scale Window-based Cross-Attention (MWCA) The core intelligence of the model lies in its fusion mechanism. Using window-based cross-attention at different resolutions allows the model to perform soft subtraction.

  • 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 work makes a significant departure from standard single-scale continuous latent representations. By introducing the first hierarchical VQ-VAE specifically for the DSA task, the authors effectively resolve the feature averaging problem common in global compression. This multi-scale approach ensures that high-fidelity details of tiny collateral vessels are preserved, even at high compression ratios that typically cause loss of signal in conventional models. 2) The integration of a multi-scale window-based cross-attention mechanism provides a sophisticated framework for soft subtraction. Unlike traditional rigid subtraction, this localized attention allows the model to remain highly robust to the non-rigid local misalignments and physiological motion that typically degrade the clinical gold standard. This ensures a cleaner vascular reconstruction without the ghosting artifacts common in manual or simple automated methods. 3) An extensive comparative analysis against a wide array of state-of-the-art baselines supports the paper. By evaluating the PH-VQVAE using a diverse set of metrics, including PSNR, SSIM, RMAE, and LPIPS, the authors provide comprehensive quantitative proof that their hierarchical discrete modeling outperforms existing continuous generative approaches in both mathematical accuracy and perceptual realism.

  • 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 Generalizability and Domain Robustness: The study relies exclusively on a single dataset for both training and testing, which fails to account for domain shift. In clinical practice, performance can vary significantly across different X-ray hardware, imaging protocols (e.g., varying kVp or dose levels), and patient populations. Without cross-dataset validation, it remains unclear whether the PH-VQVAE can maintain its accuracy when deployed in diverse clinical environments. 2) Constraints of Discrete Codebook Representation: While the hierarchical discrete codebook effectively filters noise, it introduces a risk of quantization error. Because the model forces the reconstruction to match a finite set of learned patterns, there is a potential clinical risk of pathology smoothing. Rare vascular structures might be incorrectly mapped to “normal” vessel prototypes in the codebook, leading to the loss of critical diagnostic details.

  • 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) The authors provide what is the first application of a Hierarchical VQ-VAE to the Digital Subtraction Angiography (DSA) task. 2) The overall score was heavily influenced by the extensive experimental validation. The authors did not rely on a narrow set of metrics; instead, they conducted a comprehensive benchmarking exercise against a wide array of SOTA baselines.

  • 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.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper proposes an improved version of VQVAE for digital subtraction angiography generation. The major contributions are a dual-branch encoder with cross attention for prior incorporation and hierarchical vector quantization for structural filtering. The experiments demonstrate the superiority of the proposed model in terms of reconstruction accuracy.

  • 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 proposed model is overall a sound design, as the employed strategies, such as cross-attention and vector quantization, have been successfully employed for many medical image analysis areas. The overall methodology and illustration are straightforward to understand. 2) The experiments include an ablation study and a comparative study and are thus relatively comprehensive. The ablation study is detailed and clear, successfully validating the components of the proposed model. The qualitative results are nicely visualized and presented.

  • 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 positioning of the proposed method is a bit unclear. Although VQVAE can be regarded as a generative model, its usage in this paper is more similar to a discriminative model. In fact, the proposed model is similar to a U-Net with a modified encoder and regularized latents. From this perspective, some of the compared approaches, such as the diffusion model, might not be appropriate for comparison. Besides, AaDSA, one of the state-of-the-art methods, does not use the mask image, not because the authors did not realize its importance, but because the research objective of AaDSA was to achieve subtraction-free DSA generation. However, the proposed model in this paper employs the mask image, leading to a potentially unfair experimental setup and comparison, and weakening the merit of the proposed model. 2) The motivation for the conducted improvements is insufficiently illustrated. For instance, the paper claims that the employed window-based cross-attention can perform a soft subtraction in the feature space. However, cross-attention is a general computational technique for representation learning that is widely used in deep learning. The connection between it and the soft subtraction in the feature space is unclear, and the paper shows no evidence to validate this argument. Besides, the cross-attention involves nothing like subtraction within it. To sum up, the utilization of techniques such as cross-attention and vector quantization is too general and optimized little with respect to the properties of the DSA generation task. 3) Some arguments are vague and a bit exaggerated. For instance, the paper claims, “Continuous latent spaces often encode stochastic motion artifacts and high-frequency noise indistinguishably from the vascular signal” on page 4.This is a very strong claim, and the paper does not give any theoretical or intuitive explanations to support it. In fact, most state-of-the-art research works related to vascular image analysis employ regular continuous latent representations instead of discrete ones. The paper seems to directly associate the performance improvements with some assumed functionalities of the employed techniques, which can hardly be convincing.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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 proposed model is basically sound and achieves a good performance. However, the motivation of the employed techniques is not sufficiently aligned with the characteristics of DSA generation. The motivation contains unclear and exaggerated arguments, making the method less convincing. Besides, the potentially problematic comparative experiment setup weakens the quality of the proposed method.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    The authors have sought to address some of the concerns through new experimental results. From an effectiveness perspective, the new experimental results and accompanying explanations support the arguments of the authors to some extent. However, adding new experimental results to support arguments does not comply with the rebuttal guidelines. It is unfair to other authors who comply with the guidelines and thus do not add new results. Based on the original manuscript, the decision remains a rejection.



Author Feedback

We thank the reviewers for their constructive feedback (R1/R2: 5, R3: 3). We address the major concerns below. 1.More Fair Comparison (R3) We added a comparison with the latest DSA imaging method, VCC-DSA, which was published online in MIA in April after our submission [Ge et al. MIA2026]. VCC-DSA also uses the mask image as an auxiliary input. After careful reproduction, VCC-DSA achieved PSNR 30.26, SSIM 0.899, and LPIPS 0.269, while our method performs better, especially in perceptual quality (LPIPS 0.085 vs. 0.269). For AaDSA, we agree that the input discrepancy should be addressed. In the original manuscript, we had already compared AaDSA with our contrast-only variant, where both methods use only the contrast image. Under this fair setting, our method improves PSNR 27.35 → 28.11, RMAE 2.26 → 2.00, and LPIPS 0.231 → 0.181, with comparable SSIM. We further added a dual-input comparison by upgrading AaDSA with a mask+contrast dual-branch encoder, termed AaDSA-Mask, and comparing it with our dual-branch concatenation baseline without cross-attention. Our model still performs better: SSIM 0.936 vs. 0.926, PSNR 32.33 vs. 31.90, and LPIPS 0.102 vs. 0.115.These results show that the improvement is not simply due to using the mask image, but comes from the proposed task-specific design.

2.Unclear Claims and Motivations (R1&R2&R3) a) Claims on continuous spaces. Our statement was grounded in prior work. Specifically, prior studies have reported that continuous representations can be vulnerable to error accumulation [CVPR 2022], while VQ latent space can prevent input interference, such as noise and artifacts[MICCAI 2024]. We will soften the claims in the revision and add missing citations. b) Task-specific design and quantization risks. Our cross-attention and VQ are not generic plug-ins but are specifically designed for DSA: i. The key challenge of DSA generation is to remove background structures shared by the mask image and contrast image while preserving contrast-enhanced vessels that only appear in the contrast image. Therefore, we separately encode the mask and contrast images, then perform cross-attention between their latent features. This allows the model to suppress shared background structures, e.g., bone and soft tissue, while enhancing contrast-filled vessels. In this sense, the module performs soft subtraction in latent space, avoiding artifacts caused by pixel-wise hard subtraction under mask–contrast misregistration. Thus, the dual-branch encoding with cross-attention is a task-specific design for DSA rather than a generic attention module. The ablation also supports its effectiveness: Concatenation / Ours — SSIM: 0.936 / 0.940, PSNR: 32.33 / 33.07, LPIPS: 0.102 / 0.095. ii. To avoid the fine-detail loss of standard single-scale VQ, we adopt a hierarchical VQ design to mitigate the quantization risks. With tri-scale VQ, coarse levels preserve global topology while finer levels retain vessel details. Thus, VQ is also tailored to DSA’s need for both artifact suppression and fine-vessel fidelity.

3.Generalizability and Dataset Diversity (R1 & R2) Although our data were collected from a single institution, the dataset is clinically diverse. It spans 6 years of acquisition, includes images from GE and Siemens scanners, and covers different artifact severities as well as both Anteroposterior (AP) and Lateral (LAT) projection views. To further assess generalizability, we evaluated the model separately on AP and LAT views. The results were consistent across views: AP / LAT — SSIM: 0.939 / 0.941, PSNR: 33.01 / 33.11, RMAE: 1.18 / 1.17, LPIPS: 0.100 / 0.089.These results indicate that the model maintains robust performance across different projection views.

Reference: 1.Gu et al., CVPR 2022, Vector Quantized Diffusion Model for Text-to-Image Synthesis 2.Han et al., MICCAI 2024, Non-Adversarial Learning: Vector-Quantized Common Latent Space for Multi-Sequence MRI 3.Ge et al. MIA 2026, VCC-DSA




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 proposes a prior-guided hierarchical VQ-VAE framework for DSA generation that aims to improve background suppression and vessel preservation.

    Reviewers agree the method is well designed with strong empirical performance and clear presentation. I suggest that the authors clarify the raised concerns about 1) unclear motivation and positioning, 2) potentially unfair comparisons, and 3) limited validation of key claims.

  • 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 agree that the paper is well presented and proposes a sound prior-guided hierarchical VQ-VAE framework for DSA generation. The rebuttal clarifies several concerns regarding the motivation, positioning, and role of the proposed components. Overall, I would recommend acceptance.



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.

    This paper proposes a dual-branch encoder with cross-attention between reference and contrast images, coupled with a hierarchical VQ-VAE, for DSA reconstruction. The reviewers’ majority opinion is positive. R2 is strongly positive, viewing the hierarchical discrete representation as a meaningful departure from single-scale continuous latents and crediting the extensive benchmarking against SOTA baselines. R3 is negative, judging that the method is positioned as generative but functions more like a discriminative U-Net, and that several claims are made without evidence or are overconfident. My view is that the design is well suited to the DSA task and the empirical results are convincing enough to merit acceptance. That said, R3’s concerns are valid and partly echoed by R1, and the authors are strongly encouraged to add the missing citations and to temper or substantiate the soft-subtraction and continuous-latent claims in the final version. Finally, regarding the additional results in the rebuttal, which are in violation of the ‘no new results’ policy, R3 rightfully does not sway their opinion, and the other reviewers are already positive, and thus the rebuttal did not influence my final recommendation.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    There are concerns of adding new experimental results to support arguments which does not comply with the rebuttal guidelines.



back to top