List of Papers Browse by Subject Areas Author List
Abstract
Cerebral digital subtraction angiography (DSA) is an Xray-based imaging modality that provides high-resolution, real-time visualisation of cerebral vasculature, and is an established part of the standard treatment of stroke patients. Conventionally, DSA data are acquired as 2D images where vessel structures overlap with one another due to the penetrating nature of X-ray. Given the increasing recognition of the importance of microvasculatures in stroke, there is an unmet need to utilise DSA to accurately assess microvessels, unobstructed from overlapping macrovessels. This work proposes a novel Expectation-Maximisation algorithm integrated with anatomy-informed regularisation to disentangle macrovascular and microvascular flow component overlaps in a spatiotemporal Gamma mixture model for DSA. In-vivo experiments across 108 stroke patients demonstrate that the proposed method achieves robust estimation and provides clear separation of the macrovascular and microvascular flow components. Based on the proposed method, quantitative microvascular cerebral blood volume was derived from DSA images and shown to be significantly associated with the current gold-standard reperfusion metric.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/5026_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{WuChe_Resolving_MICCAI2025,
author = { Wu, Chengchuan and Davey, Catherine and Sharma, Gagan and Ng, Felix},
title = { { Resolving the Overlap of Macrovascular and Microvascular Flow Components in Digital Subtraction Angiography for Cerebral Reperfusion Assessment } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15966},
month = {September},
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper addresses the computation of perfusion maps from DSA images. The aim is to isolate the blush due to brain perfusion, called microvascular component, in-between the early arterial phase and the late venous phase which both show more cleanly delineated blood vessels (called macrovascular components). The paper is based on a previously published approach by Scalzo et al which considers each pixel as a mixture of the 3 phases that must be separated. The proposed method extends this work by incorporating a regularizing prior on the smooth, low frequency aspect that the microvascular component should display, and handling singularity in the decomposition for pixels that are not reached by the contrast agent. The method is evaluated on a dataset of 108 patients, by comparison to the method by Scalzo et al. While the reconstruction error is better for the latter, the proposed method is visually assessed as presented with reduced superposition of the macrovasculature onto the microvascular component, and the mean CBV that it can infer is shown to be significantly associated with clinical perfusion score (eTICI).
- 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 clinical application is of the highest interest for interventional neuroradiologists and the authors present a good argument for it. Contrary to many previous works that consider classifying frames in the DSA into one of three phases, the basic approach of considering images as mixtures of components is, in my opinion, the most sound. The original method indeed presents with a regularizing issue that prevents tiny vessels from actually be considered into the macrovascular component and the proposed work is a smart way to incorporate smoothness priors and make the estimation more robust.
- 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.
In my opinion, the proposed regularizing term is an approximation, more precisely a relaxation. The correct way to incorporate a smoothness prior would be to minimize variations of the mixture parameters ($\alpha_k, \beta_k, \tau_k$) over a small neighbourhood. But that would imply a problem that would probably hardly be tractable due to induced couplings between formerly pixelwise losses. The paper lacks insight on this aspect of the algorithm. Also, experiments are missing to understand the impact of several parameters on the results.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission 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
The introduction is very well written and provide both a good overview of previous works and a good argument for the clinical interest of computing perfusion maps from DSA.
In section 3.2, I reckon that mentioning the Gaussian weighting in 3rd position, right before equation 9, would make things clearer.
The paragraph about the constraints to non-vessel pixels lacks clarity. The final mention of $p \in p_t$ is unclear, since $p_t$ was introduced as a single pixel. Same remark for $k_M$ which was introduced as a single component. Also, why use Frangi’s filter, which is known to be tricky to parameterize and was not design for vessel segmentation per se, when there now exist efficient deep learning-based methods to segment blood vessels in DSA?
In section 4.2:
- are all patient data biplane acquisitions?
- which are the frame rates for the sequences in the dataset?
- how is the temporal interpolation done? This is absolutely non trivial. What is the highest frame rate?
- the applied subsampling is quite severe. Why such a reduction in image size? What would be the impact of preserving the original resolution (around 750-1000 I guess)?
The experiments should be expanded (the description of the method in [11] could be reduced to save space):
- using independent $\tau_{pk]$ in a neighbourhood of 81 pixels imply that each pixel can be assigned 81 different value for $\tau_k$. This is a relaxation of the otherwise strong coupling that should be imposed. Images of the extend in which this parameter varies for each pixel would be informative (e.g. the standard deviation of these 81 values per pixel). If the relaxation is well grounded, then this variation should be small.
- the Gaussian weighting appears to be redundant with independent $\tau_{pk}$. Experiments should be provided without Gaussian weighting, as well as with other values for $\sigma$
- experiments should be provided to demonstrate the impact of small or larger neighbourhood sizes (which is one of the main aspect of the paper)
Typos:
- section 1, first paragraph : vascualture -> vasculature
- end of second paragraph, p.2: treamtent -> treatment
- equation 8: there is sum over p within a sum over p. Please use two different notations.
- equation 11: same thing with index i
- section 4.1: intialised -> initialised
- p. 8 and fig. 3: coronal and sagittal are used for 3D imaging. I have seen many occurrences of these terms in recent papers about DSA, but I was very recently confirmed by my clinical partners that the medical community still uses frontal and lateral terminology.
- 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 approach is very interesting and the addressed issue is important. On the other hand, more experiments should be realized to better understand the impact of parameters and design choices.
- Reviewer confidence
Very confident (4)
- [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
The authors present an approach to distinguish between macro (arterial and venous) and microvascular enhancement in 2D DSA sequences using a pixel-wise gamma-variate mixture model. The approach is based on a previously published pixel-wise gamma-variate mixture model approach, but extends this by including the local pixel neighborhood in the EM-estimation to improve the accuracy of microvascular enhancement estimation in locations with overlap.
- 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 presented approach is well founded in statistical theory.
While the inclusion of the local pixel neighborhood is a relatively small addition, it seems to clearly improve (at least visually) the results compared to previously published techniques.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper presents an interesting approach, though its structure could benefit from some refinement. Specifically, the flow between the theoretical background, theory, and methods sections could be improved for clarity. Introducing key terms earlier, when first referenced in the theory, would enhance reader comprehension. Additionally, separating the results and discussion sections would allow for a clearer distinction between observed findings and their interpretation.
While the qualitative results are presented well, the quantitative evaluation could be strengthened. The reported goodness-of-fit, while high, may not fully capture the method’s effectiveness given the model’s complexity. The statistical analysis is lacking crucial details including the type of test used, null hypothesis, and descriptive statistics (means/standard deviations or medians/IQRs), and precise p-values. Including these details would significantly enhance the rigor of the evaluation.
Furthermore, the evaluation strategy warrants reconsideration. The introduction highlights the limitations of TICI scores in predicting procedural outcomes. However, the evaluation relies on comparing the proposed method’s CBV measurements with these same TICI scores. Addressing this apparent discrepancy by providing additional context or alternative evaluation metrics would strengthen the validation of the proposed technique.
Some of the equations in the paper seem to be inconsistent with the descriptions (see detailed comments).
- 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
-
Regarding Equation 1, it appears there might be a slight discrepancy between the textual description and the equation itself. While the text explains it as the probability of a particle appearing at a specific pixel at time T, the equation seems to also incorporate a component-specific element. To ensure clarity, could the authors please elaborate on the definition of a “component” in this context? Specifically, clarifying whether it represents a binary variable, such as micro versus macro vasculature, would be very helpful for reader comprehension.
-
Conditional probabilities are sometimes noted with a semicolon and sometimes with a vertical bar. It would be helpful to stay consistent throughout the paper.
-
In the description of Equation 2, the phrase ‘N denotes the total number of sampling points’ could benefit from further clarification. While ‘sampling points’ might be interpreted as spatial locations, the equation’s context suggests they represent distinct time points. Specifying that ‘N’ refers to the total number of temporal sampling points would enhance the description’s accuracy and prevent potential misinterpretations.
-
Regarding Equation 2, the current description of the left-hand side, stating it represents the likelihood of the parameter set theta given known T and Z, appears to conflict with the right-hand side, which sums over all time points and components. To ensure accuracy, would it be more appropriate to define the left-hand side as the log-likelihood of theta, conditioned on the observed concentration at each time point? This adjustment might better reflect the equation’s structure and intent.
-
In Equation 2, the method of ‘iterating’ or summing over the contrast concentration C(t_i) at each time point t_i requires further clarification. As contrast concentration values are typically not integers, it would be beneficial to explain how this summation is performed. Specifically, detailing whether any form of conversion or estimation, such as converting to an estimated number of particles, is involved would greatly enhance the reader’s understanding of the process.
-
- 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?
While the presented approach is overall very interesting, the mathematical description of the approach and the scientific rigor of the evaluation could be improved.
- Reviewer confidence
Very confident (4)
- [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
Application of a neighbourhood based EM-GMM method to distinguish macro- and micro-vascular signals in 2D DSA.
- 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.
Comparison to CBV and eTICI for post-EVT stroke patients to demonstrate validity of the approach.
- 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.
Success of the approach is presented based on the comparison to CBV and eTICI, however this could be backed up by demonstration that baseline methods (P-EM or other) do not achieve a comparably significant association. It’s not clear from the description how the Frangi vesselness filter is integrated into the algorithm or how sensitive the method is to the short-comings in this component eg. for bifurcations. Also, how the order of Gaussian components is constrained, if at all, and justification of the choice of three components - is this optimum?
- 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
I would anticipate that this method would fail to accurately model the filling of two temporally distinct overlaid vessels. Can you comment? Also are three the optimum number of Gaussians? Of course more would improve the goodness of fit. Equivalent of Fig. 3 for P-EM or other baseline would help demonstrate the superiority of your approach.
- 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?
I don’t consider this method particularly novel but the results look visually convincing. A brief search could not throw up any similar approaches but I’d be surprised if there aren’t papers out there applying similar, if not identical approaches, for this clinical problem.
- Reviewer confidence
Very confident (4)
- [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 thank the reviewers and meta-reviewer for their constructive feedback, and we are encouraged by the overall positive assessment of the novelty [R1, R2], clinical relevance [R1, R3] and theoretical grounding [R1, R2] of our method. Below, we summarise our response to key comments:
Major comments:
- Clarity and flow [R1, R2]: We appreciate the reviewers’ feedback on improving presentation. We will make necessary revisions to clarify equations and variables, use consistent terminology, and correct for typos. We will separate the results and discussion to improve the structure.
- Algorithm insight [R1, R3]: While we focused on clinical demonstration in this manuscript, we agree that understanding the algorithm is important, particularly the parameters (e.g., neighbourhood size, Gaussian weighting, and number of Gamma components). The impact of parameters will be demonstrated in future work.
- Evaluation strategy [R2, R3]: We thank the reviewers for raising concerns regarding the evaluation strategy and suggesting comparative experimental design. While TICI scores have known limitations, their correlation with mCBV proves that our proposed approach is clinically relevant. However, we agree that additional experiments are required to assess the prognostic ability of mCBV, such as correlation of mCBV with long-term patient outcomes. They will be the focus of future work. We understand the reviewer’s concern that the current test lacks a baseline. However, correlating baseline mCBV (P-EM, K=3) and eTICI introduces interpretability issues owing to the prevalence of artifacts. The benefits of our approach is demonstrated by comparison with the literature. In Scalzo et al., weak association between CBV (P-EM, K=2) and TICI (p<0.17) was reported. In comparison, mCBV derived by our method (N-EM, K=3) has shown significant association with stronger evidence (p<0.001).
- Vessel segmentation [R1, R3]: Frangi filter was used for its interpretable and easy application. We appreciate reviewers’ concerns and will provide a more detailed description. Inaccurate vessel segmentation errors can propagate to N-EM. We agree that modern AI models (e.g. Su et al. 2024) have demonstrated high segmentation accuracies. To understand how vessel segmentation impacts N-EM, an ablation study will be presented in future work.
Minor comments:
- Biplane DSA were obtained for all patients in this dataset, with max rate 4 fps. Linear temporal interpolation was used. While advanced interpolation (e.g. Haouchine et al. 2021) can be more accurate, linear interpolation seems adequate for our purpose.
- The original image matrix sizes, typically 1024*1024, were down-sampled for computational efficiency. Conceptually, higher resolution leads to lower microvascular fluctuation in a fixed-size neighbourhood. Thus, algorithm parameters, particularly neighbourhood size, need adjustment for matrix size.
- Reviewer 1 suggested validating the smoothness regularisation based on intermediate tau values. We are thankful for this suggestion and will consider its inclusion in future work.
- We thank Reviewer 2 for raising concerns over the statistical testing. More details are provided in the revision.
- We thank Reviewer 3 for raising concerns over the choice of K (number of GMM components) and the temporal order of the components. We assumed K=3 based on the 3 phases of cerebral blood flow. Conceptually, for data depicting 2 overlaid vessels, 2 of the 3 components will be fitted to the vessel data, with the extra one fitted to noise, typically negligible. The effects can be investigated by simulations, the focus of future work. For N-EM, the order of components is not a major concern, as the microvascular component can be identified by the smoothness constraint.
We thank the reviewers again for their valuable insights. We believe the suggested revisions will significantly improve the clarity, rigor, and impact of our work.
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 presents a pixel-wise gamma-variate mixture model with a neighborhood-based EM-GMM refinement to separate macro- and microvascular flow in DSA for improved cerebral reperfusion assessment.
The paper has received three weak accepts, so there is consensus. Reviewers found the method’s clinical motivation and statistical grounding compelling. They highlighted the sound extension of Scalzo’s mixture model via smoothness priors and singularity handling, and praised the clear introduction and clinical correlation with CBV and eTICI scores.
Reviewers also suggested a few things to be address in the next revision: 1). clarity in story telling flow, 2). discussion of the smoothness-prior relaxation versus a fully coupled neighborhood model, and quantify how parameter variation (e.g., σ, neighborhood size) affects results through sensitivity experiments, 3). evaluation against baseline GMM or P-EM methods, 4). details on the statistical analysis, 5).methodological clarity in e.g., notation and variable definitions, choice of Frangi filtering, image acquisition parameters, and 6). choice of evaluation strategy.
There are also many other detailed suggestions that should be addressed, but those should be rather straightforward.