List of Papers Browse by Subject Areas Author List
Abstract
Photoacoustic (PA) imaging is an emerging biomedical imaging modality well-suited for visualizing blood vessels due to its noninvasive, label-free nature. Registering vascular PA volumetric images acquired at different times or viewpoints is crucial for tracking longitudinal changes and expanding the field of view for comprehensive vascular assessment. However, PA volumes exhibit characteristics such as sparsity, ambiguity, vascular network changes, and unavoidable body hair, which pose significant challenges and limit the accuracy and robustness of existing registration methods. We propose a robust affine registration framework to address PA registration challenges, integrating feature-based alignment, intensity-based refinement, and hair removal. We leverage 2D feature matching with reverse mapping based on maximum intensity projections (MIPs) to handle sparsity and ambiguity, enabling robust alignment. An intensity-based refinement further enhances accuracy by incorporating our feature-guided sampling strategy to mitigate the impact of vascular network changes. Additionally, we introduce a hair removal procedure to prevent hairs from affecting registration. Experimental evaluation, conducted in collaboration with medical experts, demonstrates that our method outperforms existing approaches in both accuracy and robustness on real PA volumes.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0431_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2025/supp/0431_supp.zip
Link to the Code Repository
https://codeberg.org/ljd/pa-reg
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiaJun_Vascular_MICCAI2025,
author = { Liao, Junda and Zhou, Chu and Asano, Yuta and Suzuki, Yushi and Bise, Ryoma and Imanishi, Nobuaki and Kishi, Kazuo and Aiso, Sadakazu and Sato, Imari},
title = { { Vascular Photoacoustic Volume Registration via 2D Feature Matching with Reverse Mapping Based on Maximum Intensity Projection } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15975},
month = {September},
page = {659 -- 669}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a pipeline to perform affine registration of photoacoustic (PA) volumes of vasculature. It operates in two stages: first the authors propose a simple, yet effective, method to remove artefacts caused by hairs. Then they propose a registration process, where 1) a preliminary alignment is estimated by aligning point clouds extracted from 2D maximum projections of the PAs by a foundation model OmniGlue, and 2) the authors refine this alignment by using the well-known Elastix framework on the intensity images.
- 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 objectives and contributions of the paper are clearly stated.
- The application of the proposed method is interesting, since it focuses on lesser-known photoacoustic images, which seem promising for the community.
- The paper is well written.
- Clear and beautiful figures.
- 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.
Major concerns:
- While the paper is well written, it remains very superficial. For example, the training dataset is not even introduced (only the testing one), and the authors do not explain how they set the values of key hyper-parameters (thresholding value, number of clusters, sigma for the Frangi filters, etc.). But most importantly, there are too many missing technical details for this paper to be reproducible, with no code and no explanations given about several key parts of this work: area thresholding, hair inpainting, SVD-based approach to estimate affine matrices, solving of the optimisation problem in 2.3.
- The authors rely on a lot of existing methods such as DBSCAN, OmniGlue, Elastix, RANSAC. In addition to reducing the technical novelty of this work, these external methods are not properly introduced. Citing a toolbox is far from enough, and details must be given about how these methods work, why they were specifically chosen with respect to other equivalent approaches, and how they are used in this work.
- I think the whole evaluation pipeline is biased because all the baselines are evaluated on the data with hairs, while the proposed method benefits from hair removal. I think there should be two experiments: one to study the performance of the hair-removal procedure, and another one to study the performance of the proposed registration step. Moreover, another heavy bias is introduced for the proposed approach by the fact that the evaluation vessel segmentations are obtained automatically with the same approach that was used to separate hairs/vessels during training.
- The authors cannot claim to outperform other baselines without performing statistical tests. A mean value is not enough to claim state-of-the-art results, especially with such a low number of data points (20).
- There are very little details about the baselines: what commands are used for ElastiX and ANTs? Could the authors provide brief explanations about how SIFT3D and MUVINN work and how they were implemented here?
- The literature review is very sparse and misses entire parts of the registration field (especially feature-based registration using representation learning), which is central to this paper. Instead, the authors cite nearly all their references in a single stack of references that is hard to make sense of and contain many different methods.
Minor concerns:
- I don’t see how sparsity is a problem for affine registration. A lot of methods try to extract sparse representative features from dense MRIs and CTs. So starting from sparse images might be beneficial if well exploited.
- In the introduction: we need a citation for “perform well across various image domains”
- Why would the hairs correspond to the biggest clusters? This is far from obvious and must explained/motivated.
- The authors are wrong in saying that RANSAC performs point-cloud matching. Rather, it is used on top of other algorithms (SVD-based, etc.) to enhance the robustness of the predicted solutions. So now the question is, which algorithm?
- Figure 2 seems to indicate that DBSCAN is applied to the thresholded labels, while the text suggests it is done using intensities. Could the author clarify this point?
- The methods seem to rely on manual cropping around the region of interest. Can the authors make this requirement explicit in their abstract/introduction, and can they discuss this aspect (time required, potential for automation)?
- 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 provide sufficient information for reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors tackle interesting problem, but the paper is not mature enough for publication because of sparse literature review, many missing technical details, and heavily biased experiments. These flaws are too fatal to recommend acceptance.
- Reviewer confidence
Very confident (4)
- [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.
I appreciate the author’s response, which has clarified some of my comments. However several points remain unaddressed:
-
If the authors do not use a training or validation dataset, which data do they tune their hyper-parameters on? It has to be on the testing dataset, which is the only data used in this study. Therefore the results are heavily biased since they were obtained by a method fine-tuned on the test set (which is not the case of the off-the-shelf baselines). This goes against all principles for the evaluation of machine learning models.
-
As all reviewers agreed, there is a clear problem of reproducibility. For example, even though the intensity-based registration is a major step in the proposed pipeline (section 2.3), we still have no information whatsoever about how it’s performed. Apart from stating the optimisation problem, no method is neither proposed or cited to solve it.
-
Reporting mean scores and standard deviation is not enough, statistical tests must be provided, especially considering the limited size of the dataset (20 cases). The authors have refused to add statistical tests.
As a whole, I think this is an interesting paper, but it is not mature for publication, as the rebuttal has unveiled a major flaw in the experimental design (the proposed method is fine-tuned on the test set) and there is a clear problem in reproducibility (lot of missing methodological aspects).
-
Review #2
- Please describe the contribution of the paper
This paper proposes an affine registration paradigm for vascular Photoacoustic (PA) volumes. The proposed method comprises a clustering-based hair removal and feature-based alignment, followed by intensity-based refinement. The feature-based alignment uses OmniGlue and reverse-mapping techniques to extract and match 3D key points from the input. Comparing the intensity-based registration methods, results on 20 pairs of vascular PA volumes demonstrate that the proposed method significantly improves target registration error and is more robust to hair artefacts and vascular structure changes.
- 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 proposed method is technically sound and promising for Photoacoustic volume registration. The registration pipeline can address the unique challenges in PA volume registration: sparsity, ambiguity, vascular structure changes, and hair artefacts.
The core idea of hair removal is interesting and practical. It uses a combination of simple image-processing techniques to address the hair artefacts. This step does not involve learning, showing great generalizability to the problem.
The reverse mapping technique is novel. It uses maximum intensity projection of the images and off-the-shelf keypoint matching to perform accurate and fast 3D keypoint extraction.
A comprehensive evaluation of the proposed method is provided. Quantitative and qualitative results show that the proposed method is superior to classical intensity-based methods, feature-based approaches, and INR-based methods for PA registration. It achieves an average TRE of 1.3mm—1.34 mm, which is impressive in PA registration.
- 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.
There are no significant weaknesses in this paper. Below, summarize the minor weaknesses of the paper.
Limited reproducibility. Since the method consists of many components (Feature-based Alignment, hair removal, and intensity-based refinement) and the experiments are conducted with a private dataset, it is difficult to reproduce the method and the experiment with the source code and the proposed method. The submission does not mention open access to the source code.
The selection of hyperparameters could be task-specific. The hair removal procedure contains a couple of thresholding methods. However, the authors did not mention how the threshold is computed/estimated.
- 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
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 is an interesting and practical solution for affine PA volume registration. A comprehensive evaluation is provided. This paper is of great interest to the MICCAI community. If the authors open-source their source code and PA volumes, it would create a long-lasting impact in PA volume registration.
- 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 rebuttal addressed the concern about reproducibility. While the proposed method combined off-the-shelf methods from computer vision, the whole pipeline is a practical solution for the affine registration of photoacoustic (PA) vasculature volumes. Therefore, I will recommend “Accept.”
Review #3
- Please describe the contribution of the paper
The authors introduce an approach for 3D affine registration of photoacoustic images of the vascularity of the skin and possibly deeper tissues. The goal is to align 3D vascularity images acquired at different points in time or in different orientations. The approach combines several methods that all use a Frangi filter, including a hair removal step, a feature-based registration step for handling large displacements with 2D projection and reverse projection, and an intensity-based registration step to refine the results by focusing on regions with vascularity. The approach has been successfully applied to 20 pairs of 3D images, and it shows superior results in comparison to standard approaches. An ablation study demonstrates that the main reason for the successful results is the feature-based registration step since the image pairs seem to be not well aligned.
- 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 application is relevant and the proposed approach is novel and sound
- The paper is mostly well written and the methods are clear
- Feature mapping in 2D maximum intensity projections of sparse vesselness maps with 3D reverse projection is interesting and is the key reason for the successful result
- The experiments are based on a sufficient amount of 20 pairs of 3D photoacoustic images, quantitative values are presented for TRE and Dice, a comparison with four standard approaches is performed, and an ablation study performed
- 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 description of the application could be more specific (which tissues are considered, how deep is laser penetration, what kind of US transducer is used, is tissue submerged in water), only Fig. 1a gives a hint
- Since all standard approaches struggle with the large displacement between the image pairs, my impression is that the comparison is not fully fair in such that the chosen standard approaches are not well suited for these images. Maybe the authors could comment why those methods fail for large displacements? For a future journal publication, I suggest to use some pre-registration for all approaches, or apply the standard approaches also after the feature-based mapping.
- Presentation of paper could be improved (see optional comments below)
- The authors do not discuss how robust the reverse mapping step is since it is ambiguous if two or more vessels are mapped to the same pixel in the maximum intensity projection. It is also not clear how OmniGlue is used for reverse mapping and why it is strong generalizable to PA images
- 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
- More details about application would be helpful
- Fig 1b is not referenced and not useful, I recommend to remove it and use the space to better elaborate on application (note that Fig 1a is not referenced as well, should be done in Sect. 1)
- Table 1 mentions pre-registration “Pre-reg”, but this is not mentioned in the paper elsewhere. Has there been a pre-registration performed? If yes, which one?
Some minor comments:
- “(MIPs)” introduced in abstract but not used
- Section 1: “PA” not introduced (only “PAI”)
- Section 2: Matrix A* should be 4x4, not 3x4 (same homogenous matrix is defined in Sect. 2.2)
- 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?
Novel approach for relevant application that yields promising results based on quantitative measurements for 20 3D image pairs. Though none of the components of the approach represents a significant novelty, the approach as a whole is novel.
- 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 clarified some of my concerns in their rebuttal, and I believe that it is possible to address them for the final paper.
Author Feedback
We sincerely thank all reviewers. Two reviewers (R2&R3) recommend Accept and commend the novelty of our registration pipeline based on 2D feature matching and reverse mapping. Some of R1’s major concerns appear to stem from a misunderstanding of our proposed method, which we believe can be clarified below:
- Lack of training dataset and related learning-based references: We would like to clarify that our work does not involve developing or training any learning-based models. The only external deep model is OmniGlue, used off-the-shelf with the authors’ released weights. We will cite representation learning based methods and clarify that our approach is not learning-based in the final version.
- Biased evaluation (only using data with hair): Contrary to R1’s comment, we respectfully clarify that Table 1 separately reports results for data with and without hair (“hairy” vs. “normal”), ensuring identical input conditions across all methods. Table 2 presents an ablation study (“Ours” vs. “w/o HR”) to isolate the hair removal effects.
- Novelty & external methods: We would like to emphasize that our main contribution lies in overall pipeline design rather than individual modules, as supported by R2 (“reverse mapping is novel”) and R3 (“approach as a whole is novel”). Further details on external methods and our selection rationale will be provided in the final version.
- Limited statistical evidence & small dataset size: PAI is relatively new, with limited clinical adoption, no public datasets, and restricted access due to privacy constraints. We provide mean and std in Tables 1&2 to demonstrate average performance and robustness. We would like to highlight that R2 considered our evaluation comprehensive, and R3 regarded our dataset as sufficient.
- Literature: We mainly discussed PA-related registration methods and baselines; other registration methods were briefly cited together.
- Minor concerns: Sparsity motivated our use of feature matching; intensity-based methods rely on non-smooth metrics with many local minima, worsened by sparsity, making optimization difficult. We will add missing citations on feature matching. As requested, we will clarify the following details in the final version: 1.After size filtering, most regions are hair, thus largest clusters primarily select hair. 2.RANSAC is applied on a total least squares method using SVD (skimage.transform.AffineTransform). 3.DBSCAN clusters regions based on intensity histograms and orientation. 4.ROI cropping takes 5 min manually and can be automated via whole-volume feature matching.
Additional clarifications for the final version include:
- (R1&R2) Reproducibility, hyper-parameter settings, baseline details: We will release code, hyper-parameter settings, baseline settings upon acceptance. Key settings for hair removal, thresholding is adjusted based on intensity distribution. We normalize intensities to [0,1] and use a threshold of 0.1 for our dataset. Size filtering retains regions with 30–400 voxels, reflecting typical hair sizes of our dataset. We retain clusters covering 80% of these regions.
- (R3) Application: Our method registers vasculature in PA images, regardless of tissue type. The dataset used [21]’s PA system designed for human limbs, featuring a hemispherical detector array, ~2 cm penetration depth, and requiring tissue submersion in water during imaging.
- (R3) Baseline performance: Elastix, ANTs, and MUVINN rely on non-smooth similarity metrics; large displacements further complicate optimization. SIFT3D fails due to limited keypoint detection/matching.
- (R3) Reverse mapping & OmniGlue: OmniGlue finds matched 2D keypoints on MIPs; reverse mapping recovers their 3D positions. Errors of reverse mapping are handled by RANSAC during affine estimation. OmniGlue leverages a vision foundation model for strong generalization.
- (R3) “Pre-reg”: it means before registration. We will revise to “before-reg”.
- (R3) Writing: We will revise writing and math notation.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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”.
N/A
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This is a tricky paper to evaluate, as it is really at the boundary between a method paper and an application paper. On one hand, the proposed pipeline is clearly useful for the problem at hand, and would find a readership. On the other hand, it is closer to a pipeline of existing methods than to a novel methodology, and as pointed by R#1, the experimental design of the validation is flawed. All things considered, I find that this paper is just below the level of maturity required for presentation at MICCAI, and I recommend rejection.
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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper proposes a method for affine registration of vascular Photoacoustic volumes, where an initial alignment is realized by feature-based reverse mapping, followed by an intensity-based refinement. Overall, the proposed method is practical, novel, and accurate while tackling a much less explored topic for medical image registration. Although there are some concerns regarding reproducibility, the paper itself has a good value for photoacoustic imaging, which offers unique insights into the biological tissues.