List of Papers Browse by Subject Areas Author List
Abstract
Integrated photoacoustic and ultrasound (PAUS) imaging has significant potential in clinical applications due to its ability to provide complementary pathological information. By integrating a standard 1D-array US transducer with laser fibers, PA images can be acquired for vascular disease assessment. However, reconstructing 3D vascular structures remains challenging due to intrinsic limitations. While additional hardware, such as a motorized stage or motion tracker, can sequentially align image frames in 3D space, these approaches introduce additional costs and operational constraints. To overcome this limitation, we propose a free-hand PAUS framework based on windowed Bi-directional Cross-Attention with Mamba (wBCAM) for scan motion estimation and 3D US volume reconstruction. Given sequential PAUS frames, the model first encodes adjacent frame pairs into temporal local tokens. Using overlapping sliding windows, wBCAM performs local-to-global cross-attention that writes spatial evidence from local tokens into a set of learnable global tokens. The global tokens are then updated across windows using a Mamba-based selective state-space module, enabling sequential motion memory and long-range temporal reasoning. Subsequently, global-to-local cross-attention feeds back the updated global context to refine local representations within each window. The proposed method is validated on dynamic in-vivo datasets acquired from the human forearm. Both quantitative and qualitative results demonstrate accurate motion estimation and high-resolution 3D reconstruction of vascular structures. The source code of this study is publicly available at (https://**.).
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3737_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/pnu-amilab/wBCAM
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LeeSiY_wBCAM_MICCAI2026,
author = { Lee, SiYeoul AND Park, SeongKyu AND Seo, MinKyung AND Seong, EonSeung AND Lee, DongEon AND Kim, Doyun AND Kim, Sehoon AND Kim, DoYeon AND Jin, MinJu AND Lee, Hyeyun AND Kim, GyuBin AND Kim, MinWoo},
title = { { wBCAM: Windowed Bi-directional Cross-Attention with Mamba for Enhanced 3D Vascular Reconstruction in Free-Hand Photoacoustic and Ultrasound Imaging } },
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 main contribution is a sensor-free free-hand 3D PAUS reconstruction framework that adapts recent sensorless freehand US trajectory-estimation ideas to the PAUS setting, where stable long-range motion estimation is especially important because small trajectory errors visibly break vascular continuity. The paper’s specific technical contribution is the proposed wBCAM architecture: consecutive B-mode frame pairs are encoded with a ViT-style encoder, processed in overlapping temporal windows, written into learnable global tokens through local-to-global cross-attention, updated with a Mamba state-space module, and then fed back to local tokens via global-to-local cross-attention before relative/window/sequence-level motion regression. This is meant to explicitly model long-range dependencies and reduce accumulated drift. A second contribution is the application framing: unlike earlier sensorless work focused mainly on freehand 3D ultrasound reconstruction, this paper targets 3D PA vascular reconstruction from free-hand PAUS without external tracking during inference, which is clinically interesting because PA is more sensitive to motion inconsistency in thin vascular structures. The paper also contributes an in-vivo human forearm evaluation with both PAUS and US datasets, using EM-sensor trajectories as supervision during training/testing.
- 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 strongest aspect is the good problem choice. Free-hand 3D PAUS without external tracking is clinically meaningful and technically harder than ordinary freehand US reconstruction because broken trajectory consistency directly degrades 3D vessel continuity. The paper explains this motivation clearly and the qualitative figures on page 8 support that the task matters. A second strength is that the method is reasonably well matched to the stated failure mode. The paper argues that frame-to-frame/local methods drift over long sequences, then proposes a design that explicitly couples local evidence with global memory. That architectural logic is coherent: local frame-pair tokens preserve discriminative short-range motion cues, while global tokens plus Mamba provide a mechanism for longer-range temporal consistency. The multi-level heads for relative, window, and full-sequence motion are also sensible for this task. A third strength is the empirical pattern of results, especially on absolute metrics. On the US dataset, wBCAM is best on all absolute metrics; on the PAUS dataset it is again best on all absolute metrics, even though it is not best on relative metrics. That pattern is consistent with the paper’s central claim that the method mainly helps by suppressing accumulated drift rather than by maximizing short-range alignment. The ablation also supports that global tokens, windowing, and token regularization each contribute incrementally. A fourth strength is the translation toward PAUS rather than only ultrasound. Prior sensorless freehand 3D reconstruction literature is mostly ultrasound-focused, including early deep learning approaches by Prevost et al. and DCL-Net, later online/self-supervised variants, IMU-assisted models, and more recent local+global or Mamba-based US reconstructions. Extending this line into free-hand PAUS and showing actual reconstructed vascular volumes is valuable.
- 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 biggest weakness is limited methodological novelty relative to recent freehand US reconstruction literature. The paper’s individual ingredients are close to existing trends:
- sensorless freehand 3D reconstruction itself is not new, going back at least to Prevost et al. 2018 and DCL-Net 2020.
- explicit local/global context for sensorless 3D ultrasound is already central in DualTrack (2025).
- Mamba/state-space modeling for long-range freehand 3D ultrasound reconstruction was already used in FiMA (MICCAI 2024), which explicitly argues for SSMs to mine long-range spatiotemporal information.
- 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?
I would place this paper at Weak Accept because it addresses a meaningful translational problem, proposes a technically coherent architecture, and shows a convincing empirical advantage on the metrics that matter most for long-sequence 3D reconstruction, namely the absolute drift-sensitive metrics and qualitative vascular continuity. The PAUS application angle is also a real plus. What prevents a higher score is that the contribution feels incremental relative to very recent sensorless freehand US literature. FiMA already introduced Mamba/SSM-based long-range modeling for freehand 3D reconstruction, and DualTrack already emphasized joint local/global context. So the paper’s novelty lies more in the exact architectural synthesis and the PAUS use case than in a fundamentally new idea. Also, the validation is still limited to forearm data and does not yet fully demonstrate robust clinical generalization. So overall I see this as a solid and relevant paper, but not yet a top-tier standout.
- Reviewer confidence
Somewhat confident (2)
- [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 addressed well to the comments
Review #2
- Please describe the contribution of the paper
This paper presents an extension on previous work where a photoacoustic+ultrasound (PAUS) imaging probe was operated in freehand, and a DL algorithm was trained on just US data to extract the pose and position of the probe, and subsequently use those data to generate a 3D PA image from stacks of 2D PAUS images acquired in freehand. This work builds upon previous work by first extending the participant pool to 19 volunteers, second by training a network to incorporate both US and PA images to improve on the 3D registration, and third by introducing a new framework wBCAM offering greater robustness across both short- and long-distance motion inaccuracy.
- 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 methods are described clearly, and the results are compelling. Combined with the code that will be released digitally, I have no doubt others could reproduce the work. The work is timely, relevant, and of high quality - both in experimental approach and its description. The inclusion of not one, but multiple, alternative DL networks is commendable - even if the comparison of their performances is somewhat limited. (Where are the data for the “SOTA methods”, for instance?)
- 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 a limited number of places, the text is a little short on detail. Specifically:
- Section 2.1, the beamformer algorithms for both US and PA have not been described or defined
- The PA acquisition was stated at 10 FPS, but what was the acquisition rate for US? The text confirms this was higher, allowing for more accurate pose estimation, but not HOW much higher.
- Section 3.1, ratio 12:3:4 - it is unclear to me whether this means the training set contains data from all 19 volunteers, or whether data from some volunteers were used exclusively to test the algorithm. The former might have introduced bias in performance, as arguably vasculature and anatomy do not change much between consecutive scans (which I assume were performed in rapid succession?) - meaning the desired geometry for all test cases would be near-identically included in the training data. How severe is this risk of bias?
- Section 3.1, EM tracker: if these data were available, and indeed used for supervised learning, a comparison of the tracking performance of all algorithms would have been very useful. I appreciate the 3D US and PA images look convincing, but I cannot judge the positional accuracy of the scan pattern based on just the plots in Fig. 3.* Fig. 3: the US images contain multiple trajectories (the yellow, green, white, pink, … squiggles) but it is not explained what these represent. As they all deviate strongly from the final volume, which one(s) is closest to the “truth”? In addition, the addition of scale bars in each panel would be very helpful.
- 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 - ethics approval was sought and approved.
- 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?
This work is timely, relevant and convincing - and my comments above are in my opinion readily addressed.
- 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.
My comments were relatively minor to begin with and have been addressed. Reading comments from the other reviewers has not generated additional concerns on my part.
Review #3
- Please describe the contribution of the paper
The author proposed a new framework to perform scan motion estimation and 3D PA vascular reconstruction in the integrated photoacoustic and ultrasound(PAUS) system. The method incorporates a local and global attention module and the Mamba module to update the global information to improve motion estimation. Through evaluating in-vivo human arm datasets, the method outperforms other methods on absolute error performance, which shows better performance on long-range trajectory estimation.
- 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 strengthen for this paper is that it will not need an extra device for motion recording. The trajectory can be predicted only from the B-mode scanning sequence, which supports better reconstruction.
The authors provide real in-vivo data for testing the model, which was collected from human arms. The device setup is comprehensive.
- 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.Miss literature citations for “Many existing deep learning (DL)-based scan motion estimation methods primarily rely on local frame-to-frame alignment.” 2.The writing and figures are confusing, especially in the “Sequence Modeling”. Since there are different modules, many notations in text are not shown in the figure. I cannot find the temporal self-attention block that produces the temporally enhanced feature V. The V is also not shown. In Figure 2, I cannot infer from the text what the number above the arrow means.
3.The description of modeling is vague. The method of encoding the sequence frames using ViT is unknown; either treat the frame as a channel or another way. The number of learnable global tokens and the number of frames in the B-mode sequence are not explicitly stated in the text. The Mamba block is not closely explained in terms of its structure and fit for this task. The choice of coefficients of loss components is not stated.
4.If I understand correctly, the relative-level motion is a kind of frame-to-frame alignment, which may be vulnerable to long trajectories, as mentioned above.
5.The proposed wBCAM’s relative error performances are lower than those of the two other models, which raises the question about the local motion estimates, which may be considered noisy. And, how the absolute poses remain accurate. The absolute error is computed independently or by composing with relative steps.
6.No qualitative comparison with other methods.
7.The authors did not provide quantitative evaluation metrics on 3D vascular reconstruction, which is one aim in their abstract.
- Please rate the clarity and organization of this paper
Poor
- 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.
- 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.
(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?
1.The writing and presentations are so confusing without clearly explaining the key parts of the innovation. 2.No evaluation support on one of the paper’s tagert task, reconstruction. 3.The performance is lower than two other models in relative error metrics
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #4
- Please describe the contribution of the paper
The paper introduces a deep learning framework (wBCAM) for sensor-free, free-hand 3D photoacoustic and ultrasound (PAUS) reconstruction. They combine windowed bi-directional cross-attention with a Mamba-based State-Space Model (SSM) to capture long-range temporal dependencies for more stable scan motion estimation.
- 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 addressing the need for sensor-free 3D reconstruction in handheld PAUS is relevant for clinical translation. Validation on in-vivo human datasets provides more clinical relevance than phantom-only studies. Also, the ablation study demonstrates the contribution of each component (G, W, R) to the final performance.
- 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.In Table 1, for the PAUS dataset, WBCAM’s relative metrics (rAE, rCS, rFE) are notably worse than SOTA methods like MoGLO-Net and FIMA, though under low frame rate, if the model is not good to capture local frame-to-frame motion as accurately as existing method then how reliable is the final 3D topology.
2.The architecture combines a standard ViT encoder with a Mamba block. Is the novelty sufficient or is this simply a “plug-and-play” replacement of an LSTM/Transformer with Mamba? could authors explain the novelty more in detail please.
3.Figure 3 shows “depth color encoding” which is visually impressive, it masks the underlying reconstruction artifacts. A direct comparison of a single vessel’s cross-section versus the ground truth EM-tracked reconstruction may be better for judging the distortion
4.the provided URL is redacted with asterisks (https://****). In the absence of an anonymous repository or supplementary material containing the implementation details
- 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 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.
(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 weakness part.
- 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’ rebuttal addresses the primary concerns raised in my initial review.
They provided a technically justification for the discrepancy between the relative and absolute metrics. In continuous trajectory estimation, prioritizing long-range stability (absolute metrics) over myopic frame-to-frame alignment (relative metrics) is indeed the correct approach to minimize cumulative drift. Since 3D vascular topology is highly sensitive to this accumulated drift, their design trade-off is well-justified.
The rebuttal also clarified the architectural novelty, differentiating wBCAM from FiMA and DualTrack. Using Mamba strictly to propagate compact global-memory tokens across windows while preserving attention-based local motion estimation demonstrates a thoughtful, problem-specific design rather than a naive “plug-and-play” backbone replacement. Since the authors have logically defended their design choices and clarified the methodological distinctions, I am upgrading my score to recommend acceptance.
Author Feedback
We thank the reviewers for their constructive comments. We address the main concerns below.
Method novelty (R1,R4) Prior sensor-free freehand 3D reconstruction exists (Prevost’18, DCLNet’20). Our contribution is a long-range trajectory modeling mechanism for drift-sensitive 3D PAUS vascular reconstruction. Unlike DualTrack (separate local/global streams + fusion), wBCAM uses learnable global tokens as structured motion memory, iteratively written/read via windowed bi-directional cross-attention (L2G/G2L), enabling unified local/global modeling and drift suppression. Unlike FiMA (SSM on backbone features), wBCAM uses Mamba only to propagate compact global-memory tokens across windows while preserving attention-based local motion estimation. This design explicitly targets cumulative trajectory inconsistency that degrades long-range vascular continuity in freehand 3D PAUS reconstruction.
Relative vs. absolute motion (R3,R4) We do not regress absolute poses independently. We compose predicted relative transforms to obtain the trajectory and compute absolute metrics against EM-tracker-derived ground truth. Better frame-to-frame metrics do not guarantee long-range stability: small but temporally biased/inconsistent step errors accumulate into drift. wBCAM enforces long-range consistency via windowed global-memory propagation, yielding better absolute metrics and more coherent 3D vascular reconstruction. This is particularly important in long freehand scans, where small but biased local errors accumulate and distort reconstructed vascular topology.
Fig. clarity / qualitative comparison (R2,R3,R4) Fig. 3 overlays trajectories from multiple methods for space-efficient qualitative comparison; the legend was unintentionally omitted. In the US panels, the grayscale volume is the EM-tracker reference (“truth”) and colored curves are estimated scan trajectories. In Fig. 3, yellow, purple, blue, green, and red trajectories correspond to DCL-Net, MoGLo-Net, DualTrack, FiMA, and the proposed wBCAM, respectively. We will add legends/scale bars in Fig. 3 and clarify arrow numbers (processing order) in Fig. 2.
Acquisition/beamforming details (R2) US/PA images use delay-and-sum (DAS) beamforming; US additionally uses envelope detection and log compression for B-mode, whereas PA is reconstructed as initial-pressure maps without log compression. The PAUS sequence operates at 10 Hz (20 Hz laser PRF with interleaved full-aperture acquisition). The US-only dataset was acquired at 20 Hz (no laser).
Split / bias (R2) The 12:3:4 ratio is subject-level (12/3/4 subjects out of 19). All scans from a subject belong to a single split; test subjects are disjoint from train/val, so within-subject similarity cannot bias test performance.
EM tracker comparison (R2) Quantitative evaluation against EM ground truth is reported in Table 1, and Fig. 3 provides a trajectory-overlay comparison; we will make the EM reference and legend explicit.
Missing citations (R3) We will add representative citations for DL-based frame-to-frame motion estimation/registration.
Method details (R3) Temporally enhanced tokens V are produced by temporal self-attention applied after GAP (GAP → temporal self-attention → V); we will label this block and V in Fig. 2/caption. ViT encodes adjacent frame pairs (not the full sequence as channels). We will explicitly state hyperparameters (M=16, N=65 with random sub-sequences of length 64, w=16) and loss weights (λ1, λ2, λ3).
3D vascular metrics & code (R3,R4) We agree there is no dedicated quantitative metric on 3D vasculature; since 3D continuity is primarily governed by accumulated pose error, EM-GT motion metrics serve as a surrogate and are supported by consistent trajectory overlays and preserved vascular continuity in Fig. 3.Future work will validate geometry using external 3D references (e.g., CTA/MRA). The URL was redacted for double-blind review; we will expand implementation details and release code upon publication.
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.
The authors present an interesting method for vascular 3D reconstruction. The novelty of the approach wrt SoA should be better explained.
- 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 authors managed to highlight the novelty of the approach and its effectiveness wrt SoA
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.
All reviews are positive after rebuttal, and the remaining concerns have been adequately addressed. The authors clarified the key methodological points, provided useful implementation details, and the reviewers’ final opinions now support acceptance.
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.
All reviewers have reached a consensus to accept the paper.