Abstract

Lissajous confocal laser endomicroscopy (CLE) is a promising solution for high-speed in vivo optical biopsy for handheld scenarios. However, Lissajous scanning traces a resonant trajectory and samples only the visited pixels per frame; at high frame rates, many pixels remain unvisited, creating structured holes. In this work, we introduce the first benchmark for 10 Hz Lissajous CLE, consisting of low-quality video clips paired with high-quality reference images. The reference images are wide-FOV mosaics obtained by stitching stabilized, slow-scan frames of the same tissue, enabling temporally aligned supervision. Using this dataset, we propose MIRA, a lightweight recurrent framework for Lissajous CLE restoration that iteratively aggregates temporal context through feature reuse and displacement alignment. Our experiments demonstrate that MIRA outperforms both lightweight and high-complexity baselines in restoration quality while maintaining a favorable computational efficiency suitable for clinical deployment.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4278_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)

https://huggingface.co/datasets/2mm/MaLissa

BibTex

@InProceedings{LeeMin_Multiframe_MICCAI2026,
        author = { Lee, Minhee AND Lee, Sangyoon AND Lee, Jiwook AND Hong, Minki AND Kim, Kyuyoung AND Kim, Won Hwa AND Lee, Jaeho},
        title = { { Multi-frame Restoration for 10 Hz Lissajous Confocal Laser Endomicroscopy } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a novel method for multi-frame restoration in Lissajous Confocal Laser Endomicroscopy (CLE). The authors propose a deep learning framework that reconstructs high-quality (high-quality, low-noise) images from low-quality, high-frame-rate acquisitions. A key contribution is the creation of training pairs by matching low-quality frames to samples extracted from a large mosaic of high-quality images. The model is based on a U-Net architecture and incorporates historical scanning information as additional input features. Experiments are conducted on a self-collected dataset of porcine stomach tissue.

  • 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 paper is among the first to leverage mosaicing and frame matching to construct paired training data for CLE image restoration.

    2)The data curation pipeline is carefully designed and provide a meaningful technical contribution.

    3)The approach is well-motivated by the inherent trade-off in CLE systems between temporal resolution and image quality.

  • 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 evaluation scope: The method is only evaluated on a single dataset. It is unclear how well the model generalizes to other tissues, acquisition settings, or devices.

    2)Potential data leakage / overlap: The paper does not clearly describe how training and validation splits are constructed. If frames are sampled from the same spatial regions or video sequences, there may be overlap, which could overestimate performance. This is particularly important for assessing generalization to unseen data.

    3)Lack of comparison clarity: The differences between the proposed method and state-of-the-art video restoration approaches such as BasicVSR++, eMVD, and Turtle are not sufficiently discussed in introduction or method.

  • 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 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

    1)Frame formulation (Section 2): In the description “Given a window of preceding frames (I_{t−τ}, …, I_{t−1}), we estimate…”, should the current frame I_t also be included as input? The subsequent equation appears to include I_t, which creates ambiguity.

    2)Interpolation of missing pixels (Section 3): The manuscript states that missing pixels are filled using interpolation to obtain dense images. Could this introduce artifacts or bias during training? Why not mask out missing pixels and compute the loss only on observed regions?

    3)Template matching details: The paper mentions replacing phase correlation with template matching due to motion artifacts. How exactly is template matching implemented (e. g. , search window size, similarity metric)? Is it robust to large motion or deformation?

    4)Feature alignment reliability: The method aligns cached features {E^l_{t−i}} (i=1 to 4) and d̃_{t−1} to the current frame I_t. How accurate is this alignment in practice? Since alignment depends on feature quality, how stable is this process during early training before the encoder has converged?

  • 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 see the main weakness.

  • 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 several of my main concerns, and I now consider the paper to be marginally above the acceptance. I recommend that the authors further clarify and justify their evaluation choices (e.g. comparison to state-of-the-art methods, down stream evaluation), as this would better position the paper’s contribution. I also hope the authors will release their code and data as promised and make them easily accessible, which would enhance the reproducibility of the work.



Review #2

  • Please describe the contribution of the paper

    This paper addresses the undersampling challenges in high-rate Lissajous-scanning confocal laser endomicroscopy (CLE) by proposing a multi-frame iterative restoration framework named MIRA. The approach utilizes a flow-free registration module based on phase correlation to align features from sparse pixels and incorporates a recurrent feature caching mechanism to reduce inference latency. The authors also introduce MaLissa, the first public paired dataset for Lissajous CLE restoration. Experimental results report a PSNR of 22.88 dB at a 10Hz sampling rate with an inference latency of approximately 236.6 ms.

  • 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.

    Dataset Contribution: The construction and public release of the MaLissa dataset provide a foundational resource for future research in the specialized field of Lissajous-scanning CLE.

    Algorithmic Design: The introduction of a phase-correlation module specifically tailored for extremely sparse pixel alignment effectively bypasses the limitations of traditional optical flow methods when dealing with substantial missing data.

    Efficiency Exploration: The use of feature caching and ResFFT modules demonstrates a meaningful attempt to balance restoration performance with computational latency for potential clinical applications.

  • 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.

    Rationale for the Scanning Strategy and Task Setup: The introduction claims that conventional raster scanning is mechanically limited at high speeds. However, mature clinical CLE systems (e.g., Cellvizio by Mauna Kea Technologies) utilizing high-frequency resonant scanners (e.g., 4kHz or 8kHz CRS) already achieve high-quality, full-sampling imaging at frame rates of 12-15 fps.

    Clarification Required: What is the specific clinical advantage of the Lissajous scheme over these established raster-scanning solutions at the 10Hz scale? Why choose a scheme that results in 70% missing pixels at 10Hz only to necessitate complex algorithmic repair, rather than utilizing existing full-sampling hardware?.

    Lack of Quantitative Resolution Validation: The core clinical value of CLE is cellular-level resolution. Current evaluation relies solely on computer vision metrics like PSNR and SSIM, which do not confirm whether the algorithm authentically restores physical high-frequency structures or merely performs visual smoothing.

    Generalization to Higher Frame Rates: While the title emphasizes “High-rate,” experiments only validate performance at 10Hz. As frame rates increase to 30-60 fps, the sparsity of the Lissajous trajectory will increase further and the Signal-to-Noise Ratio (SNR) will drop precipitously. The robustness of the algorithm under such extreme undersampling remains unproven.

  • 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.

    (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?

    1.Summary of Recommendation Factors The recommendation is primarily influenced by the tension between the paper’s methodological innovation and its clinical feasibility. While the work offers a novel application for Lissajous scanning in CLE , it faces significant challenges regarding its practical impact. 2.Major Strengths (Positive Weighting)Novelty of Application: The paper introduces the first open benchmark for Lissajous CLE restoration, which is an original way to use data for this specific imaging modality. Algorithmic Contribution: The MIRA framework’s use of flow-free registration and recurrent caching addresses a technically difficult problem (sparse pixel alignment) in an interesting way.Contribution to the Community: By releasing the MaLissa dataset, the authors contribute a resource that others in the MICCAI community may find valuable for benchmarking. 3.Major Weaknesses (Negative Weighting) Clinical Feasibility and Significance:A significant concern is the clinical impact. Current commercial systems (e.g., Cellvizio) already achieve high-quality 12-15 fps imaging without the 70% pixel loss inherent in the authors’ 10Hz setup.The end-to-end latency (~340ms) significantly exceeds the requirements for real-time surgical guidance, which undermines the claim of “clinical deployment”. Evaluation Rigor:The paper lacks a particularly strong evaluation of physical resolution. Relying solely on PSNR/SSIM fails to prove cellular-level fidelity, which is the primary value of CLE.The results are mostly incremental over previous work by focusing on a relatively low 10Hz rate, rather than demonstrating a substantial contribution at truly “high” rates (30+ fps). 4.Weighing of Factors Methodology vs. Impact: While the U-Net architecture and phase-correlation modules are sound , the clinical significance of the task setup is currently insufficient to outweigh the methodological gaps in resolution validation.Reproducibility: The paper provides sufficient algorithmic detail , but the lack of “hard” physical evidence (like a USAF 1951 target test) limits the informed expert assessment of its true performance. Conclusion: The paper is an interesting contribution , but its recommendation is contingent on the authors’ ability to substantiate their “High-rate” and “Intraoperative” claims with evidence of real-world resolution improvement and latency reduction.

  • 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.

    Overall Assessment: The authors have successfully addressed my concerns in the rebuttal. They have explained the advantages of constructing a Confocal Laser Endomicroscopy (CLE) system based on a Lissajous scanning scheme compared to existing commercial counterparts (e.g., Cellvizio, CONVIVO). Specifically, this approach achieves high-frame-rate scanning while effectively avoiding the honeycomb artifacts inherently introduced by fiber bundles. To address the sparse sampling caused by high-frame-rate scanning—which is an inherent characteristic of Lissajous trajectories—the authors have developed an innovative image reconstruction algorithm and made their dataset publicly available.

    Constructive Comments for Future Work: However, the current reconstruction performance was validated only at a 10 Hz frame rate. I strongly recommend that future work test the algorithm at higher frame rates to fully demonstrate the high-speed imaging advantages of Lissajous scanning. If the scope remains restricted to 10 Hz, current commercial systems (such as Cellvizio from France, VieStar and Hisky from China) can already achieve comparable temporal and spatial resolution. Additionally, the authors are encouraged to evaluate the reconstruction quality using a standard 1951 USAF resolution test target in future studies. For system developers in microscopy, physical metrics like line pairs per millimeter (lp/mm) are critical, as they objectively reflect the comprehensive spatial resolution of the system. Relying solely on downstream task performance may not provide a sufficiently objective validation.

    Conclusion: While certain aspects of the study could be further extended, the work exhibits a high degree of innovation and represents a significant contribution to the field of CLE. Therefore, I recommend acceptance of this manuscript.



Review #3

  • Please describe the contribution of the paper

    The paper introduces the task of high-rate Lissajous CLE restoration and presents MaLissa, which appears to be the first open paired benchmark for this setting. Because the low-quality high-rate frames and high-quality low-rate frames do not share the same field of view directly, the authors build the dataset through an HQ stitching and LQ-to-mosaic matching pipeline, producing 2,060 curated LQ–HQ pairs from 16 porcine stomach clips. On top of that, they propose MIRA, a lightweight recurrent U-Net-based restoration model that combines feature caching, phase-correlation-based global registration, and lightweight refinement modules, and they evaluate it against several video-restoration baselines with both restoration and downstream tasks.

  • 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 dataset is a real contribution. High-rate Lissajous CLE is a meaningful setting, and paired supervision is clearly hard to obtain because the LQ and HQ acquisitions do not naturally align. The matching-based construction pipeline is therefore not just bookkeeping; it is a practical workaround to a real data bottleneck, and it makes the paper more than just a model paper. The overview in Fig. 2 on page 4 makes this pipeline quite clear.

    2.The method is reasonably well matched to the task. The choice to use global phase-correlation registration instead of optical flow makes sense for sparse frames with many missing pixels, and the recurrent feature cache is a sensible way to reuse temporal information without paying the cost of heavy transformer-style processing at every frame. The architecture on page 5 is also fairly clean and easy to follow.

    3.The empirical tradeoff is strong enough to be interesting. In Table 1 on page 6, MIRA gives the best PSNR, SSIM, and MS-SSIM while staying much lighter than the transformer baselines, and it clearly outperforms the very lightweight EMVD. The downstream results on page 8 also help: retrieval improves quite a bit, and segmentation metrics are at least directionally consistent with the restoration results. That makes the paper feel more practically grounded than a purely pixel-metric comparison.

  • 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 evidence for some of the model-specific claims is not very strong. The ablation table on page 7 only weakly separates the full model from the variants: removing registration or FAM barely changes PSNR/SSIM, and the clearest gain comes from rejection sampling, which is more of a data/training pipeline contribution than a core architectural one. There are also several important design choices that are left unablated, including the 5-frame window, the loss weight, the correlation threshold, the rejection threshold, and the block size in rejection sampling.

    2.The supervision is useful but still somewhat narrow and indirect. The HQ targets are matched mosaic crops rather than synchronized ground truth, and the paper itself notes brightness inconsistencies and alignment issues severe enough to motivate rejection sampling. The data are also limited to one CLE system, one tissue type, and only 16 raw clips. Since 2,060 pairs are selected from 3,988 raw frames, it would have helped to know more about whether the retained subset is biased toward easier or lower-motion cases.

    3.The deployment and clinical framing goes a bit beyond the evidence. The paper presents MIRA as a practical step toward real-time intraoperative use, but its measured latency is still 236.66 ms, while the conclusion itself says a 10 Hz target would ideally require below 100 ms per frame. The downstream segmentation experiment is also based on MedSAM pseudo-labels built from bounding-box prompts rather than clinician annotations, and the paper reports 95% upper-trimmed means without also showing the untrimmed results. So I think the practical story is promising, but not fully established yet.

  • 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.

    (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 lean weak accept because I think the paper makes a worthwhile contribution at both the dataset and method level. The task is interesting, the dataset construction is useful in its own right, and MIRA shows a convincing quality–efficiency tradeoff with some downstream support. My reservations are mainly about how strongly the paper justifies its model-specific choices and how broadly the results can be expected to generalize, given the indirect supervision and narrow data source. So for me this is a borderline accept: the contribution is real, but the empirical case still has some important limitations.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    I maintain my weak accept after the rebuttal. The authors addressed several of my main concerns, particularly regarding the data split, matching procedure, and potential bias from retained frames. Although some weaknesses remain, including limited ablations, pseudo-label-based evaluation, and latency/clinical validation concerns, I do not think they outweigh the paper’s contribution. The dataset and task are valuable, and the work provides a useful first benchmark for Lissajous CLE restoration.



Author Feedback

We appreciate the thoughtful feedback from the reviewers and AC. Rationale for Lissajous CLE (R2): Cellvizio is not a “clean raster” but a post-processing dependent platform with a structural pixel ceiling. Cellvizio drives a 30k-fiber bundle with a MEMS raster scanner, and thus has HW-bound limits: (1) honeycomb blanks between fiber cores, which Cellvizio addresses with dedicated post-processing (Le Goualher’04, Bedard’12); (2) a ~30k pixel ceiling fixed by core count, with FOV shrinking to 240 μm at high-res. Single-fiber raster systems (e.g., Convivo) avoid bundle artifacts but are limited by low frame rate from single-axis resonance, causing severe clinical drawbacks (Xu’25). Lissajous CLE bypasses both ceilings: single fiber removes the bundle pixel ceiling, and dual-resonant scanning decouples frame rate from per-axis line count. Just as honeycomb removal is standard for Cellvizio, Lissajous reconstruction is intrinsic to this architecture. Our work tackles this challenge. 10Hz vs. High-rate (R2): We agree that this can be confusing. We’ll revise “high-rate” to “10Hz.” We target 10Hz due to its clinical relevance, as established in prior works [6,7]. Details on train/val split (R1, R3): MaLissa is splitted at the clip level, where each raw clip was acquired from a distinct non-overlapping spatial location on the tissue. Each clip exhibits significantly distinct visual characteristics (Figs 4,5). Through this split, we aim to assess generalization under a challenging setup, without leakage. Narrow data scope (R1, R3): This is why our dataset—MaLissa—is a meaningful contribution. We contribute the first open dataset for Lissajous CLE, and we could not evaluate on other datasets as there is none. Also, we use only one device, as there is only one Lissajous CLE system on the market. Comparison clarity (R1): To clarify, the comparison is in the last paragraph of sec.2.We’ll relocate them to the introduction, as suggested. Latency (R2, R3): We clarify that Tab.1 results correspond to the “end-to-end latency,” whereas matters for 10Hz deployment is the “throughput.” In fact, we can parallelize multiple frame-wise processing threads via pipelining—i.e., start processing the next frame whenever we’re done encoding (not full processing) the previous frame. This can reduce the per-frame gap to under 100ms. We’ll clarify this point in the revised manuscript. Questions on ablation (R3): We optimized for visual clarity over PSNR/SSIM, and the current configuration works clearly the best in this regard (Fig. 5). On MedSAM pseudolabels (R3): The key takeaway of segmentation results is that our method is effective in preserving structural consistency. Although not fully-clinical, we believe these results are still meaningful. Biased to low motion? (R3): It turns out that the ratio of large-motion frames (>100px) remains similar after matching (>15%). Thus, the bias does not look very severe. We’ll clarify this in the revised version. Too focused on PSNR/SSIM (R2): While we acknowledge the lack of physical resolution validation, we believe our downstream task evaluation can be a complementary evidence for functional resolution; indeed, several recent microscopy works (e.g., MiHATP, Joint EM) utilize similar PSNR/SSIM+Task evidence as main evaluation metrics. Retrieval and segmentation tasks can be a good proxy, as these tasks require recovering high-frequency details to achieve a high accuracy. Equation typo (R1): Thank you. Will fix! Potential interpolation artifact? (R1): Measured pixels in 2Hz images are quite dense with <10% pixels unmeasured, and thus do not introduce such artifacts. Template matching details (R1): We used OpenCV TM_SQDIFF_NORMED with 2048^2 window. Empirically, this is robust to large motion; the deformation is not significant in our data [10, 18]. Feature alignment stability in early stage (R1): It is stable even at early stages, as the displacement estimation relies on raw model input, not encoder features.




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.

    All reviewers are inclined to finalize their decisions following the rebuttal. The authors are advised to address the concerns raised by Reviewer 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 reviewers have reached a consensus to accept the paper.



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.

    After the rebuttal, the three reviewers have converged towards acceptance. Despite certain identified limitations, notably regarding the lack of a quantitative resolution metric and a somewhat limited validation, the reviewers agree on the innovative value of the contributions: the first dataset for Lissajous CLE restoration and the introduction of a phase-correlation module for sparse pixel alignment.



Meta-review #3

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

    Accept

  • Please justify your recommendation.

    The paper addresses an interesting problem and includes both a new dataset and a restoration framework for Lissajous CLE imaging. The reviewers raised concerns about evaluation scope and clinical motivation, but the rebuttal clarified several important points, particularly regarding the scanning setup, data split, and deployment considerations. Overall, the paper provides sufficient contribution and evaluation to support acceptance.



back to top