Abstract

Augmented reality from preoperative 3D model registration is promising to assist navigation in minimally-invasive liver surgery. The current registration methods are either accurate, but require surgeon interactions to annotate anatomical landmarks, or are fully automatic, but inaccurate. We propose a two-step automatic and accurate registration method. Step 1) segments the registration landmarks with a neural method. Step 2) estimates the 3D model deformation from the landmarks. The task is challenging because of the defects of the automatically segmented landmarks and the impossibility to label registration for training. We handle it by combining supervised training from synthetic transformations with domain adaptation and a novel robust Run-Time Optimisation (RTO). Our method outperforms existing ones, both with manual and automatic landmark segmentations, improving both automation and accuracy.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1316_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2025/supp/1316_supp.zip

Link to the Code Repository

https://github.com/EmilienGad/ADeLiR.git

Link to the Dataset(s)

https://encov.ip.uca.fr/ab/code_and_datasets/datasets/llr_reg_evaluation_by_lus/index.php

BibTex

@InProceedings{GadEmi_Automatic_MICCAI2025,
        author = { Gadoux, Emilien and Bartoli, Adrien},
        title = { { Automatic Deep Deformable Registration using Domain Adaptation and Run-Time Optimisation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15968},
        month = {September},
        page = {65 -- 74}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper present a method for registration in minimally-invasive liver surgery. The main contribution in a deep learning scheme that first estimate a global transformation then refining it to capture the deformation. The network learns to map landmarks with 3D transformation by regression. Domain adaption is used to generalize to unseen landmarks. Results on synthetic and real data are shown.

  • 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.

    -Using landmarks for registration is a suitable solution. -The related works covers the recent work. -The DA architecture for the coarse registration is interesting

  • 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 method is incremental w.r.t [10] and [11], it is technically sound but does not bring novelty to the problem. Using silhouettes and training a regressor is common place. -It is unclear why Domain Adaption is required. Using landmarks is an abstraction to using the images. Training on all possible rigid and non-rigid landmarks transformation should suffice to generalize to unseen data transformation. Table 1 suggests a limited impact of the DA to the overall results. Data augmentation could solve the issue of missing/perturbed points. -Results don’t address the cases where the silhouette is partially visible. A sensitivity study would have been informative. -Why using FEM for deformation simulation, ARAP or FFD could be an option? Does the DA also learns to adapt to the deformations or only for the coarse registation? -Minor: text in figures is too small to read.

  • 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

    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 — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This work tackles an important and challenging problem. The contribution of this paper is however too incremental w.r.t previous work. The addition of the DA module in innovative but is not justified when using landmarks that abstract images and results are not clear for the DA impact on the overall results.
    In addition, the DA seems to be used only for global registration, reducing thereby its usefulness.

  • 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.

    Authors addressed my comments. To improve the overall quality of the paper, I suggest AUthors to improve the clarity of the papers and the readability for the figures.



Review #2

  • Please describe the contribution of the paper
    • This paper presented a 2D-3D nonrigid registration method for liver navigation. The method consists of two main components: domain adaptation to accommodate noisy, automatically detected landmarks, and run-time optimization component for deformable registration.

    • The method is evaluated on a public 2D-3D liver dataset with ground truth tumor positions provided, the results demonstrate superior performance than baseline methods. Ablation study presents the effectiveness of the designed components.

  • 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 innovative in the field of 2D-3D deformable liver registration.

    • The method outperforms baseline methods in terms of TREs for most patients. The inference time is acceptable

  • 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.

    General:

    • While using domain adaptation to compensate for imperfect segmentation is interesting, it may be considered a workaround for suboptimal automatic landmark detectors adopted in this pipeline. Also, more work integrated with foundation models [1] came out, and they seem to have achieved good segmentation results. Could the author justify the significance and necessity of the proposed domain adaptation in this context?

    • The text requires careful editing for better demonstration. For example, please introduce the full name when the abbreviation first appears in the text. In related work: “DA aims at reducing the domain gap,…”, introduce DA. In sec 4.2, “to visualize the ADeLiR results”, introduce the full name of AdeLiR here.

    The readability of figures and tables can be greatly improved. For example, in Fig 3, the laparoscopic images are obviously squeezed, and the algorithm part is somehow blurry.

    Method:

    • From the demonstration in Fig.1, the liver mesh first undergoes nonrigid registration then the deformed mesh is affine transformed to match the intraoperative 2D image. In my view, this design is unintuitive compared to methods that apply rigid alignment first. Moreover, the nonrigid deformation is achieved by minimizing the Hausdorff distance between the projected liver landmarks and intraoperative landmarks, so the predicted deformation can be unrealistic as it attempts to cover the rigid part. Could the author justify such a design, or is it a wrong delineation in the figure?

    • Following the above, the liver meshes in Fig. 3 exhibit “spikes” on the top part. Is this caused by the design described above?

    • It is unclear how the domain adaptation is trained. What are the real images and segmentations? Fig. 2 denotes them as “Test”, are the images from the test set involved in training? If other automatic segmentation methods that yield different types of segmentation errors are deployed, does the DA need to be re-trained?

    • Hausdorff distance is quite sensitive to outliers, which often exist in the results from automatic segmentation algorithms. Could the author elaborate more on this? Would other loss functions help, e.g. pixel-wise difference?

    Experiment:

    • If I understand correctly, w/o RTO means only rigid alignment is applied and no deformation is predicted. According to the ablation study, the discrepancy between the results of rigid registration (without RTO) and the results of nonrigid registration (with RTO) is trivial - most of them are within 1mm. S,o one could raise concerns that nonrigid registration may contribute little, and the main error decrease comes from rigid alignment. Could the author justify this?

    • Details of the training set are missing. What kind of liver models are used in affine and deformable simulation? From which dataset? Public or private? What is the initial position of the 3D liver models?

    • The result tables are somewhat difficult to parse. I would suggest grouping the results in a different way for Tab. 1 and 2 so that corresponding columns will appear adjacent, and the best is easier to find.

    • In Figure 4, the bottom images should be the P2ILF dataset if I checked them correctly, but there is no reference. Also, Fig. 4 is not referred anywhere in the text.

    misc:

    • In patient 2 results in Tab 2, the lowest error for manual ones is the “ADeLiR” column, not the “w/o DA, w/o RTO” column.

    Reference: [1]. Pei J, Cui R, Li Y, Si W, Qin J, Heng PA. Depth-Driven Geometric Prompt Learning for Laparoscopic Liver Landmark Detection. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention 2024 Oct 3 (pp. 154-164). Cham: Springer Nature Switzerland.

  • 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.

  • 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.

    (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 proposed method is interesting in combining domain adaptation with run-time optimization for deformable 2D–3D registration, and it shows better results on a public dataset than baseline methods. While some design choices and experimental details require clarification, the overall contribution could be valuable to the 2D-3D registration community.

  • 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.
    • Most of my main concerns have been addressed pretty well by the authors, particularly the structure of the designed method.

    • Although the gains in deformable registration are a bit limited, the introduction of runtime domain adaptation is quite novel in 2D-3D registration.

    • However, careful manuscript polishing is highly recommended for better clarification, especially the figures and the method descriptions.



Review #3

  • Please describe the contribution of the paper

    The paper proposes an approach for registering preop 3D model of the liver onto real time 2D images. The method has 2 steps. First, a coarse landmark-based affine registration is performed by a neural network that has been trained with adversarial unsupervised domain adaptation. A non-rigid registration is then performed by optimizing over the latent space of an autoencoder trained on realistic deformations of the patient’s liver. The proposed method outperforms SoTA approaches on a standard public dataset. An ablation study is communicated to evaluate the impact of each of the two contributions.

  • 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 describes clearly the issues it aims to solve and how the solutions it proposes actually solves them. The use of UDA techniques to handle imperfect automatic inputs is interesting and has a significative effect. The experiments are thorough with a comparison with 2 state of the art methods and an ablation study. The results on the standard benchmark are very good.

  • 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 low performance of Vision Transformers that is communicated raises the question of whether they were optimally trained. Some more details on the autoencoder used in the RTO step would be welcome (what type of layers does it use?). It would be interesting to have comments about the contribution of RTO in the automatic landmarks setting, since it only reduces the TREs by 0.5mm on average. Does it make a difference in practice?

  • 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.

  • 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

    In other domains, approaches optimizing over the latent space have been outperformed by conditioning diffusion models. Is it something worth investigating?

  • 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?

    The paper proposes an interesting approach that outperforms state of the art on the problem. The use of UDA techniques to solve the issue of imperfect automatic input labels is of particular interest.

  • 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 rated this paper Accept pre-rebuttal




Author Feedback

We thank reviewers for their assessment. We respectfully point out misunderstandings in R1. We will revise the final paper to clarify and resolve all points.

R1

Comment: The method is incremental wrt [10] and [11], it is technically sound but does not bring novelty to the problem.

Resp: Our method combines neural inference and numerical optimisation (Sec 3.1, Fig 1); in contrast, [10] does not use neural inference at all and requires clean manual landmarks and silhouette; [11], while being automatic, is far in accuracy from manual methods (Sec 2). [11] does not address the domain gap between simulated training data and real test data, and does not use numerical optimisation, two major technical differences with our work. See Tab 1, column “LMR/A” for [11], with TREs of 33.3 and 29.2 mm (with and w/o P2), dropping, in column “ADeLiR/A” (proposed), to 22.08 and 13.96 mm resp. (ablation study in Tab 2 and 3 for the benefits of each component).

Comment: It is unclear why Domain Adaptation is required […] Table 1 suggests a limited impact of the DA to the overall results

Resp: We use training data simulated from the liver model prior to surgery (Sec 3.2). We extensively sample the simulation parameters and randomly perturb the result to achieve data augmentation. The landmarks form a geometric abstraction of the images but the ones obtained manually or automatically from real images are different from the simulated ones. Reducing this domain gap is critical to handle imperfections, see R2 and R3. See ablation study in Tab 2 and 3: DA consistently improves registration with Automatic landmarks, eg, by 26% in Tab 2 w/o P2.

Comment: Results don’t address the cases where the silhouette is partially visible.

Resp: Indeed, the silhouette is almost never entirely visible (liver is big and occluded). We disagree that this is not addressed: all the tested cases show a partially visible silhouette. See, eg, Fig 2 or 4, with a fraction of the silhouette (in orange).

Comment: Why using FEM for deformation simulation, ARAP or FFD could be an option? Does the DA also learns to adapt to the deformations or only for the coarse registration?

Resp: FEM uses the liver parenchyma’s mechanical parameters (all details given in Sec 4.1); in contrast, ARAP is an approximate isometric model and FFD is a smoothest deformation field from control points. See comparison in Labrunie’s PhD thesis (Dec 2024) for superiority of FEM in reproducing realistic deformations.

R3

Comment: Significance of DA

Resp: Segmentation models are improving but may never be perfect (large inter-patient liver appearance and shape variability). Hence, stress is put on registration inference, making DA relevant to cope with imperfections; with the current segmentation method, DA brings a significant improvement (resp to R1).

Comment: Unintuitive design of Fig 1 Resp: Our method estimates the rigid part first and the deformation part second (Sec 3.1), freezing the rigid part. Fig 1 shows that deformation is in liver model coordinates. When estimating the deformation, the rigid part is used, so the deformation does not try to fit the rigid part (Fig 3). We will clarify Fig 1.

Comment: It is unclear how the domain adaptation is trained

Resp: DA is trained with a supervised loss for the simulated landmarks and an adversarial loss for the real landmarks (Sec 3.2), which we named training and test in Fig 2. DA training is specific to the segmentation method.

Comment: Hausdorff distance and RTO

Resp: This is accurate: the improvement brought by RTO is currently limited. We are exploring other distances instead of Hausdorff, including L2, L1 and M-estimators.

Comment: Details of the training set are missing

Resp: We use liver models reconstructed from CT scans using MITK (Sec 4.2). Part of the dataset is public [16] and part is private. Details and IRB given in Sec 4.2. There is no specific initial liver position: the models are expressed in some CT coordinate frame.




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.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Two of the three reviewers recommend acceptance (one upgraded). The third reviewer finds the contribution incremental wrt to refs 11, although with better accuracy.



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’

    N/A



back to top