List of Papers Browse by Subject Areas Author List
Abstract
Intraoperative tumor tracking is critical for accurate targeting and safe margin control in X-ray-guided liver interventions. However, accurate tracking is difficult because fluoroscopy often provides weak or missing tumor appearance, and single-view X-ray images are insufficient to reliably constrain dynamic 3D deformation under severe structural overlap. This motivates a two-stage dynamic 3D/2D registration framework that first constructs a static Gaussian organ-target prior from preoperative CT and then deforms it under intraoperative X-ray sequence guidance. In Stage I, we build a label-aware static 3D Gaussian model with task-driven non-uniform sampling and constrained initialization, which preserves global liver morphology while improving the stability of internal tumor representation. In Stage II, we extract explicit inter-frame 2D motion cues from X-ray images and map them to 3D Gaussian deformation using a decomposed motion formulation with rigid-like sliding and residual deformation, parameterized by a sparse 3D control-point lattice. The deformation model is trained in a closed registration loop with joint 2D and 3D supervision for accurate and stable dynamic alignment. Experiments on simulated and retrospective clinical data demonstrate state-of-the-art performance in intraoperative 3D tumor tracking and deformation estimation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5208_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{GenHai_Intraoperative_MICCAI2026,
author = { Geng, Haixiao AND Fan, Jingfan AND Ai, Danni AND Song, Hong AND Zhang, Jinying AND Duan, Feng AND Yang, Jian},
title = { { Intraoperative X-Ray-Guided Tumor Tracking via Dynamic 3D/2D Registration with Task-Driven 3D Gaussian Priors } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16889},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The liver and tumor anatomy is represented using a Gaussian Mixture Model (GMM) of X-ray attenuation, initialized from the patient’s preoperative 3D CT scan. At inference time, a deformation field is estimated over the liver volume by optimizing displacements at a sparse set of control points. These displacements warp the GMM attenuation representation, which is then differentiably rendered as a digitally reconstructed radiograph (DRR) and compared against the observed intraoperative X-ray frame to drive the optimization.
Motion Estimation and Deformation Prediction: Intraoperative motion is estimated from successive X-ray frames using a dedicated 2D motion prediction network. Its output serves as initialization for a 3D deformation prediction network, which regresses control-point displacements within the liver region. The warped GMM is rendered as a DRR and matched against the observed frame, closing the optimization loop.
Training: Both networks are trained on synthetic data derived from 4DCT acquisitions. For each CT phase, DRRs are generated by forward projection, providing paired 2D–3D supervision. Training is supervised with a combined loss comprising a 2D intensity consistency term in the projection domain, a 3D anatomical label consistency term, and a spatial smoothness regularization term. Inference At test time, given an unseen preoperative CT and a sequence of intraoperative X-ray frames, registration proceeds iteratively: the 2D motion prediction network estimates displacement from the current DRR rendering and the observed X-ray; this estimate initializes the 3D deformation network, whose output warps the GMM; the updated GMM is re-rendered and fed back into the 2D network for the next iteration.
Evaluation: The method is quantitatively evaluated on real X-ray sequences on predicted versus segmented 2D tumor contours and against 3D registration baselines on synthetic data derived from 4DCT.
- 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.
Use of a GMM as a differentiable attenuation representation: The use of a Gaussian Mixture Model to encode 3D X-ray attenuation — rather than a voxel grid or mesh — is a compact and differentiable representation that facilitates both efficient DRR rendering and smooth deformation.
Deformation modelisation The paper explicitely models deformations to include a sliding component, to better respect real biomechanical behavior of the liver.
Evaluation: The evaluation is conducted on both real and synthetic data: quantitative assessment of 2D tumor contour prediction on real X-ray sequences (clinically relevant and challenging) and 3D registration accuracy on synthetic 4DCT-derived data (providing 3D ground truth, hard to obtain with real 2D X-Ray sequences). This dual evaluation strategy partially compensates for the inherent difficulty of obtaining intraoperative 3D ground truth.
Ablation study: The paper includes an ablation study that isolates the contribution of individual components. This justifies and measures the contribution of each design choice.
Domain gap mitigation: The integration of a CycleGAN-based approach to reduce the domain gap between synthetic DRRs and real intraoperative X-ray images is an important aspect of the method, even though it is not fully described in the paper. Bridging this gap is one of the central challenges in DRR-based X-Ray to CT 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.
GMM-based differentiable X-ray rendering has been explored in prior work, including Gaussian splatting approaches for X-ray novel view synthesis and volume rendering. The paper does not adequately differentiate its rendering formulation from these existing methods. Relevant references include:
Cai et al. , Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis, ECCV 2024. Gao et al. , DDGS-CT: Direction-Disentangled Gaussian Splatting for Realistic Volume Rendering, NeurIPS 2024.
The experimental comparison is limited to 3D registration baselines. The paper does not benchmark against dedicated 2D/3D deformable registration methods from the literature, which represent the most directly relevant competing approaches.
The architectures of both the 2D motion prediction network and the 3D deformation prediction network are not described in the Methods section. Given that these networks constitute the core technical contribution of the paper, their omission is a significant reproducibility and evaluation concern.
The distinction between training and inference is insufficiently described. In particular, it is unclear whether any optimization loop is performed at test time, and if so, how many iterations are used and what convergence criterion is applied.
- 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 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.
(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?
The central algorithmic contributions — the 2D and 3D motion prediction networks — are not described with sufficient detail to assess their design choices or to reproduce the results. The GMM-based DRR rendering, while a useful component, is not novel to this work and is not adequately contextualized against closely related prior art. Finally, the experimental evaluation relies solely on 3D registration baselines and does not include comparisons with dedicated 2D/3D deformable registration methods, which limits the assessment of the method’s relative merit. These three factors collectively justify a weak reject, pending clarification in the rebuttal.
- 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.
Even though the main weakness of the paper, limited validation, cited by R2 and R3, is not adressable in the rebuttal, the contributions remain relevant and worth being presented at MICCAI, as pointed out by all reviewers. The authors agreed to add required clarifications, which should increase the reproducibility of the paper.
Review #2
- Please describe the contribution of the paper
This paper proposes a two-stage framework for intraoperative liver tumor tracking from fluoroscopic X-ray sequences. In Stage I, the authors build a label-aware static 3D Gaussian representation from preoperative CT and liver/tumor masks, using non-uniform sampling to preserve global liver structure while assigning higher representational density to the tumor. In Stage II, they estimate inter-frame 2D motion cues from X-ray images and map them to 3D Gaussian deformation through a sparse control-point lattice with a decomposed motion model that separates rigid-like sliding from residual nonrigid deformation. The method is trained in a closed registration-by-rendering loop with joint 2D projection consistency, 3D label/volume supervision, and smoothness regularization. The main practical contribution is showing that this combination can improve tumor localization and deformation estimation under challenging single-view fluoroscopy, where soft-tissue contrast is weak and depth cues are limited.
- 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 targets a clinically meaningful and technically difficult problem: continuous intraoperative liver tumor tracking from single-view X-ray, where poor soft-tissue visibility and depth ambiguity make the task highly ill-posed. The proposed formulation is well aligned with interventional workflow because it leverages preoperative CT priors rather than attempting unconstrained reconstruction.
- The two-stage design is sensible and technically coherent. In particular, the label-aware non-uniform Gaussian sampling gives more capacity to the tumor region, while the motion decomposition into sliding and residual deformation is a thoughtful way to stabilize dynamic registration. The ablation study supports that both design choices matter, especially the explicit 2D motion guidance.
- The experimental section is encouraging: the method improves over several learned registration baselines on simulated data and also shows promising retrospective clinical results across a respiratory cycle, including physician-annotated tumor contour evaluation.
- 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 novelty is somewhat incremental at the component level. Differentiable X-ray rendering for 2D/3D registration has already been explored by Gopalakrishnan et al. [8], radiative Gaussian modeling for X-ray synthesis by Cai et al. [3], sparse-control deformation ideas by Li et al. [13], and single-view volumetric/tumor tracking or reconstruction by Dai et al. [5], Shen et al. [16], and Zakeri et al. [20]. The main contribution here is therefore the integration and task-specific adaptation rather than a fundamentally new formulation.
- The baseline comparison is not fully convincing for this specific task. VoxelMorph, TransMorph, MambaMorph, and DiffuseMorph are strong generic registration methods, but they are not the most directly matched baselines for single-view X-ray-guided dynamic 3D/2D registration. Comparisons against more task-relevant methods such as Nakao et al. [14], Dai et al. [5], or DSC-Recon [7] would strengthen the claims.
- The retrospective clinical validation is promising but still limited: only 7 CT/X-ray test cases are used, the real-image pipeline depends on CycleGAN-based domain translation, and runtime / robustness to initialization are not clearly reported. These issues make clinical readiness harder to assess.
- 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.
- 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.
(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?
I recommend 5 (Accept) because the paper addresses an important intraoperative imaging problem with a well-motivated and technically consistent framework. The integration of label-aware Gaussian priors, explicit 2D motion cues, and decomposed sparse-control deformation is pragmatic and appears effective. The results are strong on simulated data, the ablations are informative, and the retrospective clinical demonstrations suggest real translational potential. My enthusiasm is moderated by somewhat incremental novelty relative to prior differentiable rendering / 2D-3D registration literature, limited task-matched baselines, and a relatively small clinical evaluation. Even with those caveats, I believe the paper is above the acceptance bar and would be a valuable MICCAI contribution.
- 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.
The rebuttal clarifies several implementation details, including the network backbones and intended task-specific design, but it does not sufficiently resolve the main concerns. The novelty remains largely an integration of existing Gaussian-splatting, X-ray synthesis, and deformable registration ideas. More importantly, the baseline comparisons are not fully convincing for this task, and the retrospective clinical validation is still too limited to establish robustness or clinical readiness.
Review #3
- Please describe the contribution of the paper
The paper proposes a two-stage dynamic 3D/2D registration framework for intraoperative tumor tracking for liver interventional procedure, which leverages a label-aware 3D Gaussian prior constructed from preoperative CT and deforms it using motion cues extracted from X-ray sequences. By combining a compact Gaussian representation with motion-guided deformation and DRR-based alignment, the method enables more stable and accurate tracking of liver tumors under challenging single-view fluoroscopy conditions.
- 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.addressed a clinical relevant problem 2.introduced a label-aware 3D Gaussian representation, which allocates higher sampling density to tumor regions, enabling more stable modeling of weakly visible internal targets compared to standard voxel or mesh representations 3.Proposed two-stage formulation (static prior + motion-driven deformation) is conceptually clean and reduces the ill-posedness of 3D/2D registration by separating representation from dynamic alignment. 4.leverages explicit 2D motion cues and decomposed motion modeling (rigid-like + residual) to guide 3D deformation, improving temporal stability under respiratory motion 5.shows consistent improvements over selected baselines
- 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.dependence on preprocessing via domain translation (CycleGAN) to convert X-ray images into DRR-like images before registration. The impact of this step is not validated 2.Each compnent of the methods are not novel. However the combination of the methods are designed to fulfill the clinical workflow 3.Only 7 case clinical validation 4.limited comparison with baselines. Further comparison studies should be conducted
- 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.
- 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?
For a CAI paper the focus of the review is on the fit to real clinical problem statement and workflow. As the data for interventional cases are per se less available, 7 case validation is some how sufficient although it is still a small cohort. In total the design of the entire framework is clear and reasonable. The clinical set up is well understood and considered. The choice of the methods building blocks are well motivated and described.
- 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
Author Feedback
We thank the reviewers and AC for their constructive comments. We are encouraged that the clinical motivation, coherent framework, and promising simulated/clinical results of our work were recognized. We address the major concerns below. Q: Method details (R1). A: We agree that the backbone details should be made explicit. The input-output flow of the 2D motion and 3D deformation prediction networks is defined in Eqs. (3)-(5) and illustrated in Fig. 1.We will specify after Eqs. (3)-(4) that the 2D motion network adopts a VoxelMorph-style U-Net encoder-decoder to estimate the 2D motion field, while the 3D deformation network adopts a TransMorph-style Transformer backbone to regress the 6-DoF sliding parameters and residual displacements of the control-point lattice. During inference, the model performs feed-forward deformation prediction without test-time optimization. The last paragraph of Sec. 2.3 describes training optimization and will be renamed from “Optimization” to “Training Optimization” to avoid ambiguity. Q: Novelty and task-specific design (R1-R3). A: Our contribution is not generic Gaussian splatting, X-ray novel-view synthesis, or standalone Gaussian rendering. Gaussian rendering serves as the differentiable projection operator in our 3D/2D deformable registration task. The core contribution is a clinical-task-driven, two-stage framework for tumor tracking from single-view X-ray images. It constructs a label-aware 3D Gaussian prior from preoperative CT and liver/tumor masks with tumor-focused non-uniform sampling, improving the representation of small tumors and subtle local deformation. It then drives tumor-aware 3D deformation using intraoperative 2D motion guidance and decomposed sliding/residual deformation estimation, enabling accurate localization and deformation tracking of weakly visible tumors in X-ray-guided interventions. We will make the contribution statement more explicit in the Introduction. Q: Task-matched baseline comparisons (R1-R3). A: We agree that task-matched 3D/2D comparisons are important. For a fair comparison of single-view X-ray-guided tumor tracking via 3D/2D deformable registration, we adapted VoxelMorph, TransMorph, MambaMorph, and DiffuseMorph from generic 3D/3D registration to the 3D/2D protocol (source: CT with liver/tumor labels; target: DRR/X-ray images). All methods use the same projection geometry, differentiable rendering, and clinical preprocessing. Thus, the current baseline comparisons provide a controlled and task-adapted evaluation, supporting the superior 3D/2D deformable registration performance of our method. This adaptation protocol will be made explicit in the revision. Dai et al. [5], Nakao et al. [14], and DSC-Recon [7] are important task-specific references, but a direct comparison is non-trivial. [5] uses patient-specific training, whereas our method learns a cross-patient deformation model. [14] and [7] mainly estimate surface/mesh-based organ motion, while our task requires volumetric deformation for intrahepatic tumor tracking. These differences will be emphasized in the related-work discussion. Q: Retrospective study and CycleGAN in clinical validation (R2, R3). A: The 7-case retrospective study provides feasibility validation on real X-ray images, with physician-annotated tumor contours enabling clinically meaningful quantitative assessment, as shown in Fig. 2.Fig. 4 presents good agreement between registered tumors and physician annotations. Together with simulated evaluation using available 3D ground truth, these results support the effectiveness and clinical potential of our method. Broader clinical CT/X-ray validation will be our future focus. CycleGAN is a shared pretrained X-ray-to-DRR translation module used in clinical testing. It provides the same image-domain input for all methods and is independent of our registration mechanism. Thus, the clinical validation evaluates registration performance under the same domain-adaptation setting.
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 paper proposes a two-stage framework for intraoperative liver tumor tracking from fluoroscopic X-ray sequences, offering a clinically relevant and well-structured approach with promising results, but limited by insufficient methodological detail, incremental novelty, and incomplete baseline comparisons. Please respond and clarify these major concerns in the rebuttal. The AC recommends proceeding to rebuttal before making a final decision.
- 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.
Although the validation remains limited and some concerns about novelty and baseline comparisons persist, the framework is clearly designed, clinically well motivated, and relevant to a real interventional workflow, with reasonable methodological choices and promised clarifications that should improve reproducibility. Therefore, the AC recommends acceptance.
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.
The paper received mixed recommendations following the rebuttal. Reviewer 2 lowered their assessment from Accept to Reject, citing limited novelty and an insufficiently comprehensive evaluation. Reviewer 1 upgraded their score to Accept after the authors clarified aspects of the architecture, although this did not substantially address Reviewer 2’s broader concerns. Reviewer 3 maintained a Weak Accept recommendation, viewing the work as potentially suitable from a CAI-oriented evaluation perspective. After further consideration, I find that the experimental validation remains insufficient. In particular, the paper lacks comparisons against strong baselines specifically designed for 2D/3D registration [1], which should be included and discussed. The current comparative methods are not dedicated to the 2D/3D registration setting, making it difficult to assess the true competitiveness of the proposed approach. As a result, the evaluation is not yet rigorous enough for a CAI-focused contribution, while the level of methodological novelty also appears insufficient for a MIC-oriented paper. In light of the post-rebuttal discussion and the overall reviewer feedback, I recommend rejection.
[1] Intraoperative 2D/3D image registration via differentiable X-ray rendering （CVPR2024）
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.
I thank the authors and reviewers for this fruitful discussion that clarified most of the points raised by the reviewers. First of all, despite most components not being novel per se, the authors managed to integrate them in a new innovative way to form a novel framework design for a specific clinical task, thus clearing novelty concerns.
Another concern was the limited clinical evaluation. This field is extremely demanding in terms of manual annotations, so even if 7 subjects might seem like not much, it is actually pretty consequential. Moreover, most CT/X-ray methods are evaluated on the same Pelvic dataset, but here the authors evaluate on a new clinical dataset, which I think is a very positive point.
I also think the authors satisfyingly addressed the concerns about missing baselines, showing that the suggested competitors are not directly comparable. As such, these missing baselines do not constitue a major flaw, even if it would have been preferable to make the little adjustments necessary to include them as competitors.
Finally, the authors clarified some missing details about architecture and test-time inference process.
Overall, this is a convincing rebuttal, and I recommend acceptance for this elegant framework with clinical evaluation. However, I strongly encourage the authors to also include some points that they did not mention int heir rebuttal: missing description of the cycleGAN framework for domain adaptation, as well as runtimes, to which I personally add missing statistical tests.