List of Papers Browse by Subject Areas Author List
Abstract
Accurate CBCT–IOS registration is essential for dental diagnosis and implant planning, yet the modality gap between CBCT volumetric intensity and IOS surface geometry substantially increases correspondence uncertainty. Conventional segmentation-based pipelines reduce the modality gap by converting CBCT into a surface-like representation, but their accuracy is highly sensitive to unstable segmentation and does not explicitly enforce reliable correspondences. We reformulate CBCT-IOS registration as direct keypoint correspondence estimation in feature space, and propose a correspondence-centric end-to-end framework. Our method learns modality-aligned embeddings, selects registration-relevant keypoints by favoring points with peaked cross-modal matching distributions (i.e., unambiguous matches) while regularizing intra-modality distinctiveness, and predicts correspondences jointly with their validity. To further suppress residual near-tie mismatches in repetitive dental geometry, we refine the rigid transform by iterative SVD while annealing the matching criterion from feature similarity to geometric consistency. Experiments on a clinical CBCT–IOS dataset demonstrate that our approach consistently outperforms previous segmentation-based approaches.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5399_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{ParJeo_Keypoint_MICCAI2026,
author = { Park, Jeonglok AND Jo, SeungKwan AND An, Jonghoon AND Chung, Minyoung},
title = { { Keypoint Correspondence Learning for CBCT–IOS Registration } },
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
This paper presents a registration framework for aligning 3D CBCT scans with IOS. The method directly matches keypoints in a shared feature space, removing the need to extract skull surface meshes. It uses intra-modality saliency and cross-modal matching confidence to establish correspondences. The approach applies iterative SVD-based refinement with a gradually tightened matching criterion.
- 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 reformulates CBCT–IOS registration as direct keypoint matching in a learned feature space, avoiding the need for surface mesh extraction. This design supports joint optimization of feature alignment, keypoint selection, and correspondence estimation. The model estimates rigid transformations using SVD and progressively shifts the matching criterion from feature similarity to geometric consistency. Experimental results show clear improvements over segmentation-based 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.
- The training process relies on ground-truth transformations, but the paper does not explain how these transformations or the corresponding keypoint matches are defined.
- The framework combines cross-attention for fusing point sets with volumetric features and uses iterative refinement. The computational and memory costs of these components are not discussed.
- The evaluation is on 92 CBCT–IOS pairs from OSSTEM IMPLANT. The model’s ability to generalize to other datasets is not addressed. In addition, CBCT scans often contain metal artifacts that degrade image quality, it would be helpful to explain how the method handles correspondence estimation in these regions.
- 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?
This work addresses CBCT–IOS registration without relying on bone or teeth segmentation. The use of shared feature embeddings, entropy-based keypoint selection, and iterative refinement appears effective and leads to improved performance in dental registration tasks. Further clarification of the training setup, along with discussion of computational cost and generalization, would strengthen the paper.
- 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 #2
- Please describe the contribution of the paper
The paper presents a novel CBCT–IOS registration framework that addresses orthodontic cross-modal registration without requiring modality conversion. The method combines modality-specific feature extractors, carefully designed objectives at different stages, and appropriate pre- and post-processing to achieve highly accurate registration. In particular, the proposed keypoint matching strategy is well motivated, including the entropy of the cross-modal matching distribution and an intra-modality saliency term for selecting registration-relevant keypoints. The final visualized keypoints are also interpretable, which strengthens the practical value of the approach. Quantitative comparisons and ablation studies further support the effectiveness of the proposed framework.
- 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 very well written. The problem is clearly motivated, the method is described in sufficient detail, and the experimental and ablation results are comprehensive and convincing. 2.The proposed framework is novel. It addresses registration between two substantially different modalities, namely CBCT volumes and IOS point clouds, without relying on explicit modality conversion to a shared representation. This is a meaningful conceptual advance over conventional pipelines. 3.The proposed loss design appears effective, especially the keypoint extraction objective combining cross-modal entropy and intra-modality saliency. The inference-time SVD refinement is also well justified and is shown to further improve the final registration performance. 4.The method provides interpretable correspondence visualizations, which is valuable for clinical trust and practical deployment.
- 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 loss weights are not given. 2.The paper discusses segmentation-based registration pipelines, including ICP-style alignment strategies, but does not provide a direct comparison with simpler traditional registration baselines. Such a comparison would help further contextualize the empirical gains of the proposed 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 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
I would like to see more details about the training, for example, on which devices and the specific hyperparameters.
- 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 paper addresses an important and practically relevant registration problem and proposes a clear, technically strong, and well-validated solution. I find the core idea novel, especially the correspondence-centric formulation that avoids explicit modality conversion, and the proposed loss design and refinement strategy appear both effective and well motivated. The paper is also very well written, with strong experimental and ablation support. My only minor reservations are the lack of loss weights and a direct comparison with simpler traditional registration baselines, which would further strengthen the evaluation. Overall, the contribution is strong, and I recommend acceptance.
- 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 #3
- Please describe the contribution of the paper
The paper introduces an end-to-end framework for CBCT–IOS registration that bypasses the conventional segmentation-based paradigm. Rather than converting CBCT volumetric data into a surface representation to reduce the modality gap, the method learns modality-aligned feature embeddings directly through cross-attention fusion, selects registration-relevant key points using a combination of intra-modality saliency and cross-modal confidence, and predicts correspondences jointly with their validity. The rigid transform is then estimated via iterative SVD with an annealing strategy. The resulting method becomes a geometric-consistency based approach, making it robust to metal artifacts that are often found in CBCT volumes. The framework is trained end-to-end under a single joint objective on 64 CBCT–IOS pairs and evaluated on 28 test pairs, reporting a mean TRE of 0.30 ± 1.90 mm — nominally outperforming three segmentation-based baselines.
- 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 reframing of CBCT–IOS registration as direct cross-modal keypoint correspondence estimation in feature space is the paper’s most significant contribution. Unlike segmentation-based pipelines that reduce the modality gap by converting CBCT into a surface representation, the proposed method learns to bridge the modality gap directly in a shared feature space. This reformulation is principled, well-motivated, and an interesting methodologically distinct direction for multimodal dental registration research. A combined keypoint selection criterion — intra-modality saliency with inter-modality confidence derived from the sharpness of the matching distribution — is an interesting design choice, and the ablation study provides evidence that the combination is necessary. Training all pipeline components under a single joint objective is an important practical and methodological strength, as it avoids the error accumulation and suboptimal component interactions that can arise from staged training pipelines. The monotonic ablation results provide support for the effectiveness of this training strategy. The qualitative results show stable registration under heavy metal artifact conditions — where segmentation-based methods demonstrably fail —which is a clinically meaningful demonstration.
- 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 evaluation dataset of 46 patients (92 CBCT–IOS pairs) is critically small, and patient inclusion or exclusion criteria are missing. It is unclear how patients were selected, whether the train/test split was performed at the patient level, and whether the dataset captures clinically relevant variability. No demographic information, scanner specifications, or acquisition protocols are reported. The absence of IRB approval documentation and patient consent statements for a commercially sourced clinical dataset is a significant ethical concern that must be explicitly addressed. The paper makes several implicit assumptions about the dataset that should be stated. Would the rigid transform assumption be violated in any cases, or will the assumption of anatomical overlap always hold? A limitations section is recommended. It would also be ideal to have runtime figures, hardware specifications, and inference time comparisons against baselines. The pipeline involves multiple computationally demanding components, all operating on high-dimensional volumetric and point cloud data. Without this information, so the practical feasibility of clinical deployment should be evaluated. The manuscript contains spelling errors, inconsistent capitalization, and typographical mistakes throughout that will need to be corrected prior to publication.
- 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 weakness section outlines several issues that should be addressed to make the paper acceptable to the main conference.
- Reviewer confidence
Very confident (4)
- [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 the meta-reviewer for their constructive comments and for recognizing our main contribution: direct cross-modal keypoint correspondence learning for CBCT–IOS registration without segmentation-based modality conversion. We will revise the camera-ready manuscript to improve reporting of data, training, computational cost, baselines, and limitations.
The train/test split was performed at the patient level; maxillary and mandibular pairs from the same patient were not shared across training and test sets. Although the dataset is single-source, it contains clinically meaningful variability, including full-arch scans across a wide age range, metal-artifact cases, and cases with reduced remaining teeth. The dataset was provided through an industry collaboration after de-identification, and the authors did not have access to personal identifiers. We will clarify the de-identification/data-use process and include the applicable ethics/data-use statement where available.
Regarding the training setup, ground-truth keypoint correspondences were not manually annotated. They were derived from the ground-truth rigid transformation: after applying the transformation, a candidate CBCT–IOS keypoint pair was treated as a positive correspondence if its Euclidean distance was below the predefined inlier threshold, and otherwise treated as a non-corresponding/outlier candidate. We will explicitly add this definition to the method section. We will also add the missing implementation details, including loss weights, optimizer, learning rate, batch size. We agree that computational cost is useful for reproducibility and practical reference; therefore, we will report the hardware environment and inference time in the implementation details, while keeping the main evaluation focused on registration accuracy.
Regarding metal artifacts, metallic structures appear in CBCT but do not have corresponding structures in IOS. Therefore, artifact-dominated CBCT regions tend to produce low cross-modal matching confidence. Our entropy-based confidence term and inlier/correspondence supervision encourage the model to select keypoints that are both salient and consistently matchable across modalities, reducing reliance on CBCT-only artifact regions. We will make this explanation clearer in the revised manuscript.
Regarding the baseline comparison, it is clarified that the baseline of this study includes the existing ICP/FPFH method registration following the standard segmentation/surface transformation steps. Directly applying existing registration methods to the original CBCT volume and IOS point cloud is not effective because the geometric representations of the two input data differ. For this very reason, the existing CBCT-IOS pipeline performs segmentation or surface extraction before applying geometric matching.
Finally, we agree that the rigid-transform and anatomical-overlap assumptions should be explicitly stated. The rigid-transform assumption is standard in CBCT–IOS dental registration when the target structures are teeth or hard-tissue-related dental surfaces. Our method assumes sufficient overlapping dental anatomy between CBCT and IOS to define correspondences; completely non-overlapping scans are outside the scope of rigid correspondence-based registration. We will add a limitations paragraph clarifying that this assumption may be weakened by large acquisition time gaps, treatment-related anatomical changes, tooth movement, severe IOS stitching distortion, or insufficient overlap. We will also carefully proofread the manuscript and correct typographical/table notation errors, including the RRE direction indicator, which should be ↓ because lower rotation error indicates better registration accuracy. The reported numerical results remain unchanged.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
This paper presents a novel approach for CBCT–IOS registration. The main strengths is that the paper presents a clear and well-motivated reformulation of cross-modal dental registration as direct keypoint correspondence learning in a shared feature space, avoiding explicit segmentation or modality conversion. Reviewers also appreciate the empirical results showing improved accuracy over segmentation-based baselines, which I agree based on my own reading. The work is also praised for being well written and experimentally reasonably well validated. There are also some concerns, including limited experimental and reproducibility details, incomplete information on loss weights, hyperparameters, training setup, and computational cost, as well as the relatively small and single-source clinical dataset without strong discussion of generalization, patient selection, or dataset variability. One reviewer also notes the lack of comparison to simpler classical registration baselines (e.g., ICP-style methods), and another raises concerns about dataset transparency and ethical documentation. The reviewer recognize the novelty of the work, while the limitations are addressable. Overall, the paper receives a mixed-but-positive recommendation profile with a consensus leaning toward acceptance. I therefore recommend this paper to be accepted to MICCAI, though the authors shall address the mentioned limitations when preparing their final papers (clarifications on experimental details, additional baselines, and improved reporting of data and training protocols).
