List of Papers Browse by Subject Areas Author List
Abstract
Precise 3D reconstruction of alveolar bone defects using Cone-Beam Computed Tomography (CBCT) is essential for predictable dental implant surgery and bone graft planning. However, automated restoration remains challenging due to the inherent anatomical complexity and the scarcity of paired clinical datasets. In this paper, we propose Diff-SABR, a novel diffusion-based symmetry-aware framework to reconstruct alveolar bone morphology. Our approach employs a hybrid learning strategy that integrates a synthetic dataset generated through hyperbolic-paraboloid simulation with a clinical dataset for weakly supervised training. To handle anatomical variations, we incorporate AlignNet, a spatial alignment module that registers healthy mirrored anatomy to the defect site. Experimental results demonstrate that our framework achieves high anatomical fidelity, with a Dice Similarity Coefficient (DSC) of 0.82 and a 95th percentile Hausdorff Distance (HD95) of 1.67 mm on real-world clinical scans. Quantitative and qualitative ablation studies further confirm that the synergy between hybrid supervision and spatial alignment is crucial to restore clinically viable ridge geometries. By providing objective and automated 3D bone restoration, Diff-SABR offers a robust tool to improve surgical stability and optimize bone graft material selection in clinical practice.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1226_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{ChoHo_DiffSABR_MICCAI2026,
author = { Choi, Ho Yoon AND Han, Ji Yong AND Lim, Sang-Heon AND Kim, Sujeong AND Yun, Hyunbin AND Ahmadi, Mobin AND Oh, Sungho AND Lee, Taeyeon AND Han, Jeong Joon AND Yi, Won-Jin},
title = { { Diff-SABR: Diffusion-Based Symmetry-Aware Reconstruction of Alveolar Bone Defects from CBCT } },
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 proposes Diff-SABR, a novel symmetry-aware latent diffusion framework for 3D reconstruction of alveolar bone defects from CBCT scans. The main contributions lie in three aspects. First, a hyperbolic-paraboloid-driven defect simulation strategy mathematically replicates asymmetric bone resorption patterns, generating paired synthetic data to address clinical data scarcity. Second, a spatial alignment module, AlignNet, registers mirrored healthy contralateral anatomy to the defect site, enforcing global morphological consistency during the diffusion process. Third, a hybrid training framework combines full supervision on the synthetic dataset with a weakly supervised learning objective on clinical data, effectively bridging the synthetic-to-clinical domain gap. Experiments on the clinical dataset demonstrate that the proposed method achieves promising reconstruction performance.
- 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 introduces a mathematical model of bone resorption using a hyperbolic paraboloid, which enables realistic simulation of complex, asymmetric multi-site bone defects and provides a principled alternative to simple masking techniques. The method effectively incorporates anatomical priors by leveraging the jawbone’s natural bilateral symmetry. AlignNet is used to correct pose-induced misalignments before applying the symmetry prior in the diffusion model, ensuring structurally consistent reconstructions. Ablation studies demonstrate the contribution of each component, showing that combining weak supervision with alignment improves the clinical DSC from 0.04 ± 0.06 to 0.82 ± 0.04, highlighting the effectiveness of the proposed strategy.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
In Figure 2(b), the input to AlignNet is the mask (fixed mask and moving mask), whereas in Figure 2(c) within the Diff-SABR module, AlignNet’s input becomes the defect CBCT and the symmetry CBCT. This inconsistency raises a concern: Was AlignNet trained on masks? If so, directly applying it to CBCT volumes may affect its performance or reliability. Please clarify this mismatch. The clinical dataset comprises 43 cases, with merely 10 cases reserved for testing. This relatively small sample size may not fully capture the extensive structural variability of alveolar bone defects present across the broader patient population. The approach relies on aligning the defect region with contralateral healthy bone, but the manuscript does not discuss how midline-crossing or bilateral defects are handled. The authors should clarify the applicability and limitations of their method in cases with midline-crossing, bilateral, or highly asymmetric defects. The method relies on aligning the defect with the contralateral healthy bone, but the paper does not explain how midline-crossing or bilateral defects are managed. In such cases, a corresponding contralateral patch may be unavailable, which could reduce alignment accuracy and compromise defect reconstruction. The proposed method employs a 3D latent diffusion model, which is potentially memory- and computation-intensive. The paper does not report GPU memory usage or inference time, leaving its practicality and feasibility in clinical scenarios unclear.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
The paper receives a “Weak Accept” because it presents some interesting contributions, including a novel defect simulation strategy and a symmetry-aware diffusion framework. These aspects demonstrate technical innovation. However, there are several limitations that temper the overall impact. The reliance on contralateral symmetry restricts the method’s applicability to unilateral defects and may fail for bilateral or midline-crossing cases. The experimental comparison includes only older baseline methods, so it is unclear how the proposed approach performs relative to more recent techniques. Furthermore, the clinical test set consists of only 10 cases, which is too small to convincingly demonstrate generalizability across the broader patient population. Taken together, while the work is technically sound and partially novel, these concerns prevent a stronger recommendation.
- 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
Review #2
- Please describe the contribution of the paper
The authors have taken a very niche area of research focusing on the alveolar defects and its reconstruction methodology with their solution DIff-SABR- a symmetry aware latent diffusion framework using CBCT data. They appear to have addressed the problem of data scarcity for this solution by using the Hyperbolic-paraboloid driven defect simulation that generated geometrically stable synthetic data. Although there is always asymmetry in Bilateral symmetry of alveolar ridges, they utilized the AlignNet- a symmetry aware conditional latent diffusion model. The use of hybrid weakly supervised training combining supervised for synthetic labels and Huber -Penalized pseudo labels on unannotated clinical data seems promising for bridging the synthetic- clinical data gap trying to eliminate the annotations protocols.
- 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 hyperbolic-paraboloid formulation for defect synthesis is a major shift from naive masking or random cropping. By parameterizing saddle-shaped surfaces with stochastic geometric parameters (curvature, offset, depth, translation), the simulation generates diverse and anatomically consistent resorption patterns that align with the clinical morphology of alveolar ridge defects. PCA-estimated ridge centerlines used to anchor the local coordinate system grounds the simulation in dental ridge anatomy.
2.AlignNet module addresses raw contralateral mirroring which could introduce pose-induced misalignment that would otherwise mislead the conditioning signal giving a impactful rise of DSC point gain from 0.72-0.82 3.Hyperparameter transparency is provided with sensitivity analysis focussing on non-monotonic clinical performance curve revealing that Huber weighted weak loss must be balanced and reduction of over-regularization.
- 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 primary clinical evaluation is performed on only 10 cases which is very small for statistical reliability 2.Three baselines used namely 3D U-Net, WGAN-GP, and VQ-diffusion are relatively weak and not fully comparable as they are training on identical data regimes via hybrid supervision. The extent of leveraging the AlignNet conditioning isn’t clear with all the baselines 3.More algorithms like nnU-Net could be a strong supervised baseline trained with the same symmetry conditioning. 4.Technical requirements could include Inference time 5.Quality of the soft pseudo labels are not characterized or documented wrt generation, quality, sensitivity, Issues pertaining to systematic errors would propagate into the subsequent training. 6.Alveolar defects are caused not just by trauma, tooth removal but also by the gum diseases- periodontitis and are labelled as wall defects based on the remaining cortical wall which is not discussed. 7.in cases of bilateral defects, severe asymmetry or prior surgical remodeling on the contra side, the contralateral reference is not valid. 8.Ethical clearance and dataset source for clinical dataset involving patients with surgical procedures. 8.Architectural detail in Figure 2 needs to be legit and readable and better moved to a supplementary document in exchange for larger, more legible clinical figures. 9.Highlight the clinical images for easy conviction and understanding of the research.
- 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?
43 clinical CBCT scans from patients undergoing bone grafting and implant surgery. No IRB approval statement, patient consent declaration, or data governance information is provided anywhere in the manuscript. While this may simply reflect an anonymization convention for blind review, the omission is notable given the clinical nature of the data. The concern should be flagged to the AC with a request for clarification from the authors in rebuttal.
- 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
Including a clinical perspective of translation of this methodology and clinical benefits, application and limitations
- 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?
Diff-SABR addresses a clinically meaningful and technically non-trivial problem with a coherent and well-motivated framework. The hyperbolic-paraboloid simulation, AlignNet-based symmetry conditioning, and hybrid supervision are each principled contributions, and their synergy is convincingly demonstrated via ablation. The clinical results (DSC 0.82, HD95 1.67 mm) are promising.
The primary concerns revolves around small clinical test set, weak baselines, missing inference time analysis, and underspecified pseudo-label generation which are addressable. The research is above acceptance threshold given the strength of its ablation evidence and the novelty of the clinical problem formulation but score is conditional on the authors clarification
- 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.
There is clarity on the diffusion framework incorporates anatomical symmetry to reconstruct alveolar bone defects were thorough and mathematically sound.
Review #3
- Please describe the contribution of the paper
The paper proposes a diffusion-based generative framework for alveolar bone reconstruction. The major contributions are the introduction of a simulation pipeline for data synthesis and a symmetry-aware conditional latent diffusion model. The experiments demonstrate the superiority of the proposed framework in terms of surface reconstruction.
- 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 develops a data simulation pipeline aligned with the clinical background. For instance, the cropped shapes (hyperbolic paraboloid) and sizes (6-9 mm) are considered to be clinically plausible. 2) The proposed generative reconstruction framework is overall reasonably intuitive, as the framework is essentially a conditional latent diffusion model learning the distributions of geometries conditioned on imaging and masking information. 3) The experiments are detailed and relatively comprehensive. There are both qualitative and quantitative experimental results. Ablation studies and sensitivity analysis are also presented to support the design of the framework.
- 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) It is still quite unclear about the contribution of the simulation pipeline. The paper does not compare it with existing simulation methods, such as random cropping. As the simulation part is one of the major contributions of the paper, this unclarity weakens the solidity of the method. 2) The method part regarding the reconstruction framework is unclear and difficult to understand. For instance, there are four types of masks in Fig. 2, namely Hard Mask, Soft Mask, Fixed and Moving mask, and ROI Cuboid Mask CBCT. These masks lack clear definitions and are extremely confusing, as they all have masks in their names but look very different. There are no explanations for the clinical soft pseudo label. It is unclear where it comes from and what it is. Besides, this quantity is termed as mask in Fig. 2 but as label in the text, which makes it even more confusing. Another problem is the subscripts, where s is used for the clinical soft pseudo label on page 4 and the simulated dataset on page 5, and the subscript to denote the clinical dataset on page 6 is c. Why is the subscript for the clinical soft pseudo label s instead of c under such semantics? To summarize, the incomplete and confusing definitions and descriptions strongly hinder an accurate understanding of the methodology. 3) The explanation for the introduction of AlignNet is insufficient, as the paper only mentions the misalignment induced by anatomical asymmetry on page 7.A more detailed explanation of the originality and the consequences of such misalignment can better justify the motivation to introduce AlignNet. 4) The way to encode the ROI Cuboid Mask CBCT as a condition through the Mask 3D Encoder is questionable. The Mask 3D Encoder is previously trained to compress the soft and hard masks, so it is questionable whether it could encode the ROI Cuboid Mask CBCT. It might be somewhat reasonable if the ROI Cuboid Mask CBCT is a binary or a soft mask, but it has CBCT in its name. Similar to weakness 2, these masks are extremely confusing. Besides, even if the ROI Cuboid Mask CBCT is a binary or a soft mask, its distribution is significantly different from that of the masks encoded by the Mask VAE, making this usage less convincing. 5) The paper utilizes concatenation, FiLM, and cross attention for conditioning simultaneously, but does not give any justification. Normally, a latent diffusion model will not employ all these approaches at the same time for conditioning. The necessity for doing so remains questionable.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The 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.
(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 proposed framework is fundamentally reasonable and achieves a decent performance. However, the confusing methodology descriptions and the lack of justification for some designs decrease the fidelity of the approach.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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 authors have tried to address the concerns. However, the rebuttal clearly reflects several flaws of the paper. For instance, the very important acquisition procedure for clinical pseudo labels, which involves signed distance functions, was never mentioned in the paper and is finally provided in the rebuttal. Clear definitions and details of the methodology and the quantities used in the paper remain insufficient. The authors repeat the motivation of the simulation strategy and complex conditionings, but there are no experiments to prove them. Although the proposed method has merits, the way it is represented in the manuscript format still has observable flaws. Therefore, the decision is maintained as a rejection.
Author Feedback
We thank reviewers for their constructive comments. [AlignNet & symmetry guidance – R1,R4] We clarify that AlignNet is trained and applied on thresholded bone-only CBCT representations, not raw intensity volumes. In Fig. 2, the “fixed” and “moving” masks denote binarized bone maps extracted from the defect-side and mirrored contralateral CBCT volumes for affine alignment. The estimated affine transform is then applied to the mirrored contralateral CBCT volume before it is used as denoising guidance. Thus, AlignNet does not perform deformable registration or anatomical replacement. It only reduces pose- and morphology-induced mismatch so that the contralateral anatomy provides stable, patient-specific symmetry guidance. [Limited Clinical Data & Simulation Strategy – R1,R2,R4] We acknowledge that paired CBCT scans with ground-truth alveolar reconstruction are inherently difficult to obtain. Our evaluation included maxillary, mandibular, anterior, and posterior defects to cover diverse anatomical locations. This clinical constraint motivated our anatomically informed simulation and weakly supervised framework. The proposed hyperbolic-paraboloid simulation reflects clinically observed asymmetric ridge resorption patterns, including bucco-lingual collapse and vertical loss, rather than arbitrary random masking. Unlike random masking, our simulation preserves clinically plausible morphology and structural continuity, helping reduce the synthetic-to-clinical gap. [Mask definitions & framework clarity – R4] We clarify the mask terminology. The hard mask Mh denotes the simulation-derived ground-truth defect label, whereas the soft mask Ms denotes the clinical pseudo label. The fixed/moving masks in AlignNet are thresholded bone-only CBCT maps used only for affine registration. The ROI cuboid mask Mr is a coarse localization prior for the candidate defect region. The Mask-VAE is used to encode spatial mask representations, allowing Mr to provide structural guidance during denoising. We also acknowledge that the notation in the manuscript could have been clearer, particularly the distinction between masks used for registration, supervision, and localization. [Clinical Pseudo Label Generation – R2,R4] Clinical pseudo labels are generated by measuring the discrepancy between defect-side anatomy and the aligned contralateral reference after symmetry registration. They are represented as signed distance maps, yielding continuous soft masks rather than discrete binary masks. Since dense voxel-wise annotations are unavailable, these labels are used as weak anatomical guidance rather than exact ground truth. [Conditioning design – R4] The three conditioning paths are complementary. Concatenation injects spatial priors, FiLM modulates intermediate features according to anatomical context, and cross-attention models structural correspondence between reconstruction latents and the aligned contralateral anatomy. This design is intended to combine local defect localization with global symmetry guidance. [Baseline Comparison – R1,R2] We compared with 3D U-Net, WGAN-GP, and VQ-based diffusion as representative supervised and generative baselines under the same training protocol. WGAN- and VQ-based models remain relevant references for medical image synthesis and reconstruction. We used 3D U-Net as a controlled supervised baseline for architectural consistency. [Applicability & limitations – R1,R2] Diff-SABR is primarily designed for unilateral defects where reliable contralateral anatomy is available and can serve as patient-specific symmetry guidance. The more complex defect patterns require mechanisms beyond direct symmetry guidance and are an important direction for future work. [Inference Time – R1,R2] Our latent-mask diffusion framework required 1.99 seconds per case with 3.6 GB GPU memory usage using DDIM-50 sampling on a single NVIDIA RTX A6000 GPU. [Ethical Approval & IRB – R2] IRB details were omitted only to preserve review anonymity.
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 has strong novelty and clinical relevance with regard to the symmetry-aware diffusion framework, but the paper requires clarification before acceptance. Please respond to the small clinical evaluation set, limited/older baselines, and lack of comparison to simpler simulation strategies, unclear or confusing methodology details around the different masks, AlignNet inputs, pseudo-label generation, and conditioning design, as well as the limitations of relying on contralateral symmetry for bilateral or midline-crossing defects. The rebuttal should also clarify inference time/computational cost and provide the missing ethics/IRB/consent information for the clinical CBCT dataset.
- 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’s core contribution is sufficiently strong for acceptance. The work addresses an important and clinically meaningful reconstruction problem and proposes a symmetry-aware latent diffusion framework that leverages contralateral anatomy for alveolar bone defect reconstruction. The framework is mathematically sound, and the rebuttal clarified the diffusion model’s use of anatomical symmetry. The remaining concerns about manuscript clarity, especially the need to better define the clinical pseudo-label generation procedure, mask terminology, and conditioning strategy, should be addressed in the final version.
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.
While the paper addresses an important clinical problem and proposes a potentially interesting diffusion framework, the methodological presentation remains insufficiently clear, with key components such as the mask definitions, pseudo-label generation, AlignNet inputs, and conditioning mechanisms either inconsistently described or inadequately justified. Although the text should be self-contained, providing an anonymous link to code (the submitted paper only has an obscured link pointing to the code) could help disambiguate critical elements.
The rebuttal did not resolve the central reviewers’ concerns, and the claimed benefits of the simulation strategy and complex conditioning design were not supported by direct experimental comparisons.
Given the very small clinical test set, limited baseline comparisons, and unresolved reporting issues around clinical data use and practical deployment, the paper does not yet provide a sufficiently rigorous or reproducible basis for acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The post-rebuttal opinions remain mixed. Reviewer #3 recommends acceptance, finding the clarified symmetry-aware diffusion framework mathematically sound and clinically meaningful, while Reviewer #4 maintains rejection due to unresolved concerns about incomplete methodological definitions, missing pseudo-label generation details in the original manuscript, and insufficient experimental justification for the simulation and conditioning designs. Reviewer #1’s initial review was weakly positive, acknowledging the novelty of the defect simulation, AlignNet-based symmetry prior, and hybrid supervision, but also raising concerns about the very small clinical test set, limited baselines, and restricted applicability to unilateral defects.
Overall, the final version must clearly address the remaining limitations regarding methodology clarity, pseudo-label generation, conditioning design, small-scale clinical validation, contralateral-symmetry assumptions, computational cost, and ethics/IRB information. The authors should provide the full open-source implementation to improve the reproducibility.