Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Accurate segmentation of dental plaque from intraoral scan (IOS) is crucial for assessing oral hygiene practices. However, the scarcity of 3D dental plaque datasets and the limitations of relying solely on either 3D geometric or 2D appearance information present significant obstacles. To address this, we introduce DentalPSAM, a hybrid 2D-3D model for 3D dental plaque segmentation. DentalPSAM features an interleaved dual-branch architecture with an appearance-enhanced 3D mesh branch and a geometry-enhanced 2D SAM branch, built upon the zero-shot learning capabilities of the Segment Anything Model (SAM). These branches process 3D IOS and its 2D rendered images in parallel, integrating both geometric and appearance information. To achieve better alignment between 2D and 3D features, we employ model stitching to connect SAM encoder with both branches through lightweight adaptation. Moreover, we present the first open-source 3D intraoral scan dataset for dental plaque segmentation. The code is available at https://github.com/HKU-HealthAI/DentalPSAM.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4344_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/HKU-HealthAI/DentalPSAM

Link to the Dataset(s)

N/A

BibTex

@InProceedings{JiaHao_DentalPSAM_MICCAI2026,
        author = { Jiang, Haoning AND Fok, Rachel Melissa AND Zhang, Xiangcheng AND Sun, Ruitong AND YEUNG, Jonathan Hok Kan AND Deng, Ke AND Qu, Liangqiong},
        title = { { DentalPSAM: Lifting SAM to Dental Plaque Segmentation with Hybrid 2D-3D Knowledge Fusion } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16883},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses dental plaque segmentation on 3D intraoral scans and proposes a hybrid 2D–3D framework, DentalPSAM. The core idea is to jointly exploit appearance information from 2D rendered images and geometric information from 3D meshes. The authors construct the first open-source 3D IOS dental plaque segmentation dataset, PlaqueIOS, containing 400 scans from 200 patients. The quantitative experiment shows the effectiveness of DentalPSAM.

  • 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 tackles a clinically meaningful and relatively underexplored task: 3D dental plaque segmentation from intraoral scans. 2.The proposed method has a 2D–3D design. Instead of relying only on geometry or only on rendered images, the model combines 2D appearance information and 3D geometric information. 3.The paper provides a new annotated dataset, PlaqueIOS, with 400 scans from 200 patients, which is potentially useful for this task since the lack of 3D dental plaque datasets is one of the main challenges in this area

  • 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 paper would benefit from a clearer explanation of what dental plaque looks like in this setting, how the plaque labels are obtained, how the dataset is constructed, and what the exact segmentation target is. 2.For a segmentation paper, it is difficult to judge the effectiveness of the method without visual examples. Qualitative comparisons and visualization of predicted masks would provide much stronger support for the claimed improvements. 3.The paper does not clearly explain what makes this task intrinsically challenging. It is unclear whether plaque regions are mainly distinguishable by color differences, or whether additional factors play a major role. If color contrast is a dominant cue, the authors should justify why simpler color-based baselines such as thresholding or rule-based segmentation would not already work well。 4.How does the SAM decoder receive the implicit prompts generated by BMT, and at which specific stage these prompts are injected.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (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?

    Although it addresses a meaningful task and presents a reasonable 2D–3D framework, the current submission still has several weaknesses that limit its overall impact. In particular, the problem setup and dataset construction are not described clearly enough, and the paper lacks qualitative visualizations to support the claimed effectiveness. In addition, the paper does not sufficiently clarify what makes the task challenging in practice, especially whether simpler color-based baselines could already be competitive. Overall, the idea is interesting, but the current evidence and presentation are not yet strong enough for acceptance

  • 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.

    After reading the author’s feedback, considering the restrictions on adding images in the camera-ready version, the current version cannot well address my concern.



Review #2

  • Please describe the contribution of the paper

    The main contributions of this paper are twofold. First, they introduce PlaqueIOS, a open-source 3D IOS dental plaque dataset. Second, it proposes a dental plaque segmentation network that combines 2D texture information with 3D geometric information. The method outperforms representative pure 2D and pure 3D approaches, demonstrating promising potential for practical application.

  • 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.

    A major strength of this paper is that the problem formulation is clear and the technical design is well motivated. It leverages the strong 2D SAM model to introduce 2D appearance information and combines it with 3D geometric information for dental plaque segmentation. The comparison experiments and ablation studies are well designed, demonstrating the performance of the model and the effectiveness of each design component.

  • 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 main weakness is that the paper includes many design components and is relatively complex, but the presentation is not sufficiently clear within the limited space. In particular, several modules in Fig. 1 are not clearly explained. For example, it is unclear how the geometric tokens are obtained, what the yellow blocks above them represent, and although the dashed lines denote the bidirectional stitching adapter, the specific meaning of the different dashed lines is not clearly explained.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper addresses an important and practical problem with a well-motivated technical design. A major strength is that it combines 2D appearance information from a strong SAM model with 3D geometric information for dental plaque segmentation, and the experimental results provide solid support for its effectiveness. The introduction of the open-source PlaqueIOS dataset also adds value to the community. My main concern is that the method contains many components and is not always explained clearly enough. In particular, several elements in Fig. 1 are difficult to follow.

  • 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 thank the authors for the detailed rebuttal. The rebuttal provides clearer explanations of these components, including how geometry tokens are obtained, how geometric prompts are injected into the SAM decoder, and how the 2D/3D branches are fused. The rebuttal provides clearer explanations of these components, including how geometry tokens are obtained, how geometric prompts are injected into the SAM decoder, and how the 2D/3D branches are fused. While these details should be clearly added to the final manuscript, I believe the major concerns have been sufficiently addressed. I therefore support acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper addresses the critical clinical task of 3D dental plaque segmentation from IOS data. Key contributions include: (1) The authors propose DentalPSAM, a novel hybrid 2D-3D dual-branch architecture that fuses geometric information from 3D meshes and appearance information from 2D rendered images. The model leverages the zero-shot segmentation capability of SAM while enhancing it with 3D geometric priors. (2) The first open-source 3D IOS dataset (PlaqueIOS) for dental plaque segmentation with 400 expert-annotated scans.

  • 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.Clinical significance: Dental plaque segmentation addresses a widespread global health issue, with clear translational value for clinical dentistry and oral healthcare.

    1. Technical innovation: The BMT for implicit geometric prompting and lightweight spatial alignment decoder are well-designed and technically sound.
    2. Dataset contribution: The release of PlaqueIOS, the first public 3D IOS dataset for plaque segmentation, provides a valuable benchmark for future research in this domain.
  • 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.Need explanation on what dental plaque is, preferably with an image 2.Lack of visual comparisons of cases 3.UV unwrapping can result in uneven vertex distribution and localized high density, leading to distortion in the 2D image. How is this addressed? 4.How to combine back-projected 2D and direct 3D predictions? It looks like a union-intersection in Fig.1 ?

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The first open-source 3D IOS dataset (PlaqueIOS) for dental plaque segmentation.

  • 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.

    Accept



Author Feedback

We thank the reviewers and ACs for the valuable feedback.

  • Dental plaque/Dataset annotation/Visual comparison (R1, R3). Dental plaque is a microbial biofilm attached to tooth surfaces. In its natural state, it is usually colorless and therefore difficult to detect visually. In clinical practice, a plaque disclosing agent is applied before scanning. After disclosure, plaque appears as stained pink regions on the 3D intraoral scan mesh. Our task is to assign a binary plaque or non-plaque label to each vertex/face on the 3D tooth surface, as shown by the IOS input and final output in Fig. 1. As reported in Sec. 4.1, each scan was labeled in MeshLab by two clinicians with more than 5 years of clinical experience. They first annotated the 3D mesh independently. The annotations were accepted only when their overlap reached an IoU of at least 0.85.Otherwise, the case was jointly reviewed and corrected until consensus was reached. We will clarify this in the revised manuscript and add representative visual examples and qualitative comparisons with different methods.

  • Task difficulty (R1) Plaque segmentation is not determined by color alone. IOS colors are easily affected by illumination, saliva, specular reflection, and partial staining near the gingiva. Moreover, plaque boundaries often lie on curved tooth surfaces, where 3D geometry and spatial continuity are important. Thus, reliable segmentation requires both appearance cues and geometric information. This is supported by our results: appearance-dominant H-SAM and geometry-dominant PointNet are both inferior to our method. This also highlights the novelty of DentalPSAM, which explicitly fuses geometry-enhanced 2D appearance and appearance-enhanced 3D geometry instead of relying on color or geometry alone.

  • BMT prompt injection & geometry tokens (R1, R2) For each 3D IOS, the SAM image encoder extracts 2D appearance embeddings from the 2D UV-rendered image. For each UV-rendered image, we obtain a paired inverse-UV-mapped 3D coordinate array I^G, where each valid pixel in I^G stores its corresponding 3D position on the original IOS mesh. We use I^G as the geometry tokens. BMT encodes I^G into implicit geometric prompts (F^G). F^G are then adapted to match the 2D appearance embeddings and added element-wise to produce the geometry-enhanced image embeddings. These enhanced embeddings are finally fed into the SAM mask decoder for 2D mask prediction. We will clarify this in the revision.

  • Yellow blocks & dashed lines in Fig. 1 (R2) We introduce the stitching adapter because the pre-trained 2D SAM and 3D branches have complementary but misaligned feature spaces. The three dashed lines indicate how the adapter bridges them: the left line feeds the selected SAM intermediate features into the stitching adapter, which aligns them with the 3D branch features; the middle line feeds the adapter-refined dense prompt into the 2D branch; and the right line sends adapter-refined appearance features to the 3D branch. The yellow blocks without text labels denote learnable prompt tokens. We will clarify these in the revision.

  • UV distortion (R3) We mitigate UV distortion in three ways: 1) Each IOS mesh is split into upper, inner, and outer regions, with planar mapping for the upper surface and cylindrical mapping for the inner/outer surfaces. A 3D mesh–2D UV–3D mesh round-trip test on ground-truth plaque labels achieved a plaque Dice score of 0.94, indicating good mapping consistency. 2) The 2D branch uses not only appearance features but also 3D geometric prompts, reducing sensitivity to UV distortion. 3) The final 3D prediction fuses the 2D and 3D branches, further improving performance.

  • 2D/3D fusion (R3) The fusion is neither union nor intersection. Fusion is performed on probability maps, not binary masks, before binarization: P_fuse = (P_2D + P_3D)/2, followed by thresholding P_fuse at 0.5.



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 reviewers are split (3 / 4 / 4) on this paper. The technical idea and the new PlaqueIOS dataset are recognized as valuable, but several concerns recur across reviewers and warrant clarification in the rebuttal: 1) Dataset and task clarification. Two reviewers found the description of what dental plaque looks like on 3D IOS, how the plaque labels are obtained, and how the dataset is constructed insufficiently clear. Authors should explicitly clarify the annotation protocol and the exact segmentation target, ideally with an illustrative example. 2) Lack of qualitative results. Two reviewers noted that, for a segmentation paper, the absence of qualitative visualizations of predicted masks (vs. baselines) makes it hard to judge the claimed improvements. 3) Unclear method / Fig. 1 details. All three reviewers had trouble following parts of the architecture from Fig. 1.Specific points to address: (i) how the implicit prompts from BMT are injected into the SAM decoder and at which stage (R1); (ii) how geometric tokens are obtained and what the yellow blocks / different dashed lines in Fig. 1 represent (R2); and (iii) how the back-projected 2D predictions and the direct 3D predictions are fused (R3 suspects a union/intersection). A clearer walk-through is needed. 4) Justification of task difficulty vs. simple baselines. Raised by only one reviewer but substantive: it is unclear whether plaque is mostly distinguishable by color, in which case simple color-based / thresholding baselines might already be competitive.

  • 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 rebuttal substantively addresses the main concerns. The main remaining gap—qualitative visualizations—is a relatively minor issue I find the rebuttal satisfactory. The reviewers are split (2 accept / 1 reject), and it resolves most technical concerns



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.

    R1 (Faculty) maintains reject: core request for visual comparisons cannot be resolved under camera-ready restrictions. For a segmentation paper, lack of visual examples is a substantial presentation flaw.



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.

    Based on the outcome of the rebuttal phase, two reviewers assigned a “Accept” score, while one reviewer voted to “Reject.” During the rebuttal, the authors provided clearer explanations that enhanced the paper’s persuasiveness. Consequently, as highlighted by Reviewers R1 and R2, this information should be incorporated into the final camera-ready version to further improve the paper’s quality. The overall comment is positive, and the inclination is to “Accept” this paper.



back to top