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

Abstract

Automatic tooth segmentation and identification from intra-oral scanned 3D models are fundamental problems in digital dentistry, yet most existing approaches rely on task-specific 3D neural networks trained with densely annotated datasets, resulting in high annotation cost and limited generalization to scans from unseen sources. Thus, we propose TSegAgent, which addresses these challenges by reformulating dental analysis as a zero-shot geometric reasoning problem rather than a purely data-driven recognition task. The key idea is to combine the representational capacity of general-purpose foundation models with explicit geometric inductive biases derived from dental anatomy. Instead of learning dental-specific features, the proposed framework leverages multi-view visual abstraction and geometry-grounded reasoning to infer tooth instances and identities without task-specific training. By explicitly encoding structural constraints such as dental arch organization and volumetric relationships, the method reduces uncertainty in ambiguous cases and mitigates overfitting to particular shape distributions. Experimental results demonstrate that this reasoning-oriented formulation enables accurate and reliable tooth segmentation and identification with low computational and annotation cost, while exhibiting strong generalization across diverse and previously unseen dental scans.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0737_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/znshje/TSegAgent

Link to the Dataset(s)

Teeth3DS dataset: https://osf.io/um96h/ Our private dataset: https://drive.google.com/file/d/1pA32rYoPWe4hvIsAQxaIw8Maa4jXszOs/view?usp=sharing

BibTex

@InProceedings{ZhuSha_TSegAgent_MICCAI2026,
        author = { Zhuang, Shaojie AND Yin, Lu AND Wei, Guangshun AND Li, Yunpeng AND Wang, Xilu AND Zhou, Yuanfeng},
        title = { { TSegAgent: Zero-Shot Tooth Segmentation via Geometry-Aware Vision-Language Agents } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces TSegAgent, a framework for automatic tooth segmentation and identification from intra-oral 3D scans as a zero-shot inference that integrates foundation models (SAM3 + ChatGPT 5.2). It further incorporates explicit geometric priors (e. g. , dental arch structure, symmetry, and volumetric cues) within a multi-stage vision-language reasoning pipeline, enabling interpretable and anatomically consistent predictions without task-specific training.

  • 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 major strength of this paper is its fundamental shift in how tooth segmentation and identification are approached, moving away from conventional supervised learning toward a reasoning-driven, zero-shot paradigm. Rather than training task-specific models on large annotated datasets, the authors reformulate the problem as one of structured geometric reasoning, where predictions are inferred through a combination of visual evidence and explicit anatomical constraints. This reframing is both novel and significant, as it directly addresses the high annotation cost and limited generalization that characterize existing dental AI methods.

    Building on this formulation, the paper introduces a geometry-aware vision-language agent that performs tooth identification through a sequence of interpretable, multi-stage reasoning steps. Instead of relying on a single-pass prediction, the method decomposes the task into sub-problems such as non-tooth filtering, central incisor identification, and full-arch classification, followed by an explicit error detection and correction mechanism.

  • 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.Lack of transparency in prompt design and reproducibility concerns. The proposed framework relies heavily on vision-language prompting and multi-round conversational reasoning, yet the paper does not provide sufficient details of the prompts used in the experiments. While the authors mention that “detailed prompt design can be found in the source code”, the absence of prompt specifications in the paper itself limits reproducibility and scientific scrutiny. Given that prompt engineering is a core component of the method, this omission prevents readers from: (a) Understanding how geometric priors are operationalized in the VLM. (b) Assessing robustness to prompt variations. (c) Reproducing or extending the approach without relying on external resources.

    2.Dependence on a single proprietary VLM without comparative analysis. The method is evaluated using a single vision-language model (“ChatGPT 5.2”), which introduces several concerns: (a) Lack of ablation across VLMs: It is unclear whether the performance gains are due to the proposed framework or the specific capabilities of this particular model. (b) Limited generalizability: The method’s effectiveness with other VLMs (e. g. , open-source or smaller models) is not demonstrated. (c) Reproducibility and accessibility issues: The reliance on a proprietary, potentially costly API raises barriers for adoption and independent validation.

    A more complete study would include comparisons across multiple VLMs (e. g. , GPT-4V-like systems, open-source multimodal models) or at least a sensitivity analysis.

    3.Use of non-standard metrics without sufficient explanation. The paper reports evaluation metrics such as TLA, TSA, and TIR, but does not clearly define their full forms or provide sufficient intuition regarding their interpretation. (a) These metrics are not standard in segmentation literature, and their clinical or methodological significance is not adequately explained. (b) The paper introduces abbreviations (e. g. , TLA, TSA, TIR, GT) without consistently defining them upon first use. (c) Readers unfamiliar with the Teeth3DS benchmark may find it difficult to assess the practical relevance of the reported improvements.

    4.Unclear reporting of variability (± values). Performance metrics are reported with mean ± values, but the paper does not explicitly state: (a) What these values represent (standard deviation, standard error, or confidence intervals). (b) How they computed. Without this clarification, it is difficult to assess the stability of the approach and whether improvements over baselines are meaningful.

    5.Absence of statistical significance analysis. Although the paper reports quantitative improvements over prior methods (Table 1, page 8), it does not include any statistical significance testing (e. g. , p-values, confidence intervals). Given that some improvements appear incremental, the lack of statistical analysis makes it unclear whether: (a) The observed gains are statistically significant. (b) The method consistently outperforms baselines across samples. This is particularly important in medical imaging, where rigorous validation standards are expected.

    6.Limited transparency and characterization of the private dataset. The paper evaluates generalization on a private dataset (340 scans), but provides insufficient details regarding: (a) Data distribution (scanner types, patient demographics, pathology diversity) (b) Differences relative to Teeth3DS (domain shift characterization) (c) Annotation protocol and quality control Furthermore, the dataset is not publicly available, limiting reproducibility. There is no quantitative analysis of how domain shift affects performance. While the inclusion of external validation is a strength, the lack of transparency reduces its impact.

    7.Methodological clarity and completeness of description. While the overall pipeline is conceptually clear (Fig. 1, page 3), several components lack sufficient technical detail: (a) The mask merging strategy across views is described heuristically but lacks formalization or algorithmic clarity. (b) The interaction between geometric priors and VLM reasoning is not rigorously specified. (c) The error correction mechanism is described at a high level without precise implementation details.

    As a result, it may be difficult for readers to fully understand how the system behaves in edge cases and reproduce the pipeline without access to code.

    8.Lack of clear architectural specification of the proposed framework. Although the paper presents an overall pipeline (e. g. , Fig. 1, page 3), it lacks a precise and formal description of the framework architecture. The interactions between components—such as SAM3-based segmentation, geometric processing, and the vision-language reasoning agent—are described at a high level but not specified in a rigorous or modular manner.

    In particular: (a) The data flow between modules is not formally defined. (b) The interfaces between geometric priors and VLM reasoning remain implicit. (c) The multi-round conversational process is not represented as a structured algorithm or state machine

    As a result, it is difficult to reconstruct the full pipeline without relying on external code, understand how different components contribute quantitatively to the final output, and compare the framework with existing architectures in a systematic way.

    A clearer architectural description (e. g. , formal pipeline diagram, pseudo-code, or modular breakdown) would significantly improve clarity, reproducibility, and scientific rigor.

  • 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 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.

    (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?

    This paper presents an interesting and forward-looking idea, namely, reframing tooth segmentation and identification as a zero-shot geometric reasoning problem using a geometry-aware vision-language agent. The integration of foundation models with explicit anatomical priors and multi-step reasoning is conceptually novel and aligns well with emerging trends in medical AI toward more interpretable and data-efficient approaches. In particular, the training-free pipeline and the demonstrated cross-domain generalization on a private dataset are strong and clinically relevant aspects.

    However, despite these strengths, the paper falls slightly below the acceptance threshold due to several concerns related to experimental rigor, clarity, and reproducibility.

    First, the reproducibility of the method is limited. The approach heavily depends on prompt engineering and multi-round interaction with a vision-language model, yet the prompts are not described in sufficient detail in the paper. Additionally, the reliance on a single proprietary VLM (ChatGPT 5.2) raises concerns about accessibility, cost, and whether the reported performance is attributable to the proposed framework or the specific model used. The absence of comparisons with alternative VLMs further weakens the empirical validation.

    Second, there are issues with evaluation transparency and rigor. The use of non-standard metrics (TLA, TSA, TIR) without a clear explanation reduces interpretability for a broader audience. Moreover, while improvements over baselines are reported, the paper does not include statistical significance analysis, making it difficult to assess whether the gains are meaningful. The reporting of mean ± values is also not clearly defined.

    Third, the external validation on a private dataset, although valuable, lacks sufficient detail regarding data characteristics and domain differences relative to Teeth3DS. Since the dataset is not publicly available, it is difficult to independently verify the claimed generalization benefits.

    Finally, while the overall idea is compelling, some components of the methodology are not described with enough precision (e.g., mask merging, reasoning interaction), which further affects reproducibility and clarity.

  • 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 core limitations identified in the original review remain unresolved after rebuttal and require substantial revision. Therefore, I maintain my original recommendation.



Review #2

  • Please describe the contribution of the paper

    This paper proposes TSegAgent, a zero-shot framework for automatic tooth instance segmentation and FDI identification from intra-oral scanned 3D models. The key idea is to bypass the supervised-learning paradigm by reformulating the task as a geometric reasoning problem.

    Experiments on Teeth3DS and a private clinical dataset show that TSegAgent achieves either state-of-the-art or competitive performance compared to fully supervised methods.

  • 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.Well-motivated problem formulation: The paper makes a compelling case for reformulating tooth segmentation and identification as a zero-shot geometric reasoning task rather than a supervised learning problem. This is a meaningful conceptual shift that addresses real limitations of existing methods: high annotation cost and poor generalization across scanners.

    2.Good cross-domain generalization: The proposed method’s performance on the private dataset is impressive, this private dataset represents a domain shift (scanned plaster models with defects and artifacts vs. digital intra-oral scans). Other supervised methods shows significantly more degradation.

    3.Interpretable pipeline: The multi-round conversation design that decomposing tooth identification into non-tooth filtering, central incisor identification, full-arch classification, and error correction is intuitive and interpretable, and could be clinically meaningful. Each step has a clear anatomical rationale

  • 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.Heavy dependence on a proprietary, closed-source VLM (ChatGPT 5.2): The entire tooth identification pipeline relies on it as the vision-language model. This raises several concerns:

    1.1 Reproducibility: Behavior of commercial API-based models can change without notice

    1.2 Scientific rigor: It is unclear how much of the system’s success is attributable to the proposed geometric reasoning pipeline vs. the raw capability of a state-of-the-art commercial VLM.

    1. Potentially biased evaluation on the private dataset: The private dataset is only used as a test set, with no details provided about its annotation quality beyond stating they are “scanned plaster models with more severe defects.”

    Critically, since supervised baselines are trained on Teeth3DS and tested on this private set, the comparison inherently favors the zero-shot method. A fairer comparison might include domain adaptation or fine-tuning baselines.

    3.Computational cost analysis is missing: The paper claims “low computational cost” but provides no runtime analysis, API call counts, or cost-per-scan estimates.

  • Please rate the clarity and organization of this paper

    Good

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

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

  • 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

    (1) VLM robustness: Please provide results with at least one additional VLM to demonstrate that the framework’s effectiveness is not solely due to the capabilities of ChatGPT 5.2.This would strengthen the claim that the reasoning pipeline, rather than the VLM, is the primary driver of performance.

    (2) Runtime and cost analysis: Please report inference time per scan, number of VLM API calls per scan, and approximate cost. This is essential for evaluating clinical practicality.

  • 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?

    My recommendation of Weak Accept is based on the following rationales:

    On the positive side, the paper presents a genuinely novel problem formulation that shifts dental segmentation from supervised learning to zero-shot geometric reasoning. The cross-domain generalization results, while I’d prefer to have a more detailed introduction of the private dataset, shows good generalization potential. And the proposed pipeline is intuitive and interpretable, which increase its clinical practicality.

    On the negative side, the most critical concern is the reliance on a single proprietary VLM with no evidence of generalizability to other VLMs. This makes it difficult to disentangle the contribution of the proposed reasoning framework from the raw power of the underlying model. The missing runtime analysis is also a notable gap given the claims of practical clinical utility.

  • 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.

    After carefully reading the authors’ rebuttal, I maintain a positive assessment and upgrade my recommendation to Accept. Here is my reasoning:

    1.VLM dependency (our primary concern, shared with R1):

    The rebuttal discloses that preliminary trials with open-source VLMs (GLM-5, Doubao Seed 2.0) “remained competitive to other baselines.” While I would have preferred quantitative numbers, this statement suggests the geometry-aware reasoning framework, not just ChatGPT 5.2’s raw power, is responsible for the method’s effectiveness.

    2.Private dataset evaluation (our secondary concern):

    The authors make a fair point that their experiment evaluates zero-shot cross-domain transfer, and asking for domain adaptation baselines corresponds to a different experimental setting. They also provide additional dataset details (plaster models, desktop scanner, diverse pathologies) and offer to make it publicly available. I find this response satisfactory.

    3.Runtime and cost (our tertiary concern):

    The rebuttal provides concrete numbers: ~6,284 tokens and ~120 seconds per case. This is reasonable for offline clinical analysis and much faster than manual annotation, adequately addressing our concern.

    Overall, the paper’s core contribution of reformulating tooth segmentation as zero-shot geometric reasoning remains novel and well-executed.

    The rebuttal adequately addresses our most critical concerns (VLM generalizability, runtime cost, private dataset fairness). The remaining gaps are real but do not outweigh the conceptual novelty, strong cross-domain results, and practical relevance. I believe this paper is above the acceptance threshold.



Review #3

  • Please describe the contribution of the paper

    This paper proposes TSegAgent to address the limitations of existing methods, which heavily rely on dense annotations and exhibit poor generalization ability. To tackle these challenges, the authors leverage a general-purpose foundation model (e.g., SAM3) and incorporate explicit geometric inductive biases derived from dental anatomy. Experimental results demonstrate that TSegAgent outperforms other methods in both tooth segmentation and identification tasks, while requiring lower computational and annotation costs.

  • 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 work addresses the challenge of sparse annotations, which is a common limitation in medical image analysis. 2.The proposed method effectively integrates state-of-the-art foundation models with vision-language models (VLMs). 3.The approach incorporates domain-specific prior knowledge from dental anatomy, which is well-motivated. 4.The model provides interpretability for its predictions, which is valuable for clinical applications.

  • 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 unclear whether incorporating geometric information from dental anatomy solely through text prompts is sufficient for the VLM to perform reliable reasoning. It would be helpful to clarify whether additional structured information (e.g., explicit geometric features extracted from data instead of prior knowledge) could further improve performance. 2.The reason behind the improved segmentation performance is not fully explained. As described, the agent primarily performs reasoning for tooth identification, and there appears to be no learnable component to refine the segmentation masks generated by SAM3.The authors are encouraged to clarify how the proposed framework leads to better segmentation results. 3.The error detection and correction module is not clearly defined. It is unclear whether this component targets segmentation quality or only targets tooth identification. More detailed clarification is needed. 4.For the segmentation task, reporting only IoU is insufficient. It is recommended to include additional commonly used metrics, such as the Dice coefficient, for a more comprehensive evaluation. 5.The manuscript claims that the integration of interpretable geometric priors reduces uncertainty in visually ambiguous cases. However, no qualitative or quantitative results are provided to support this claim. It would strengthen the paper to include interpretability analyses, such as examples of multi-round interactions demonstrating how the agent arrives at its decisions, or ablation studies showing how the incorporation of geometric priors influences VLM predictions and reduces uncertainty.

  • 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?

    In this work, they used their private dataset. But I’m not sure that authors will publish their data.

  • 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?

    All major concerns have been discussed in the above section (the list of major weaknesses). Here, I would like to highlight two key questions. First, it is unclear how TSegAgent refines the segmentation masks generated by SAM3.The current pipeline does not appear to include a learnable component for mask refinement, and further clarification is needed. Second, it remains unclear how the incorporation of geometric priors influences VLM predictions and reduces uncertainty. In particular, it would be helpful to understand whether this mechanism contributes to the interpretability of the agent’s decisions.

  • 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.

    The authors’ response addresses most of my concerns. However, one point remains unclear: how does TSegAgent refine the segmentation results produced by SAM3? The current manuscript mainly describes a mask-merging strategy for projecting 2D masks back onto the 3D mesh, but it is not entirely clear whether this process also helps refine or correct erroneous SAM3 segmentations. I encourage the authors to further clarify this mechanism in the camera-ready version.



Author Feedback

[R1.1,R3.1,R3.5 Prompt] The prompt consists of both textual instructions and annotated images. In images, tooth masks are colored with labeled text, indicating each instance. We ask the VLM to first identify non-tooth masks, then determine the central-incisor IDs to separate the left and right sides, and finally classify each instance’s tooth type based on its shape and anatomical position. We take into account the characteristics of dental data, including the dental arch order to guide tooth arrangement and tooth volume information to help the VLM distinguish and order individual teeth. Their effectiveness has been validated through our ablation studies. Without geometric priors, the VLM is more prone to errors in FDI ordering and instance-ID recognition. [R1.2,R2.1 VLM reliance] Preliminary trials with other VLMs (like open sourced GLM-5 and Doubao seed 2.0) remained competitive to other baselines, but ChatGPT gave the best result. However, the focus is the zero-shot geometry-aware reasoning framework rather than optimizing the choice of VLM. [R1.3,R3.4 Metrics] TLA is mean of normalized Euclidean distance between GT teeth centroids and the closest localized teeth centroid. TSA is the average F1-score over all instances. TIR is the percentage of true identification cases relative to all GT teeth in the testing set. Their formulas are defined within the paper. These metrics are appropriate for this task because they separately evaluate localization, segmentation, and identification, while mIoU provides an additional standard overlap-based reference. [R1.4,R1.5 Statistics] We use mean standard deviation across test scans, which is a common convention in machine learning. We’ll clarify that. Regarding statistical significance, we agree that p-values or confidence intervals could provide an additional statistical perspective, but current results indicate not only higher average performance but also stronger stability under domain shift. Therefore, while formal significance testing could further strengthen the statistical analysis, the reported values already support the robustness trend observed in our experiments. [R1.6,R2.2 Private dataset] The private dataset follows the same annotation scheme as Teeth3DS produced by experienced dentists. The data were obtained by scanning plaster dental models using a desktop scanner, and include diverse challenging cases such as missing teeth, crowding, bubbles in models, defects, and other dental abnormalities. We can make the data publicly available. We disagree that the evaluation inherently favors TSegAgent, the goal of this experiment is to evaluate zero-shot cross-domain transfer without target-domain annotations, while domain adaptation or fine-tuning baselines would require labeled target-domain data and therefore correspond to a different setting. The domain shift is quantitatively reflected by the clear degradation of supervised baselines on the private dataset, whereas TSegAgent maintains stronger performance and stability. [R1.7,R1.8 Details] We have clearly described the entire framework in an algorithmic and formulaic manner through Fig. 1 and Sec. 2.The geometric priors help reorder masks and are filled in text prompts, which can be found in the code. [R1.7,R3.2,R3.3 Error correction] The module only identifies and corrects wrong FDI predictions, not segmentation masks. The dental arch can imply symmetry and ordering of teeth ID, through this feature, we can propose potential errors to VLM, and let VLM reconsider its prediction to the right one. Segmentation improvements mainly come from non-tooth filtering and over-segmentation removal, while error correction mainly improves FDI accuracy. [R2.3 Cost] In thinking mode, each case costs 6284 tokens and around 120 seconds on average, which is acceptable for offline analysis and much faster than manual annotation. It would be more economical and faster without thinking mode, however, performance may be slightly reduced.



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.

    This paper introduces TSegAgent, a zero-shot tooth segmentation framework that leverages Vision-Language Models (VLMs) as reasoning agents to provide anatomical guidance for downstream segmentation models. The core idea is genuinely novel: rather than training task-specific segmentation networks, the authors propose using a ChatGPT 5.2 VLM to reason about dental anatomy, generate geometric priors, and direct a foundation segmentation model (e.g., SAM) to produce tooth delineations without any fine-tuning on target data. The reported cross-domain generalization results—are indeed promising and address a critical clinical bottleneck, namely the prohibitive cost of manual annotation across diverse imaging modalities.

    However, the paper suffers from two fundamental weaknesses that the reviewers correctly identified. First, there is a Reproducibility Crisis: the entire pipeline depends on a proprietary, closed-source VLM (ChatGPT 5.2), which is not publicly accessible, version-controlled, or reproducible by the community. As R1 and R3 both noted (scores of 3), this dependency makes it impossible for other researchers to verify, replicate, or build upon the work. The authors provide no fallback using open-source alternatives (e.g., LLaVA, GPT-4V, or other openly available VLMs), nor do they specify the exact prompting strategy with sufficient detail to allow approximate reproduction. This is a serious methodological concern for a MICCAI publication.

    Second, there is an Evaluation Rigor issue. The paper introduces three non-standard metrics—Tooth Localization Accuracy (TLA), Tooth Segmentation Accuracy (TSA), and Tooth Identification Rate (TIR)—without providing clear formal definitions, mathematical formulations, or references to established evaluation protocols. As R2 (score of 4) pointed out, while the approach is intriguing, the lack of standard metrics (e.g., Dice, IoU, Hausdorff distance) makes it difficult to compare against existing benchmarks or to assess clinical significance. R3 further highlighted that the ablation studies are insufficient, and the reported improvements lack statistical significance testing.

    Despite these limitations, the paper’s novelty and potential clinical impact are high. The idea of using VLMs for anatomical reasoning as a zero-shot segmentation strategy represents a paradigm shift that could significantly reduce annotation burden in medical imaging. The reviews collectively suggest that the core contribution is valuable but requires substantial strengthening in reproducibility and evaluation rigor.

  • 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.

    Two reviewers support acceptance (R2=4, R3=3→Accept). Rebuttal effectively addressed VLM dependency concerns (shown to be ~6284 tokens, ~120s/case), runtime cost, and private dataset fairness. Zero-shot tooth segmentation is a clinically relevant problem with promising cross-domain generalization.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Reviewers agreed the core idea is novel and the cross-domain results are strong and clinically relevant. The rebuttal addressed the main concerns well. Two reviewers found the rebuttal convincing, with only a minor camera-ready clarification (how SAM3 masks are refined) outstanding. The novelty and practical relevance outweigh the remaining gaps, and I recommend 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.

    Accept

  • Please justify your recommendation.

    This is a borderline paper with mixed reviews. While reviewers agreed that the paper introduces a novel, reasoning-driven zero-shot paradigm for tooth segmentation and identification, and that it demonstrates strong cross-domain generalization without requiring large annotated datasets, they also raised serious concerns about the paper’s soundness and reproducibility: (1)​the heavy reliance on a proprietary, closed-source VLM limits scientific rigor and makes it difficult to disentangle the contribution of the proposed geometric reasoning framework from the raw power of the model; (2)​key methodological details—such as prompt design, architectural specifications, and the segmentation refinement mechanism—are insufficiently transparent, hindering reproducibility; and (3)​the evaluation lacks statistical analysis and uses non-standard metrics without adequate explanation, while the private dataset comparison may unfairly favor the zero-shot approach. The rebuttal provided clarifications on runtime, dataset characteristics, and preliminary tests with open-source VLMs, which partially addressed these concerns. After careful deliberation, I recommend accepting this manuscript due to its technical novelty.



back to top