List of Papers Browse by Subject Areas Author List
Abstract
3D intraoral scans (IOS) are increasingly adopted in routine dentistry due to abundant geometric evidence, and unified multi-disease diagnosis is desirable for clinical documentation and communication. While recent works introduce dental vision-language models (VLMs) to enable unified diagnosis and report generation on 2D images or multi-view images rendered from IOS, they do not fully leverage native 3D geometry. Such work is necessary and also challenging, due to: (i) heterogeneous scan forms and the complex IOS topology, (ii) multi-disease co-occurrence with class imbalance and fine-grained morphological ambiguity, (iii) limited paired 3D IOS–text data. Thus, we present IOSVLM, an end-to-end 3D VLM that represents scans as point clouds and follows a 3D encoder-projector-LLM design for unified diagnosis and generative visual question-answering (VQA), together with IOSVQA, a large-scale multi-source IOS diagnosis VQA dataset comprising 19,002 cases and 249,055 VQA pairs over 23 oral diseases and heterogeneous scan types. To address the distribution gap between color-free IOS data and color-dependent 3D pretraining, we propose a geometry-to-chromatic proxy that stabilizes fine-grained geometric perception and cross-modal alignment. A two-stage curriculum training strategy further enhances robustness. IOSVLM consistently outperforms strong baselines, achieving gains of at least +9.58% macro accuracy and +1.46% macro F1, indicating the effectiveness of direct 3D geometry modeling for IOS-based diagnosis.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6442_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{XioHui_IOSVLM_MICCAI2026,
author = { Xiong, Huimin AND Meng, Zijie AND Hu, Tianxiang AND Zhou, Chenyi AND Feng, Yang AND Liu, Zuozhu},
title = { { IOSVLM: A 3D Vision-Language Model for Unified Dental Diagnosis from Intraoral Scans } },
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
IOSVLM makes two separable contributions. First, it introduces IOSVQA, a large-scale multi-source VQA dataset of 19,002 IOS cases and 249,055 QA pairs spanning 23 oral diseases and two scan types (single-arch, occluded-arches). This is the first dataset of this scope designed for multi-disease IOS diagnosis with language generation, and it captures real-world co-occurrence of multiple diseases within a single scan. Second, IOSVLM itself is a 3D encoder–projector LLM architecture that processes native IOS point clouds for generative diagnostic VQA.
- 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.IOSVQA dataset is a significant community contribution 2.Geometry-to-Chromatic Proxy (GCP) is an original and well-analysed contribution 3.Performance results are compelling 4.The two-stage design , freeze LLM in Stage 1 to stabilise geometric alignment, fine-tune with LoRA in Stage 2 on high-quality data, is principled and consistent with best practices in VLM training. The ablation comparing training modes (Table 3) provides clear attribution.
- 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.Multi-view rendering protocol for 2D baselines is underspecified 2.GCP ablation does not fully isolate the normal-specific contribution 3.LLM selection rationale is not in the main text. 4.Inference time and memory requirements are not reported
- 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?
IOSVLM makes a solid dual contribution, the IOSVQA benchmark and the GCP proxy , and the experimental results convincingly support the claim that native 3D geometric modelling outperforms multi-view rendering for IOS diagnosis. The paper is appropriate for MICCAI as an application-of-existing-methods-to-new-problem contribution. The main concerns are: (1) the underspecified 2D baseline rendering setup, which is needed to confirm the comparison is fair; and (2) the GCP ablation gap, which limits the interpretability of the GCP contribution. These are genuine weaknesses but not sufficient grounds for rejection on their own. The inference cost omission and LLM selection justification are minor presentation issues. I recommend weak accept, with a suggestion to the authors to add rendering specifications and the LLM selection rationale in their camera-ready version.
- 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’s primary contribution is IOSVLM, a dentistry-focused vision-language model developed together with a large-scale dental VQA dataset. The main novelty of IOSVLM is that it directly encodes intraoral scan data using a 3D encoder, instead of depending on 2D rendered or captured views of the scans. The paper also introduces a geometry-to-chromatic proxy to compensate for the absence of color information in IOS data. Experimental results show that the proposed approach outperforms existing multimodal large language models from both general and medical domains on this task.
- 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.A major strength of the paper is its direct use of a 3D point cloud encoder together with an LLM for intraoral scan data. This is an important direction for digital dentistry, since IOS is inherently a 3D modality. Compared with approaches based on a small number of 2D photographs or rendered views, operating directly on 3D point cloud data preserves richer geometric information and is therefore better aligned with the underlying clinical data source. This makes the formulation both technically meaningful and practically relevant.
2.Dataset contribution The paper also makes a strong data contribution by curating a large-scale dental VQA dataset. This is valuable because progress in dentistry-focused multimodal models is often constrained by the lack of domain-specific benchmark data. Although the evaluation has some limitations, the proposed model still demonstrates stronger performance than existing off-the-shelf proprietary LLMs in this setting, which already provides a compelling reason to consider this model over more general alternatives for dental applications.
3.Another strength is the thorough ablation study. The paper also provides empirical justification for key design choices and training decisions. This makes the work more convincing, since the reader can better understand which components contribute to performance and why the proposed setup is effective.
- 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.Limited evaluation against relevant dental baselines. The paper compares IOSVLM against existing multimodal large language models from both general and medical domains, but it does not compare against prior dental vision-language models such as OralGPT [1] and ArchMap [2]. This is a notable limitation because the current superiority over general-purpose MLLMs is not especially surprising, given that those models were not designed for dentistry and were not trained for this specific data setting. To support a stronger claim of state-of-the-art performance, the method should be evaluated against existing dentistry-specific VLMs or MLLMs, even if those methods operate on 2D photographic views rather than direct 3D IOS data. Such a comparison would better isolate the value of the proposed framework.
2.Insufficient clarity on the dataset contribution. The paper states that the IOSVQA dataset is curated by aggregating data from three sources, but the provenance of each component is not sufficiently clear. In particular, it is unclear whether datasets such as MaloccIOS and DiseaseIOS were newly collected and annotated by the authors or whether they are existing resources that were reorganized into a unified benchmark. This distinction matters because the contribution is substantially stronger if the work involves original data collection, annotation, or expert curation, rather than dataset aggregation alone. The paper would benefit from a more explicit breakdown of which parts are newly created, which parts are reused, and what curation or annotation effort was required.
3.Unclear quality control for GPT-generated rationales. The paper mentions the use of GPT-4o to generate rationales or chain-of-thought style explanations. While synthetic supervision is increasingly common, it also raises concerns about factual reliability and hallucination. The manuscript does not clearly explain whether the generated rationales were validated, manually reviewed, or filtered through any quality-control procedure. This is important because incorrect or low-quality rationales may affect both training quality and interpretability claims. A discussion of rationale verification, filtering criteria, or human evaluation would make this part of the dataset and training pipeline more convincing.
4.Limited discussion of broader design trade-offs and future directions. The geometry-to-chromatic proxy appears effective for compensating for the lack of color information in IOS data, but the discussion of this design choice remains somewhat limited. In particular, the paper does not sufficiently address how this proxy compares with settings where true color information is available. Since some IOS formats can include color information, it would be valuable to discuss whether models trained on true colored IOS data might ultimately outperform proxy-based approaches, even if such datasets are currently harder to collect. This would help position the proposed method more clearly within the broader development path of dental multimodal modeling.
5.Minor presentation issue. In Figures 1(a) and 1(b), the meaning of legend labels 1, 2, 3, and 4 is unclear. These labels should be explicitly defined in the figure caption or main text for readability.
References [1] OralGPT: A Two-Stage Vision-Language Model for Oral Mucosal Disease Diagnosis and Description [2] ArchMap: Arch-Flattening and Knowledge-Guided Vision Language Model for Tooth Counting and Structured Dental Understanding
- 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?
My recommendation is based on the paper’s strong technical and clinical motivation but incomplete empirical support. The proposed use of a native 3D point cloud encoder for intraoral scan data is well motivated and represents a meaningful step for dentistry-focused multimodal modeling. The associated dataset contribution also adds value. However, the lack of comparison with prior dental vision-language models significantly weakens the strength of the paper’s claims. As a result, while I view the work as promising and relevant, the current evidence is not yet sufficient to fully support a stronger acceptance recommendation. If the concerns mentioned here are addressed, I am willing to make a stronger recommendation.
- 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
1.The paper construct the first large scale multi-source IOS diagnostic VQA dataset. 2.The paper introduce the first end-to-end VLM that takes native 3D IOS geometry as input for unified diagnosis, achieving clear performance advantages.
- 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 constructs a large-scale dataset by integrating three existing datasets. The resulting dataset effectively captures real-world scenarios involving co-existing diseases and heterogeneous inputs, which is crucial for robust model training. 2.Leveraging this large-scale dataset, the paper adopts a two-stage training strategy to develop the first end-to-end vision-language model (VLM) that directly takes native 3D intraoral scans (IOS) as input for unified diagnosis. The proposed model achieves state-of-the-art overall performance, demonstrating both strong clinical applicability and significant value to the research community.
- The ablation studies are comprehensive and well-designed. The paper provides thorough experimental evaluations of key model components and offers insightful analyses that help explain the observed performance improvements.
- 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 contributions of the paper are primarily technical, and the overall model design lacks sufficient novelty. The proposed “geometry-to-chromatic proxy” essentially replaces RGB values in the point cloud with surface normals. However, the use of normal information for feature extraction has been widely explored in prior point cloud methods such as PointNet, PointNet++, and DGCNN, which limits the originality of this component. 2.The paper does not provide sufficient details regarding the training procedure, particularly for the two-stage training strategy. For example, it remains unclear whether the loss functions used in the two stages are identical or different, and how they are designed and optimized in each stage.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Although the contributions of the paper are primarily technical and may lack strong novelty, it presents the first vision-language model (VLM) for diagnosis based on native 3D intraoral scans (IOS), effectively addressing an important gap in the field. Therefore, I would be inclined to support acceptance, provided that the authors release the source code and trained model checkpoints to facilitate reproducibility and benefit the broader community.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers and AC for their valuable comments! Reviewer #1 Q1: Rendering protocol A1: All 2D baselines use the same IOS renderings. Following orthodontic conventions, occluded-arch scans are rendered into five views: frontal in occlusion, left/right buccal in occlusion, maxillary occlusal, and mandibular occlusal. Single-arch scans use four views: frontal, left/right buccal, and occlusal. Q2: GCP ablation A2: We will add ablations with constant white, random pseudo-color, raw normals, and GCP to better isolate the normal-specific contribution of GCP in future work. Q3: LLM selection A3: Qwen3VL-8B was selected because it is open-source, reproducible, moderately deployable, instruction-following, and designed for multimodal spatial understanding, suiting geometry-grounded IOS diagnosis. Our scale/version ablation supports its accuracy-cost trade-off. Q4: Inference cost A4: IOSVLM can run on a single NVIDIA RTX 3090 and uses 20765 MB GPU memory. Average inference time is 1.71 s/sample for label-only output and 4.8 s/sample for diagnosis+rationale. Reviewer #2 Q1: Dental-specific baselines A1: Existing dental VLM/MLLMs are mostly 2D models or lack executable weights. A fair IOS baseline should support multi-image IOS renderings. The only open-source dental model we found supporting multi-image input is OralGPT-Omni-7B, but checkpoint access needs author approval. We requested access but have not yet been approved, so we cannot report its result now. We will state this limitation and add such baselines once accessible. Q2: Dataset provenance A2: IOSVQA is not simple aggregation. MaloccIOS and DiseaseIOS are private datasets collected/curated by our team, while Bits2Bites is public. Our work includes registration, label harmonization, mapping report-derived 2D diagnoses to IOS labels, QA construction, and multi-disease formatting. In MaloccIOS, 557 cases were corrected by 28 orthodontists, yielding 7,628 high-quality samples; DiseaseIOS and Bits2Bites were annotated by experts. Q3: GPT-generated rationale quality A3: GPT-4o was not used to generate disease labels. Labels come from clinical reports/expert annotations; rationales are auxiliary supervision for about 50% of high-quality Stage-2 samples. To reduce hallucination, generation was label-conditioned and constrained by disease definitions. We filtered outputs inconsistent with the target label, containing extra labels, or failing format checks. No rationale is used in noisy Stage-1.We will state that full expert validation was not performed and discuss this limitation. Q4: GCP vs true-color IOS A4: We do not claim GCP is superior to reliable true color. If large-scale true-color IOS with consistent acquisition becomes available, RGB-based or RGB+geometry models may outperform GCP, especially for appearance-related findings such as caries or mucosal changes. GCP addresses the common setting where IOS color is absent, inconsistent, or unavailable after export, and where many labels are geometry-driven. We will position GCP as a practical proxy and discuss true-color IOS as a complementary direction. Q5: Fig. 1 labels A5: Labels 1–4 denote disease-specific category labels for visualization; their meanings are not shared across diseases. We will clarify this in the caption. Reviewer #4 Q1: GCP novelty A1: The novelty is not simply using normals as features. GCP uses geometry-derived local separability cues as a chromatic proxy to bridge color-free IOS point clouds and RGB-pretrained 3D encoders. Normals are one instantiation; the idea can extend to other geometry attributes and color-missing 3D domains. Q2: Training details A2: Both stages use standard autoregressive next-token cross-entropy loss. Stage 1 trains 1 epoch with the LLM frozen, using LR 5e-6 for the 3D encoder and 3e-5 for projectors. Stage 2 freezes the 3D encoder and trains 15 epochs with LR 1e-5 for projectors and 1e-4 for LoRA parameters, rank 8 and alpha 16.We use Adam and train on NVIDIA H200.
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.
All three reviewers consider the paper has strong technical contributions and the experiments can convincingly support the advantages of the proposed methods. In addition, the authors also constructed a new large-scale dataset to facilitate further investigations. All three reviewers suggest accepting this papers.
It is anticipated that the authors can address the following concerns in the final version. (1) Release the codes and data to facilitate further studies. (2) Add rendering specifications and the LLM selection rationale. (3) If possible, see if you can compare the proposed model with more prior dental vision-language models, as suggested by one of the reviewers.