List of Papers Browse by Subject Areas Author List
Abstract
Orthopantomograms (OPGs) are the standard panoramic radiograph in dentistry, used for full-arch screening across multiple diagnostic tasks. While Vision Language Models (VLMs) now allow multi-task OPG analysis through natural language, they underperform task-specific models on most individual tasks. Agentic systems that orchestrate specialized tools offer a path to both versatility and accuracy, this approach remains unexplored in the field of dental imaging. To address this gap, we propose OPGAgent, a multi-tool agentic system for auditable OPG interpretation. OPGAgent coordinates specialized perception modules with a consensus mechanism through three components: (1) a Hierarchical Evidence Gathering module that decomposes OPG analysis into global, quadrant, and tooth-level phases with dynamically invoking tools, (2) a Specialized Toolbox encapsulating spatial, detection, utility, and expert zoos, and (3) a Consensus Subagent that resolves conflicts through anatomical constraints. We further propose OPG-Bench, a structured-report protocol based on (Location, Field, Value) triples derived from real clinical reports, which enables a comprehensive review of findings and hallucinations, extending beyond the limitations of VQA indicators. On our OPG-Bench and the public MMOral-OPG benchmark, OPGAgent outperforms current dental VLMs and medical agent frameworks across both structured-report and VQA evaluation. Code will be released upon acceptance.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0308_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Zhaolin-Yu/OPGAgent
Link to the Dataset(s)
N/A
BibTex
@InProceedings{YuZha_OPGAgent_MICCAI2026,
author = { Yu, Zhaolin AND Yang, Litao AND Babicka, Ben AND Hu, Ming AND Hao, Jing AND Huang, Anthony AND Huang, James AND Jin, Yueming AND Wu, Jiasong AND Ge, Zongyuan},
title = { { OPGAgent: An Agent for Auditable Dental Panoramic X-ray Interpretation } },
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
An agentic tool OPGAgent for report generation from Orthopantomogram X-ray images is proposed. There are three modules in the agentic tool: (1) a hierarchical evidence gathering module which analyses global, quadrant level and tooth-level anomalies, (2) a toolbox which is used for spatial analysis, detection of different anomalies in different resolutions, and a collection of experts who can vote on these aspects, and (3) a consensus agent collecting the different decisions from the expert zoo and consolidating the decisions into final report with possible conflict resolution. The paper also proposes a structured report format in the form of a triplet (location, field, value) for precise reporting. Experiments show that the proposed tool achieves good precision, recall and reduced false positives for different dental parts on two benchmark tasks.
- 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: The proposed tool OPGAgent is inspired by agentic AI tools like MedAgent, DentalGPT, OralGPT etc. and tries to improve the agentic AI collaboration for OPG understanding specifically. The analysis is done in multiple scales: a global image level, quadrant level and individual tooth level. A collection of models (example: MaskDino, Dino, Medsam, Yolo) is used for analyzing different components of OPG image. Then the expert zoo of several large foundational models DentalGPT, OralGPT and GPT-5.2 act on the annotations provided by the analysis tools. The consensus is gauged by a few aligned predictions (two or three).
B: The report format is made structured.
C: Experiments show improved results on most of the anomaly types in multiple regions.
- 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 type of analysis done by OPGAgent is very similar to existing tools. How it differs from other tools is not made very evident in the paper.
The methodological novelty is quite limited. Access to all foundational models used in the OPGAgent tool is quite limited due to the paid nature of these foundational tools. This severely limits benchmarking of the OPGAgent tool against other similar types of tools.
Details of which type of anomalies are looked for by the models are missing. For example: details of which detection tools are used for various pathological models are not clear from the paper.
The set of tasks handled by utility tools is also not very clear. Some examples are given in section 3.2 but a complete list of tasks is not provided.
What opinions are collected from expert zoos? How do these opinions look in terms of output form? How are they processed by the consensus subagents? These details are also missing.
Disagreement handling is also not very clear. Is it deterministic and found by a set of rules? If so, what are these rules? Or is it based on parsing and understanding the conflicts? If natural language understanding of conflicts is done, how is it done?
Is the triplet form of report done for all compared tools? If not, how are the precision, recall, FP values for all tools computed?
What is the nature of memory? How are the different analyses stored in memory? What is the structure of the storage for each level: global, quadrant and tooth level?
Though the premise of the tool is laudable, the tool appears very cryptic as described in the paper. Many details are not clear. I am also not sure if such tool building without much methodological advances is a good fit for the conference.
- 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 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
Please check my comments on weaknesses.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The OPGAgent tool queries multiple models for analyzing an OPG image and queries various VLMs or LLMs to create reports of individual parts. The consensus handling and conflict resolution are not very clear from the paper. Details of many components of the tools are missing in the paper and leaves the reader with not much pointers to conduct a comparison study. Methodological advancements are not present in the paper.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
In this paper, the author(s) propose an agentic workflow for dental imaging to generate reports based on automated findings. 1.They breakdown the scan to a tree like structure in scope, process in every level. 2.They include a variety of tools that can be called by the agent in all the levels. 3.They unify the reporting format in a smart way. 4.Evaluate based on individual findings.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper is well written and flows nicely. Motivation and processes are clear. 2.The approach is on the cutting edge of current research and development landscape. 3.In places the author(s) go beyond “just and LLM output” like the deterministic reporting with the parser. 4.The variety of the tools ranging from other DL methods to more simple and mathematical utilities are well defined and utilized. 5.Clearly applicable to existing data, easily verifiable results (ignoring the reproducibilty concer) / auditable
- 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.
Evaluation:
- The ablation (table 3) and the main quantitative comparison (table 1) use the same dataset. However, in table 1 the authors report on 4 fronts [E.M. (Exact Match), Det. (Presence), Loc. (Location), Cls. (Grade/Type), FP (False Positives/Case)] while in table 3 they only report on the first front. At the very least the authors should explain why that is.
- There are no statistical analyses for any of the presented results such as confidence intervals or statistical testing for significance.
- The coefficients of the score table are not explained anywhere on why they are chosen.
Implementation details:
- I am missing explanation and the implementation details of the tools in the toolbox. Many of the findings depend on these tools yet there is very little individual detailed explanation on these tools.
Cost of using the proposed model:
- The authors do not mention anything about the cost of using the online models, how many tokens/queries are used. There is no information regarding any of the actual usage of online models.
- How long an analysis takes is similarly not mentioned.
There is no discussion of any kind in the paper.
- 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
Please enable hyperlink in the Latex so the citation and table/section references are clickable. There are some minor typos such as: Title of section 2.1
- 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 concept, goal, and writing of the paper is definitely worth a 5.However, the gap in the evaluation regarding the usage of online models, the lack of detail for reproducibility, and the absence of a discussion reduces my grade for this paper to a 4.I would still like to see this paper in the conference and in the proceedings as is, while it would be much better with the addressed concerned fixed.
- 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 am not sure on how much they can actually edit the paper. (They say they will add the open source model comparison and statistical results.). If they do what they say they will do I like it. I think the point of the paper is solid, the novelty in my opinion is there both in application and in small details that are related to the niche application. I would like to see this paper at MICCAI.
Review #3
- Please describe the contribution of the paper
The authors propose OPGAgent, the first agentic framework specifically designed for Orthopantomogram (OPG) interpretation. It uses hierarchical evidence gathering (from global to tooth-level), a specialized toolbox of perception models, and a consensus subagent to resolve anatomical conflicts. Additionally, the paper introduces OPG-Bench, a novel structured-report protocol for auditing pathological findings and hallucinations.
- 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.
- Architectural Novelty: The shift from single-pass VLMs to a multi-tool agentic workflow successfully balances high-level reasoning with task-specific precision.
- Hallucination Mitigation: The use of a Consensus Subagent and anatomical constraints to verify findings provides a robust mechanism for clinical auditability.
- Rigorous Evaluation: The OPG-Bench protocol offers a more clinically relevant evaluation than traditional VQA by using structured (Location, Field, Value) triples.
- 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.
- Planner Dependency: The system relies heavily on the GPT-5.2 planner, which may limit deployment in privacy-sensitive or resource-constrained clinical environments that require local hosting.
- FP Interpretation: The authors acknowledge that reported False Positive rates are likely inflated because the agent detects valid pathologies not documented in the original clinical reports; this makes it difficult to fully quantify the precision gains against human-derived ground truths.
- 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
To further enhance the quality and clinical impact of this article, I recommend the following scientific refinements: 1.The abstract should cover all parts of the manuscript so that it can be more reader-friendly. For this purpose, at the end of the abstract structure of the manuscript, the authors should express the quantitative results and the conclusion of the work based on the obtained findings. 2.It is recommended that the authors standardize the selected keywords using MeSH (Medical Subject Headings) terminology. Adhering to this established framework ensures technical consistency and improves the accuracy of cross-platform literature searches. 3.The introduction would benefit from a more structured flow. I recommend beginning the first section by clearly establishing the importance of the research topic and outlining the current challenges or gaps in the field. Once this context is well‑defined, you can then transition into discussing deep learning approaches as potential solutions. 4.The system relies on GPT-5.2 as the primary planner. While high-performance LLMs are effective, they pose challenges for clinical deployment due to data privacy and the need for substantial computational resources. The authors could discuss the feasibility of using smaller, locally hosted models (e.g., Llama-3 or Med-Gemma) for the planning module. An ablation study or a brief comparison showing the performance drop when using a more “mobile-friendly” planner would be highly valuable for clinical translation. 5.In the results section, the authors state that the reported False Positive (FP) rates may be slightly inflated because the agent detects valid pathologies not documented in the original clinical reports. To clarify this, the authors could perform a secondary expert validation on a small subset of these FP. If a significant percentage of these are indeed “undocumented correct findings”, it would further strengthen the claim of the model’s superior sensitivity over human-derived ground truths.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The integration of agentic planning with specialized perception tools addresses a critical bottleneck in dental AI—balancing versatility with diagnostic accuracy. The consensus mechanism and the new OPG-Bench protocol represent high-quality contributions to the field of auditable medical AI.
- 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 am maintaining my score of 5 (Accept). The original manuscript already presented a highly novel and clinically relevant contribution by introducing the first multi-tool agentic framework (OPGAgent) and a structured evaluation protocol (OPG-Bench) for dental panoramic X-rays. My primary reservation was the pipeline’s dependency on a proprietary, cloud-based planner (GPT-5.2). In the rebuttal, the authors successfully demonstrated the framework’s robustness by swapping the planner for a locally hostable, open-source model (Qwen3.6-27B), resulting in only a minimal (4.5 pt) performance drop. Furthermore, they provided rigorous statistical analyses (bootstrap CIs, p < 0.001) and clarified the implementation details of their tools, effectively addressing all major concerns raised across the reviews. This is a rigorous, high-quality paper that deserves acceptance.
Author Feedback
We thank the reviewers and the AC for their constructive feedback and for acknowledging OPGAgent’s multi-scale Agent design, the structured (Location, Field, Value) triplets, and its performance improvements across multiple regions and dental diseases. We would like to clarify the following points to address the reviewers major concerns.
[R1, Meta] Novel Contributions Compared to Prior Work (1) Unlike VLMs that just rely on single-pass outputs, OPGAgent achieves more accurate and auditable results via specialized dental tools – outperforming SOTA VLMs(+6.9%) on OPG-Bench, and beyond the runner-up (+3.5%) on the external benchmark MMOral-OPG VQA. (2) OPGAgent is the first Agent framework dedicated to OPG diagnosis and analysis. To meet the clinical demands for auditable and interpretable results, we explicitly link findings to FDI tooth anchors and utilize tool-assisted consensus verification, producing reports that are suitable for clinical use. Furthermore, our results highlight the limitations of general medical Agent frameworks on OPG tasks: using the same planner, tools, prompts, and a shared Parser Agent, OPGAgent achieves a +21.9% performance improvement over MedAgent-Pro (Table 1). (3) The current VQA paradigm focuses only on individual questions and cannot assess overall performance. We provide a more comprehensive approach: OPG-Bench, a novel evaluation paradigm based on these triplets, providing more evaluation dimensions than VQA on completeness, localization, and classification.
[R1, R2, R3, Meta] Detailed Tool Design and Reproducibility
Planner: (R1) The Planner follows a Global -> Quadrant -> Tooth structure (detailed in §2.2) to generate the report. (R2) Each case processes roughly 8.5k input tokens and 5k output tokens, takes 1-3 minutes, and costs approximately $0.06.To further demonstrate reproducibility (R2) and address R3’s concern regarding planner dependency, we attempted to replace the planner with an open source model(Qwen3.6-27B), the results show only a 4.5 pts Score drop on OPG-Bench.
Tools requiring weights or code: (R1, R2) YOLOv9 [23] trained on DENTEX [8] is used for quadrant/tooth detection. MedSAM [17] is used for tooth segmentation. We leverage the open-source MaskDINO/DINO models released by OralGPT-Omni [10] to detect features diseases (caries, periapical/apical findings, fillings, and impacted teeth). An IoU tool links diseases, teeth, and FDI numbers on the schema to establish consensus, additionally, the mandibular canal and maxillary sinus are paired with the Tooth Mask to identify extraction risks.
Consensus Subagent: (R1) When a conflict arises, the subagent first checks for validity against anatomical constraints. The ROI is then routed to the respective Experts for verification. If the conflict persists, Majority Voting is used to make the final determination(detailed in §2.4).
Code and model weights will be released upon acceptance as we promised in the submission.
[R2, Meta] Statistical Analysis Table 3 uses the same Exact-Match metric as Table 1, we previously thought this was sufficient, so we omitted the Det./Loc./Cls. columns due to space limitations. Per R2’s suggestion, we will include these additional metrics in Table 3 of the final version. The Score assigns weights 50% to Exact-Match because it strictly aligns with every label in the ground truth, while detection, localization, and classification are weighted 20%, 20%, 10% to further reflect lesion detection, anatomical grounding and disease classification capabilities. Paired bootstrap (B=1000, n=1006): all 9 methods σ ≤ 0.0084 (OPGAgent 0.0080). OPGAgent’s 95% CI [48.21, 51.36] is non-overlapping with the strongest baseline Gemini-3-Pro [41.47, 44.56], 8/8 pairs p < 0.001 (Holm-corrected).
[Final Version] In the final version, we will include the Qwen3.6-27B on OPG-Bench, add Det./Loc./Cls. to Table 3 per R2’s suggestion, statistical analysis align MeSH keywords, enable LaTeX links, and fix the §2.1 typo.
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 presents OPGAgent, an LLM-agentic framework for interpreting dental panoramic X-ray (OPG) images. The system integrates medical vision-language models with specialized perception modules and a consensus mechanism for auditable radiology reports. Key contributions include a hierarchical evidence gathering approach (global, quadrant, tooth-level), a specialized toolbox encapsulating spatial, detection, utility tools, and expert zoos, and a Consensus Subagent that resolves conflicts through anatomical constraints. The paper also introduces OPG-Bench, a structured-report protocol using (Location, Field, Value) triples.
Strengths:
- Reviewer 3 (Score 5) highlights the architectural novelty, particularly the shift from single-pass VLMs to multi-tool agentic workflow, hallucination mitigation via Consensus Subagent, and rigorous OPG-Bench evaluation protocol
- Reviewer 2 (Score 4) acknowledges the well-written presentation, cutting-edge approach, and auditable mechanism
- The paper addresses an important clinical need in dental imaging where automated reporting could significantly impact workflow
Limitations:
- Reviewer 1 (Score 2) raises substantial concerns about novelty, noting the methodological novelty is quite limited and many details are not clear
- Reviewer 2 identifies critical gaps: ablation and main comparison use the same dataset, lack of statistical analyses (confidence intervals, significance testing), missing implementation details of tools, and no cost/token analysis for online models
- Reviewer 3 notes planner dependency on GPT-5.2 which limits deployment in privacy-sensitive clinical environments
- Reproducibility concerns due to proprietary backbone model
Consensus assessment: The paper presents a well-motivated application with practical engineering effort. The average score of 3.67 (borderline) reflects divergent reviewer opinions. The rebuttal should address: (1) clear articulation of novel contributions beyond pipeline assembly, (2) external validation on independent datasets, (3) statistical rigor with confidence intervals and significance testing, (4) detailed cost and feasibility analysis for clinical deployment.
*ndividual reviewer scores:
- Reviewer 1 (Balamurugan Palaniappan): 2 (Reject)
- Reviewer 2 (Baris Imre): 4 (Weak Accept)
- Reviewer 3 (Ali Kazemi): 5 (Accept)
- 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 explicitly support acceptance (R2=4, R3=5). Rebuttal effectively addressed technical concerns: clarified underlying models, added statistical analysis, demonstrated open-source alternatives. The agent-based approach for auditable dental panoramic X-ray interpretation has practical clinical value.
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.
The paper presents OPGAgent, an agentic framework for generating structured reports from panoramic dental X-rays. Reviews were mixed: concerns included limited methodological novelty, unclear implementation and consensus details, reproducibility, and reliance on proprietary models; strengths included the timely clinical application, hierarchical multi-tool design, structured reporting, and promising results.
The rebuttal addressed key concerns by clarifying the tool components and consensus logic, adding statistical analysis, and reporting runtime/cost details. Overall, despite some remaining presentation weaknesses, the work is clinically relevant, timely, and sufficiently strong for acceptance. I recommend acceptance and encourage the author to include the promised changes (especially the promised links) in the final submission.
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.
I recommend acceptance. The reviews were initially mixed, mainly due to concerns about novelty, implementation details, statistical analysis, and dependence on proprietary models. The rebuttal sufficiently addressed these issues by clarifying the OPG-specific agent design, toolbox implementation, consensus mechanism, runtime/cost, statistical significance, and an open-source planner replacement. Reviewers 2 and 3 support acceptance after rebuttal. Although some reproducibility and systems-integration concerns remain, the paper is clinically relevant and suitable for acceptance.
