Paper Info Reviews Author Feedback Meta-Review Back to top

List of Papers Browse by Subject Areas Author List

Abstract

We present a dataset and baseline for ontology-grounded structured prediction from dental Cone-Beam Computed Tomography (CBCT) volumes. Building on the public ToothFairy3 benchmark (532 volumes with expert-level segmentations), we contribute (i) a total of 893 free-text clinical reports for 529 publicly available CBCT volumes, (ii) their conversion into validated RDF/Turtle (Resource Description Framework) instances aligned with a clinician-designed OWL (Web Ontology Language) ontology spanning 13 finding types and multiple qualifier axes, and (iii) a strong baseline demonstrating the effectiveness of our setup and establishing a foundation for future work. We formulate CBCT reporting as a three-stage structured prediction problem—i.e., finding detection, anatomical slot allocation, and property prediction—and introduce a hierarchical evaluation suite of six clinically interpretable metrics that decouple detection, localization, and characterization. A baseline model using frozen multi-scale VoxTell features, a structure-indexed encoder, and ontology-driven prediction heads achieves strong results under 5-fold cross-validation, with stage-decoupled analysis identifying presence detection as the primary deployment bottleneck. Dataset, ontology, and code are publicly released: https://github.com/AImageLab-zip/CBCT-Report.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0618_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/AImageLab-zip/CBCT-Report

Link to the Dataset(s)

https://ditto.ing.unimore.it/toothfairy4/

BibTex

@InProceedings{LumLuc_OntologyGrounded_MICCAI2026,
        author = { Lumetti, Luca AND Di Bartolomeo, Mattia AND Pellacani, Arrigo AND Anesi, Alexandre AND Grana, Costantino AND Bolelli, Federico},
        title = { { Ontology-Grounded Structured Prediction for Dental CBCT Reporting } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16878},
        month = {September},
        page = {pending}
}

Reviews

Review #1

Please describe the contribution of the paper

The paper proposes a new dataset, or more specifically a set of structured diagnostic labels for an existing dental CBCT dataset (ToothFairy3). They also adapt/propose a framework (using VoxTell) to this dataset-label pair.
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.

In general, the approach makes a lot of sense. With a lot of the focus in dental CBCT being on segmentation, moving towards a diagnostic assessment and automated reporting would be of interest to the end-user. The ‘ontology’ structured approach is also sensible, allowing for the use of established diagnostic metrics rather than the more fuzzy LLM output metrics that are often used. Building all of this on an existing benchmark dataset also makes perfect sense, allowing for a quick adoption without apparent barriers.
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.
- There is limited overall methodological novelty. While the diagnostic labels are undoubtedly a valuable contribution, which can benefit future research efforts, the proposed framework involves a frozen pretrained encoder (VoxTell) with a few standard components added. While similar work has been performed in other domains (most notably RadGraph), they are the first to propose this for dental CBCT.
- While some ablation was performed, there is limited overall experimentation, including a lack of a comparison with alternative models (incl. ‘conventional’ detection/classification models trained specifically on this dataset) and, essentially, a comparison with human performance.
- The reliance on segmentation is a general risk, especially for certain structures or tasks. How does one expect to detect periapical radiolucencies (assuming no root resorption) by segmenting hard tissues like the jaw bones and teeth? How does one expect to detect caries (for which CBCT is not indicated in the first place) based on a tooth mask? At the same time, many of the tasks (with third molar - mandibular canal contact being the prime example) have a rather deterministic relation between their segmentation and their diagnostic assessment, which can be put into practice with a rule-based system rather than an extra DL model. For localization in particular, the proposed approach may be ‘overkill’ or ‘too easy’, with a well-designed rule-based system being faster and foolproof.
- The reliability of labeling is insufficiently demonstrated. CBCT reporting is notoriously prone to inter- and intra-observer variability. Eight observers were included with varying experience and background (what is a ‘maxillofacial clinician’, is it a maxillofacial surgeon?), but no mentioning is made of calibration, observer conditions, or assessment of observer agreement. This is a major limitation, which I would recommend the authors to address before releasing their labels publicly.
- Results appear to be averaged over the various diagnostic tasks, obscuring possible tasks for which the model might perform poorly. At the very least, a performance range for the different tasks should be provided (I understand that a full breakdown of performance per task is not feasible due to the word limit).
- At a more conceptual level, I wonder if this is the best way to apply an AI-assisted diagnostic tool. This gets into human-AI interaction a bit, which is a domain with plenty of room for discussion, but the way that I see it is that a clinician carries full responsibility for a report, meaning that explainability is key. Therefore, they will want an AI prediction to be shown on the CBCT data (in almost every case on the original, grey level data, not on segmentations) using some kind of heatmap approach, allowing them to approve or overrule the finding. Then, it can be transcribed into a report (which is relatively trivial because the output is already structured at this point). The proposed approach can probably be turned into something like this, but currently falls short of offering sufficient information to the clinician to make an ultimate decision.
- The use of an existing public dataset, while being a strength, is also a limitation. This dataset was not selected based on diagnostic considerations, because it was released as a segmentation benchmark. Thus, it is unlikely to be representative for the types and distributions of pathology in a typical population. Some of this is a bit obscured; for example, you have ‘mandibular canal course’ for each patient (because of course they both have two canals) but how many of them have third molars? Out of those, how many third molars are impacted? Out of those, how many are in close contact with the canal? As for the other pathologies, there is plenty to discuss (‘maxillary sinus finding’ can be dozens of things, ‘pathological lesion’ can be literally anything, ‘jaw atrophy’ was found 600 times which seems concerning unless this just refers to periodontal bone loss. . . ).
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?

While the dataset is anonymized, it is unclear whether the patients provided consent for this ‘extra’ information (which is more sensitive than segmentation masks, because it concerns disease) to be released.
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.

(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?

While I believe this work has several merits, I believe it might be too incremental from a methodological standpoint, and I identified too many issues that would need a lot of extra work to address. The recommendation accounts for the competitiveness of this particular conference. My suggestion to the authors would be to consider my feedback in a constructive manner, consider subsequent steps through careful communication between technical and clinical people involved, and then work towards a full paper (I would not suggest a conference paper).
Reviewer confidence

Very confident (4)
[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.
I appreciate the difficulty in writing a rebuttal to a critical review without being able to make substantial changes to the paper (and without any room for further experimentation). This is a limitation of the MICCAI system itself, but it inevitably turns this rebuttal/re-rebuttal stage into a meaningless exercise. That being said, while I appreciate the authors’ responses, I have to stick with my original assessment due to the very competitiveness of this congress. I also will briefly follow up on a couple of issues/rebuttals, where I feel like the authors may have oversold their work or not entirely addressed the issue in their response:
- There is a fundamental issue to this framework’s clinical applicability, because most diagnostic tasks need to be assessed (by humans or AI) on the greyscale (original data) rather than segmentations. Structured reporting is a relatively trivial task for any automated pipeline, because outputs are inherently structured. All the downstream opportunities listed in the rebuttal are also quite trivial to implement, and should not be framed as something where the proposed method acts as an enabler (or ‘the key’, in the authors’ words). This is overselling at best, and misleading at worst.
- The authors also offer situations in which a rule-based approach may not work (which I suggested as an alternative for specific tasks where a deep classifier would be overkill), but the proposed method would be just as prone to the same issues (like artefacts).
- The authors mention that detailed info can be found on the dataset page, but a cursory check did not allow me to find clear breakdowns of label distributions (which are likely to be skewed and not representative) or model performance per task. The authors mention results for three handpicked tasks in their rebuttal; at face value, these are relatively ‘easy’ tasks for a model that uses segmentations as input. The fact that they do neither provide any range in performance in their rebuttal, nor the performance for some of the tasks I listed as questionable applications for this framework, makes me suspect even more that certain performance/application issues are being obscured.
- No answer was provided regarding the fundamental question of annotation reliability.

Review #2

Please describe the contribution of the paper

The authors introduce a novel dataset comprising 893 free-text clinical reports paired with 529 publicly available dental CBCT volumes. They established an OWL/RDF ontology, validated by clinicians, to map 13 specific clinical finding types and their qualifiers. They propose a three-stage structured prediction baseline model and a decoupled evaluation suite (measuring detection, localization, and characterization independently) to overcome the limitations of standard text-generation metrics.
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.

Shifting the focus from pure anatomical segmentation to ontology-grounded clinical reporting directly addresses a major bottleneck in automating dental radiology workflows. Novel Evaluation Framework: Standard free-text metrics often fail to capture clinical correctness. The proposed stage-decoupled metrics offer a interpretable and rigorous way to identify model failure points. The authors commit to publicly releasing the dataset, the ontology, and the code.
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 proposed deep learning baseline is largely a composite of existing techniques. It relies on standard components, specifically frozen VoxTell features concatenated across levels, followed by standard cross-attention mechanisms and MLPs. However, considering this is explicitly positioned as a Dataset and Baseline paper, the lack of architectural innovation is a secondary concern compared to the value of the dataset.

2.The pipeline uses adaptive max-pooling to align the segmentation mask with the feature map. Table 1 demonstrates robustness against standard nnU-Net predictions. However, in clinical dental CBCT, extreme scenarios like massive metal artifacts can lead to catastrophic segmentation failures. Conceptually, how do you expect the structure-wise attention pooling to behave when the input mask is severely degraded by such artifacts? A brief answer on this potential limitation in the final manuscript would provide valuable clinical context.

3.The dataset exhibits a significant long-tail distribution, which is a realistic challenge in medical imaging. Could you share your insights on how this imbalance impacts the model’s reliability on the rarest classes?
Please rate the clarity and organization of this paper

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

The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

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

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

(6) Strong Accept — must be accepted due to excellence
Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

My recommendation to accept this paper. The authors have curated 893 free-text reports authored by experienced clinicians for 529 publicly available CBCT volumes. However, the most significant contribution is not just the raw text, but the rigorous conversion of these reports into validated OWL/RDF ontology instances. By grounding the data in a clinician-designed schema with SHACL constraints, the authors have provided a structured resource. While the baseline structured prediction model relies on standard, existing architectural components, the public release of this dataset and the ontology framework single-handedly justifies publication. It fills a gap in the community.
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 am satisfied with the responses and recommend accepting this paper. The authors have addressed my initial concerns.

The details regarding the dataset statistics and class-wise agreements provided in the rebuttal. Pointing to the task-wise statistics on the dataset platform resolves my previous questions about the data distribution.

The baseline model and the decoupled evaluation framework serve as a reproducible starting point. Overall, this dataset and the proposed structural framework will be a valuable resource for the medical imaging community.

Review #3

Please describe the contribution of the paper

The paper mainly contributes a dataset of 893 clinical reports for publicly available CBCT volumes and a baseline model for reporting prediction. The model can be regarded as a foundation of the computer-aided reporting for dental CBCT.
Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

The paper is strong in the effective use of the existing metrics and framework for constituting the baseline model, which is appropriate for the dental CBCT reporting. The dataset presented in this paper has high clinical value and reusability.
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 validity of result is limited a single dataset of 532 public CBCT volumes. The comparison and ablation experiments with baseline methods are somewhat insufficient.
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?

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 baseline model established a foundation for the reporting prediction of dental CBCT. However, the comparison of the performance with the other methods was not made. A single dataset also limits its generalizability.
Reviewer confidence

Very confident (4)
[Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

N/A
[Post rebuttal] Please justify your final decision from above.

N/A

Author Feedback

Reviewers recognized the clinical relevance and community interest in our proposal (R1,R2). They valued the proposed ontology-based structure, addressing a major bottleneck in the literature (R1,R2), and the release of the dataset, ontology, and code, supporting the adoption, transparency, and reusability of a clinically valuable resource (All). They appreciated the rigorous evaluation design, including appropriate stage-decoupled metrics for identifying model failures (R2,R3).

Contributions [All] We clarify that the primary contribution lies in the proposed ontology, the ontology-grounded reporting dataset, and the decoupled evaluation framework. The model is intended as an efficient, reproducible baseline rather than a fully optimized architecture or deployment-ready product. As highlighted by R2, “the lack of architectural innovation is secondary […] compared to the value of the dataset”. It is true that several model components build on existing techniques. However, the structure-aware attention pooling is a practical adaptation for volumetric CBCT reporting, as it substantially reduces the number of tokens and enables training.

Clinical deployment and explainability [R1] The proposed system is intended to generate structured report suggestions for a clinician to review, verify, correct, and approve. Providing a report proposal does not remove the clinician’s responsibility; rather, it can support faster and more consistent reporting. The current localization and segmentation outputs already provide a useful basis for verifying reported findings. Future integration into end-user software could additionally show heatmaps, selected CBCT crops, and segmentation overlays to help clinicians inspect each predicted finding. This work should be seen as a methodological step toward assisted CBCT reporting, not as a final clinical tool. Structured report generation also enables other downstream uses, including retrieval, dataset curation, clinical audit, and automatic computation of finding distributions. We agree that rule-based systems might be suitable, especially when the relation between anatomy and diagnosis is deterministic; however, designing hard-coded rules that are robust across anatomical variability, acquisition differences, artifacts, and edge cases is challenging. Our approach aims to learn these associations directly from data, being easier to extend and generalize without requiring ad-hoc engineering at each iteration.

Dataset statistics [R1,R2] Due to space constraints, the paper could not include the full distribution of every finding and property. On the dataset download platform page, detailed task-wise statistics and label distribution are already reported. Agreement is high overall for finding presence and characterization (presence-F1 0.81, char acc. 0.83) and remains substantial for structured report matching (report-F1 0.67). Examples of class-wise agreements are: MandibularCanalCourse (presence/instance-F1: 0.97/0.93), EndodonticTreatment (0.87/0.72), and DentalImplant (0.83/0.59). Our ontology-based representation is the key that enables this kind of analysis, providing an additional value beyond report generation itself. For example, identifying impacted third molars in close contact with the mandibular canal is difficult to extract reliably from narrative free-text reports but straightforward from structured reports.

Ontology details [R1] The names assigned to the findings were suggested by the maxillofacial surgeons; we agree that they might sound too general, but in the file that defines the ontology, it is possible to have the details about them. In “ontology.ttl” from the code in the original submission, it is possible to Ctrl-F for a specific finding, and a comment provides the requested description. As an example, we can here report the one related to MaxillarySinusFinding: “Maxillary sinus pathology: mucosal thickening, mucositis, mucosal polyp, retention cyst, […].”

Meta-Review

Meta-review #1

Your recommendation

Invite for Rebuttal
Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

The paper offers a labeled dataset for diagnostic interpretation of CBCT dental scans. All the reviewers have appreciated the need for such a dataset that goes beyond segmentation. The dataset is still rather small and there are concerns expressed in the collection details and low methodological novelty which I invite authors to rebut.
After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

Reject
Please justify your recommendation.

The rebuttal provided is weak at best and although I believe a dataset paper should be considered a valid submission even with modest methodological novelty, the paper seems to have some experimental flaws which the reviewers continue to suspect even after the rebuttal.

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 introduces a valuable ontology-grounded dataset for structured prediction in dental CBCT reporting, along with a baseline model and decoupled evaluation framework. All three reviewers acknowledged the clinical relevance and community need for such a resource. Reviewer #2 and Reviewer #3 were supportive, emphasizing the dataset’s reusability and the novelty of shifting focus from segmentation to diagnostic reporting. Reviewer #1 raised substantive concerns regarding limited methodological novelty, insufficient demonstration of labeling reliability, lack of task-wise performance breakdown, and clinical applicability. In the rebuttal, the authors clarified that the primary contribution is the dataset and ontology rather than architectural innovation, provided pointers to detailed dataset statistics and inter-annotator agreement (presence-F1 0.81, report-F1 0.67), and explained the system’s intended role as an assistive tool for clinician review. While the rebuttal successfully addressed many of Reviewer #2 and #3’s concerns, Reviewer #1 remained unconvinced, reiterating issues about the framework’s clinical grounding and data representativeness. Nevertheless, given the strong positive assessments from two reviewers and the acknowledged value of the publicly released structured dataset, which fills a critical gap in the field, the overall consensus supports 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.

The main concerns were missing dataset details, unclear label taxonomy and annotation workflow, LLM-based clinical metrics, overstated clinical claims, and insufficient ethics/data-access information. The rebuttal addressed several of these issues by clarifying the report language, MRI sequences, exclusion criteria, annotation workflow, metric definitions, fine-tuning setup, moderated claims, and controlled-access IRB-approved data release. Reviewers 1 and 4 support acceptance after rebuttal, although Reviewer 2 remains concerned that key dataset and taxonomy details are still not sufficiently presented in the main paper.

back to top

Ontology-Grounded Structured Prediction for Dental CBCT Reporting

Author(s):