List of Papers Browse by Subject Areas Author List
Abstract
Accurate delineation of maxillofacial anatomy in Cone-Beam Computed Tomography (CBCT) is essential for dental planning, but robust automated segmentation remains challenging, due to limited public multi-structure datasets and the high computational burden of 3D deep learning models. We present and release ToothFairy3, a large-scale CBCT benchmark that extends ToothFairy2 with 102 additional fully annotated scans and an expanded taxonomy covering 77 classes, including 32 tooth-specific pulp cavities and small neurovascular structures. ToothFairy3 comprises 582 volumes (over 40000 annotated objects), with 532 released with voxel-level labels and 50 held out for leakage-free, server-side evaluation. We also introduce U-Mamba2, an efficient U-Net-style architecture that inserts a Mamba2 state-space block at the bottleneck to capture global context with favorable computational scaling. Our proposed domain-informed training further improves the learning of maxillofacial anatomies. Across CNN, Transformer, and Mamba baselines, U-Mamba2 achieves competitive Dice/HD95 scores with lower latency and, compared with training on state-of-the-art public CBCT datasets, ToothFairy3-trained models generalize best to the hidden test set, particularly for maxillary structures.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0853_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhiqin1998/U-Mamba2
Link to the Dataset(s)
ToothFairy3: https://ditto.ing.unimore.it/toothfairy3/
ToothFairy2: https://ditto.ing.unimore.it/toothfairy2/
Pulpy3D: https://ditto.ing.unimore.it/pulpy3d/
BibTex
@InProceedings{LumLuc_ToothFairy3_MICCAI2026,
author = { Lumetti, Luca AND Tan, Zhi Qin AND Borghi, Lorenzo AND van Nistelrooij, Niels AND Rosati, Gabriele AND Addison, Owen AND Li, Yunpeng AND Vinayahalingam, Shankeeth AND Grana, Costantino AND Bolelli, Federico},
title = { { ToothFairy3: Scaling CBCT Maxillofacial Segmentation to 77 Classes with U-Mamba2 } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16883},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
It introduces Maxillo-77, a large-scale CBCT maxillofacial segmentation benchmark with 582 volumes, 77 classes, 532 public labeled cases, and 50 hidden test cases for server-side evaluation, extending ToothFairy2 with broader anatomy including 32 tooth-specific pulp classes and additional small neurovascular structures. -This paper proposes M77-Mamba2, a U-Net-style 3D segmentation model with a Mamba2 bottleneck block to capture global context efficiently. -The paper introduces domain-informed training strategies, including anatomy-aware label smoothing, weighted loss for tiny structures, and left-right mirroring with label/logit swapping. The experiments show that the proposed dataset improves cross-dataset generalization relative to prior public datasets, and that M77-Mamba2 achieves competitive Dice/HD95 with favorable runtime.
- 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 dataset scale and taxonomy expansion are substantial. Maxillo-77 extends prior public CBCT resources by combining external anatomy and internal tooth anatomy into a single benchmark with 77 classes and 582 volumes, which is likely to be useful for the field beyond this paper alone. The hidden test set and server-side evaluation also strengthen benchmarking integrity. 2.The paper does more than report in-dataset numbers. The controlled cross-dataset comparison against Pulpy3D and ToothFairy2 is a strong point, since all models are evaluated on the private Maxillo-77 test set and the paper clearly explains why this is leakage-free and fair. This supports the claim that the dataset itself has value, not just the proposed model. 3.The manuscript explicitly frames latency as part of the contribution, which is appropriate for CBCT multi-class segmentation. The runtime-vs-DSC presentation is helpful and aligns with the paper’s clinical translation angle. 4.The use of a Mamba2 block only at the bottleneck is sensible for 3D memory/runtime constraints, and the architecture is described clearly enough to follow. The domain-informed training additions are also intuitive for dental anatomy. 5.The paper compares CNN, Transformer, and Mamba-based baselines, reports grouped anatomical results, and includes an ablation of the proposed dental domain knowledge components. That gives the reader a decent picture of where gains come from.
- 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.M77-Mamba2 is presented as being inspired by U-Mamba, and the main architectural idea seems to be a fairly incremental adaptation(a U-Net-like framework with a Mamba2 bottleneck). The paper itself notes consistency with prior findings that bottleneck placement works best. As written, the methodological novelty seems modest compared with the stronger dataset contribution. 2.In Table 2, M77-Mamba2 is best overall, but the improvement over strong baselines such as nnU-Net ResEnc is not dramatic. The paper also acknowledges that the overall Dice/HD95 gain is modest, even if improvements on fine structures are more clinically relevant. This weakens the claim of a clearly superior model. 3.The proposed “dental domain knowledge” package mixes several tricks, but deeper analysis is limited. Label smoothing over related anatomies, weighted loss for tiny structures, and mirrored augmentation with label swapping are all reasonable, but these are mostly training heuristics rather than a unified methodological advance. The ablation table shows that all three together help, but the paper gives limited anatomical or error-mode analysis explaining exactly which classes benefit most and why. 4.The paper would benefit from more class-wise and failure-case analysis. Since the benchmark contains 77 classes including thin canals and pulp structures, grouped averages alone are somewhat coarse. More per-class breakdown, especially for the newly added structures, would strengthen the empirical section. 5.Runtime comparison lacks some practical detail. The paper emphasizes latency, but a more explicit protocol would help: exact inference setup, whether sliding window/test-time augmentation was enabled for all methods equally, and how hardware differences were handled across experiments. The current description is good but not fully exhaustive for a paper making efficiency a key claim.
- 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
This is a well-written and potentially valuable submission, especially because the dataset contribution is substantial and likely to be useful to the broader dental CBCT community. I encourage the authors to sharpen the positioning of the paper. At present, the strongest contribution appears to be the dataset and benchmark, whereas the model contribution is more incremental. The paper would benefit from explicitly embracing this framing rather than over-claiming methodological novelty.
I also encourage the authors to strengthen the empirical analysis in the final version. In particular, more per-class results, especially for the newly introduced pulp and small neurovascular classes, would make the contribution more convincing. A richer failure analysis would also help readers understand where M77-Mamba2 improves over nnU-Net ResEnc and U-Mamba, and where it still struggles.
Finally, because efficiency is a central claim, the inference benchmark should be described even more explicitly to avoid ambiguity. Overall, this is a promising benchmark paper with a practical model component, and with some polishing it could become a strong reference for future maxillofacial CBCT segmentation work.
- 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 overall score is mainly driven by the strength of the benchmark contribution. The proposed Maxillo-77 dataset meaningfully expands the public CBCT segmentation landscape in both scale and label granularity, and the hidden test design increases its credibility as a long-term benchmark. This alone makes the paper relevant and potentially impactful for the MICCAI community.
On the method side, M77-Mamba2 is reasonable and practically motivated, but it feels more like a careful engineering adaptation than a major algorithmic leap. The empirical gains are real but mostly modest in aggregate, which limits enthusiasm from a pure methodology standpoint. Still, since the paper is not only about the model, but also about establishing a stronger public benchmark with clinically relevant efficiency considerations, I lean slightly positive overall.
- 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
1 This paper introduces Maxillo-77, a substantially expanded CBCT benchmark for maxillofacial segmentation with 77 classes across 582 volumes. 2, This paper proposes M77-Mamba2, an efficient U-Net-style segmentation architecture with a Mamba2 bottleneck and several dental-domain-informed training strategies.
- 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 most important strength of the paper is the introduction of Maxillo-77, which meaningfully expands the public CBCT segmentation landscape in both scale and label granularity. 2.The split design, with 532 public cases and 50 hidden cases for server-side evaluation, is well motivated and helps reduce train-test leakage. The paper also describes annotation workflow, quality control, scanner diversity in the private test set, and comparison with existing public alternatives. These aspects increase the value of the benchmark for the community. 3.The paper compares CNN-, Transformer-, and Mamba-based baselines. 4.M77-Mamba2 is a reasonable and practically motivated design. 5.I expect the dataset to be useful beyond the proposed model itself. In my view, this is the kind of resource contribution that can support more rigorous and broader progress in CBCT maxillofacial segmentation.
- 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 methodological novelty of M77-Mamba2 model is relatively limited.
- 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
I believe the dataset has strong potential to benefit the community.
- 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?
I lean toward accept because I believe the paper makes a meaningful contribution primarily through the Maxillo-77 dataset and benchmark, which substantially expands the public CBCT maxillofacial segmentation landscape in both scale and label granularity. The hidden test set design, the broader structure coverage, and the cross-dataset evaluation together make this a valuable community resource. The proposed M77-Mamba2 model achieves the best overall average performance among the compared methods.
- 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
Review #3
- Please describe the contribution of the paper
This paper presents Maxillo-77, a large-scale CBCT maxillofacial segmentation benchmark extended from ToothFairy2 with 102 additional fully annotated scans and 77 anatomical classes.The authors also propose M77-Mamba2, an efficient U-Net-style architecture integrating Mamba2 state-space blocks at the bottleneck to capture global context with favorable computational scaling. The dataset extension and model design show clear advantages in both segmentation accuracy and computational efficiency.
- 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.Clinically valuable large-scale CBCT dataset with 77 fine-grained anatomical classes 2.Efficient M77-Mamba2 architecture combining U-Net structure with Mamba2 for global context modeling and low inference latency
- 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.Maxillo-77 is very similar to the open-sourced toothfairy3 dataset (https://ditto.ing.unimore.it/toothfairy3/), please describe the differences. 2.Compared to nnUNet ResEnc, M77-Mamba2 shows only a slight improvement in DSC, while inference time slightly increases; overall, the performance gain is minimal. 3.I suggest discussion on the model’s generalization ability, such as using a model trained on Maxillo-77 to test P3D.
- 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?
The proposed dataset Maxillo-77 is very similar to the open-sourced toothfairy3 dataset (https://ditto.ing.unimore.it/toothfairy3/) (Bolelli, F., Marchesini, K., van Nistelrooij, N., Lumetti, L., Pipoli, V., Ficarra, E., Vinayahalingam, S. & Grana, C. (2025). Segmenting Maxillofacial Structures in CBCT Volume. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE.)
- 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?
The proposed dataset is excellent work, but data ethics must be ensured.
- 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 are encouraged that all reviewers recognized the value of Maxillo-77 as a large-scale 77-class CBCT benchmark with a hidden server-side test set, broader anatomical coverage, and cross-dataset evaluation against ToothFairy2 and Pulpy3D. This dataset/benchmark contribution is the central goal of the paper and, as noted by the meta-reviewer, the main basis for the positive recommendation. We will revise the camera-ready to sharpen this positioning: Maxillo-77 is the primary contribution, while M77-Mamba2 is intended as a strong, efficient, and reproducible baseline for this new benchmark rather than as a purely architectural breakthrough.
Contributions (All) As recognized by all the reviewers and meta-reviewers, our main contribution lies in the release of a new public dataset that supports long term benchmarking thanks to the private test set accessible through a public platform. This serves as a major contribution to the community. Regarding the model, our intention was not to claim a major architectural breakthrough, but to provide a strong, efficient, and reproducible baseline for 77-class maxillofacial CBCT segmentation. M77-Mamba2 achieves the best average Dice/HD95 in the submitted comparison and is motivated by clinically relevant small, thin, and anatomically correlated structures. We agree with R1 that label smoothing, weighting for tiny structures, and left-right mirroring with label/logit swapping are training heuristics rather than a unified new theory. Nevertheless, the submitted ablation shows that their combination improves both Dice and HD95, supporting the value of simple domain priors in this challenging setting.
Runtime protocol (R1,R3) The submitted experiments use the nnU-Net framework for all models to keep the runtime/accuracy comparison fair. The public source-code release will provide the exact reference implementation of the inference protocol. We acknowledge that the manuscript should describe this protocol more clearly, and we will add the relevant details, including hardware, sliding-window/TTA settings, and timing procedure, in the camera-ready.
Broader Analysis (R1,R3) Our submitted cross-dataset study addresses the complementary leakage-free question of whether models trained on Pulpy3D, ToothFairy2, or Maxillo-77 generalize to the hidden Maxillo-77 test set. The reverse direction is interesting, but less aligned with the paper’s goal because Pulpy3D covers a narrower lower-jaw pulp taxonomy thus it would require to collapse or completely remove most of the labels present in the Maxillo-77 dataset. We agree that per-class and failure-case analysis would be useful, especially for newly introduced pulp and neurovascular classes; within space constraints, we reported grouped metrics, while the benchmark and server-side evaluation support finer-grained future analyses. We will also provide a complete analysis of the performance for each of the 77 classes in the public code repository of the model to provide a complete overview of the models performance.
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 recognize the significance of the newly established “Maxillo-77” benchmark for maxillofacial CBCT segmentation. I concur that this public dataset represents a valuable contribution to the MICCAI community, which is the primary basis for my recommendation of early accept. To further strengthen the paper, the authors should address the concern raised by Reviewer 3 regarding the relationship between Maxillo-77 and the ToothFairy3 dataset. While ToothFairy3 may have been released too recently for inclusion in the original submission, the camera-ready version should include a systematic comparison and discussion to clarify the distinct value of Maxillo-77 and address potential reader concerns.