List of Papers Browse by Subject Areas Author List
Abstract
Dental disease detection on panoramic radiographs is challenged by extreme scale variation, tiny lesions with ambiguous boundaries, and the vertical anisotropy of tooth anatomy. To improve sensitivity to small lesions while retaining computational efficiency,we propose DentMamba, a two-stage framework: a global tooth localization stage to isolate tooth ROIs and suppress background structures, and a fine-grained detector built on anatomy-aware global–local modeling. In the second stage,we introduce the AnatoMamba block with anisotropic convolutions to encode vertical root priors while avoiding the quadratic cost of global attention, and a bidirectional MultiScale Adaptive Spatial Attention Gate fusion module (MASAG) to reduce semantic misalignment during feature fusion. Furthermore, we propose a Dynamic Scale-Aware Hybrid Loss (DSAHL) that adapts the regression objective across lesion scales and aligns classification confidence with localization quality.Experiments on a large clinical panoramic dataset (20,000 images; 18,534 used for cross-validation)demonstrate that DentMamba achieves 84.6 AP with 17M parameters, with notably improved smalllesion performance (APS 52.1),outperforming recent detectors under a strong accuracy-efficiency trade-off. The evaluation code can be accessed at:https://github.com/zhkangzhkang/Dent_Mamba.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0668_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/zhkangzhkang/Dent_Mamba
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaKan_DentMamba_MICCAI2026,
author = { Zhang, Kang AND Zhuang, Shaojie AND Wei, Guangshun AND Zhou, Yuanfeng},
title = { { DentMamba: An Anatomy-Aware Global-Local Hybrid Network for Large-Scale Multi-class Dental Disease Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Main contributions of this paper are threefold. First, a large-scale panoramic X-ray dataset with over 20,000 annotated images is built which covers 12 types of dental lesions. Second, Anatomamba Block and MASAG Module are proposed to address the vertical anisotropy of teeth and feature misalignment, which integrate anatomical features and use a bidirectional gating mechanism to improve detection of subtle lesions. Third, a Dynamic Scale-Aware Hybrid Loss (DSAHL) is designed that combines scale-aware regression with an NWD-based focal loss, enabling accurate detection of lesions across different scales
- 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.
①Good balance between accuracy and efficiency: DentMamba achieves 84.6 AP on the clinical dataset and 74.6 AP on OralXrays-9 with only 17M parameters, significantly outperforming many Transformer-based methods and demonstrating a stronger accuracy-efficiency trade-off. ②Strong performance on small object detection: Thanks to the Anatomamba Block and DSAHL design, the model achieves 52.1% on small lesion detection, outperforming existing methods by 5.9 percentage points, validating the effectiveness of anatomical priors in detecting tiny lesions.
- 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.
①The paper adopts a Mamba architecture, but the experimental section does not include comparisons with any other Mamba-based detection models. This makes it difficult to know whether the performance gains come from the Mamba architecture itself or from the anatomy-aware modules and loss function design. ②The framework relies on a first-stage tooth localization step to crop ROIs, but does not discuss how localization errors propagate to the second-stage detector. If the localization stage misses or misaligns a tooth, the subsequent detection will simply fail. The paper lacks a quantitative analysis of how sensitive the overall framework is to this coupling.
- 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.
(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?
Score of 4 (Weak Accept) is mainly based on following considerations. On the positive side, the paper builds a large-scale panoramic X-ray dataset with over 20,000 annotated images covering 12 types of dental lesions, providing a valuable public benchmark for the field. The proposed Anatomamba Block and MASAG module are specifically designed to address the vertical anisotropy of teeth, showing some novelty. The model achieves competitive results on two datasets with only 17M parameters, striking a good balance between accuracy and efficiency. However, there are also clear limitations. The paper claims to adopt a Mamba architecture, but the experimental section does not include comparisons with any other Mamba-based detection models. The framework also relies on a first-stage tooth localization step to crop ROIs, but does not discuss how localization errors propagate to the second-stage detector.
- 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 #2
- Please describe the contribution of the paper
This paper makes a practical contribution to multi-class dental disease detection on panoramic radiographs. The main idea is to build a two-stage framework that first extracts tooth regions and then performs lesion detection within these regions. More specifically, the paper introduces the anatomy-aware AnatoMamba module, the MASAG fusion module, and the DSAHL loss to improve modeling of dental structure and to improve detection of multi-scale, especially small, lesions. The reported results on both a large private clinical dataset and the public OralXrays-9 benchmark suggest that the method is competitive.
- 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.
Overall, I think the strengths of the paper are mainly in its task relevance and empirical performance. The problem is clinically meaningful, the method is built in a task-aware way, and the reported results are enough to make the paper competitive. 1) The task is clinically meaningful: dental disease detection on panoramic X-rays is both useful and realistic, naturally involving small lesions, strong class imbalance, large scale variation, and complex background. 2) The use of anatomical prior is interesting. The AnatoMamba module uses a 5×3 anisotropic convolution to better match the vertical structure of teeth and roots, which is a reasonable design choice for dental images. 3) The paper has a clear focus on small-lesion detection. The DSAHL loss combines NWD and GIoU in a scale-adaptive way, which is relevant because small lesion detection is one of the hardest parts of this task. 4) The reported results are strong. The paper reports 84.6 AP / 52.1 APS on the private clinical dataset and 74.6 AP / 24.5 APS on OralXrays-9.On the public benchmark, these numbers are competitive with recent methods, especially on APS. 5) The model is relatively efficient, with only 17M parameters, while several strong baselines such as Co-DINO and MOQAM are much larger.
- 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.
My main concerns are about fairness and transparency. The method appears effective, but the paper does not fully isolate the source of improvement, and some experimental details need clearer explanation.
1) Fairness of the two-stage framework. Does the first stage (tooth localization/segmentation) rely on extra tooth-level annotations? The ablation also lacks a key baseline—ROI cropping plus a standard detector—making it unclear whether gains come from the proposed modules or simply from ROI standardization.
2) Dataset release plan. Currently only evaluation code is provided, with no clear commitment to releasing the private dataset, which limits reproducibility.
3) Incomplete ablation study. The absence of ablation results on OralXrays-9 makes it difficult to assess the design’s stability on the public benchmark.
4) proofreading issues. Table 1 appears to contain mismatched citation indices (e.g., CFinet [2], DETR [20]). Please double-check and correct these, and also verify if similar issues exist elsewhere in the manuscript.
- 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?
I give this paper weak accept because the problem is important, the system is well developed, and the reported results are strong on both the private dataset and the public benchmark. In particular, the paper shows clear gains on small-lesion detection, which is one of the main challenges of this task.
But I do not give a higher score because the weakest part of the paper is not the final performance, but the lack of clarity about where the gain comes from and whether the comparison is fully fair. The role of the first-stage ROI module, the protocol on the public benchmark, and the transparency of the private dataset all need clearer explanation. So I currently lean toward a conservative accept rather than a strong accept.
- 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
This paper constructs a large-scale panoramic dataset with over 20,000 annotated images across 12 lesion categories, providing a valuable benchmark for future dental research.
- 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.
This paper is well-written and very easy to follow.
- 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.In Equation 2, the dimensions of $X$, $G$, $\alpha$, and $\beta$ are not clearly defined. Additionally, in Equation 4, the Wasserstein distance term $W_2^2(\mathcal{N}_p,\mathcal{N}_g)$ is introduced without a proper mathematical definition in the preceding text. 2.The authors claim to introduce a valuable clinical dataset with 20,000 images covering “12 lesion categories”. However, the names, clinical definitions, and diagnostic criteria of these 12 categories are nowhere to be found in the text. Figure 2(A) only lists categories as numerical indices (0 to 11). Without knowing what diseases are actually being detected, the clinical value and validity of the dataset cannot be assessed. 3.The framework proposes a “coarse-to-fine” two-stage pipeline, utilizing a “YOLOv8 Segmentation” model in Stage 1 to crop teeth. However, the performance, training details, and exact mAP/Dice score of this Stage 1 model are completely omitted. More critically, the authors do not address the issue of error propagation. If Stage 1 fails to detect a tooth or crops it inaccurately, how does this affect the Stage 2 disease detection? A cascaded pipeline cannot be properly evaluated without isolating the performance of each stage. 4.In Figure 1(C), the architecture diagram includes major blocks labeled “AIFI” and “CCFF”. These acronyms are never defined, explained, or even mentioned in the main text. 5.The paper claims strong “computational efficiency” and compares its 17M parameter model against heavy Transformers and older CNNs (Table 1). However, since the authors use YOLOv8 in Stage 1, they should logically compare DentMamba against modern, efficient real-time detectors like YOLOv8, YOLOv9, or YOLOv10 for the Stage 2 task.
- 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.
(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?
This paper addresses an important clinical problem and presents a large-scale dataset.
- 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
N/A
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.
The paper got 3 x weak accept, which suggests a positive feedback from reviewers. I would suggest to address the issues raised by reviewers in the camera-ready.