List of Papers Browse by Subject Areas Author List
Abstract
Segmenting the temporomandibular joint (TMJ) disc from MRI is essential for accurate diagnosis of internal derangement, yet it remains unreliable in practice due to its small size, low contrast, and morphological variability. Existing methods, primarily adapted from general segmentation architectures, often produce fragmented or anatomically inconsistent masks, leading to unstable measurements of disc position and shape for downstream diagnosis. To address these challenges, we propose TISC, a TMJ disc segmentation framework that integrates semantic anchoring with clinical metadata-guided boundary refinement. The framework first establishes robust disc localization in the foundation model feature space via a Prototypical Semantic Anchoring (PSA) module that aggregates adjacent-slice MedDINOv3 features and derives a prototype-driven similarity map. It then performs targeted boundary refinement through a Clinical-Metadata Point Refinement (C-MPR) module, with point-wise predictions modulated by Mouth Open Limitation (MOL), a clinical indicator associated with disc displacement without reduction. On a large-scale cohort of 2,488 PD MRI volumes from 1,300 patients, our method achieves up to a 4.96 Dice improvement over strong baselines across diverse architectures, delivering more anatomically coherent and clinically reliable TMJ disc segmentation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0906_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
Link to the Dataset(s)
N/A
BibTex
@InProceedings{JuDay_Anatomically_MICCAI2026,
author = { Ju, Dayun AND Kim, Chanyoung AND Jung, Sunyoung AND Jung, Hyo-Jung AND Lee, Chena AND Park, Younjung AND Hwang, Seong Jae},
title = { { Anatomically Consistent TMJ Disc Segmentation via Semantic Anchoring and Clinical Priors } },
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
The PSA module and C-MPR are well designed and integrated to cooperate to solve the targeted problem. The PSA module is responsible for anatomical consistency by utilizing the MedDINOv3 to extract semantic features to achieve stable localization of disc. And C-MPR is responsible for boundary refinement with the help form PSA module. In addition, clinical prior of MOL is cleverly utilized in both modules to improve the preformance.
- 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 targeted problem is practical and challeging. The module designs are very clear and interpretable, for example the use of MedDINOv3 in PSA module. MOL is effectively used to help with the performance improvment.
- 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 explanation of some method designs and experimental settings is insufficient. , such as the temperature in softmax function, the ratios of sampled points. The proposed framework is a 2D segmentation model in essence, although the PSA module aggregate adjacent slices. As we know, 3D segmentation models are advantageous over 2D models in anatomical inconsistency performances. I suggest the authors add some metrics of continuity or topology accuracy to further prove the effectiveness of their method, such as, number of connected components, fragmentation score.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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 method proposed in this paper exhibits a certain degree of novelty. And experiments results show the effectiveness
- 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.
The authors’ reply addresses my questions adequately. Although their method still has some limitations, the paper presents certain novelty.
Review #2
- Please describe the contribution of the paper
In this paper, the authors proposed TSIC, a TMJ disc segmentation framework that integrates semantic anchoring with clinical metadata-guided boundary refinement. Previous methods struggle with spatial uncertainty due to small target localization and precise boundary delineation after coarse localization. To address above problems, the proposed PSA module aggregates adjacent-slice MedDINOv3 features and generates a similarity-based localization map，stabilizing predictions and reducing false positives under pathological displacement. The C-MPR module adaptively adjusts boundary predictions at the point level conditioned on clinical metadata, enabling correction of ambiguous regions. Moreover, MOL, a binary indicator of limited mouth opening as auxiliary metadata, guides spatial attention in PSA and modulates refinement features via FiLM layers, promoting adaptive localization and boundary correction across pathological states. The quantitative and qualitative experimental results show the effectiveness of the proposed method.
- 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 writing of this paper is easy to understand and there are few spelling or grammatical error. b) The quantitative and qualitative experimental results show the effectiveness of the proposed method, and the article achieves SOTA performance in comparative experiments. c) The method tries to describe the network architecture as well as the objective functions thoroughly. d) Considering the specific difficulties of the TMJ task, the proposed TSIC utilizes object detection foundation model and adjacent-slice features to produce an enhanced similarity based localization map, overcoming the challenge in localizing small targets. Besides, it adaptively adjusts boundary predicions at the point level to resolve the issue of boundary ambiguity. e) The innovative introduction of auxiliary metadata effectively enchance the performance.
- 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.
a) In ‘Introduction’ section, the review of current methods is insufficient, lacking more detailed introduction of relevant specific approaches on TMJ task. Meanwhile, foundation models such as MedSAM and MedDINOv3 are mentioned. What’s the advantages and disadvantages of them compared to general model. b) The comparison methods in experiments are not up-to-date. It only compares against old methods like the standard UNet and its very basic variants and new but less effective SAM-based model. If possible, more state-of-the-art methods specifically for TMJ Segmentation tasks should be included for comparison. c) In ‘Application’ section, predicted masks and images are concatenated for downstream diagnostic tasks, the setting should be discussed. Besides, more state-of-the-art medical image classification methods should be taken into account, rather than very basic ResNet. Actually, more detailed experimental elaboration and corresponding tabulated comparison results should be included. d) Although the writing of this paper is easy to understand, there are still some minor writing issues can be corrected. For example, the first line of ‘Ablation Study’ section references a non-existent ‘Table 2(a)’. The architecture diagram uses two different representations for the ground truth: ‘Mgt’ and ‘GroundTruth’, Although they are used in different modules, the notation should remain consistent if there is no change. e) The submission does not mention open access to source code or data. Although it’s allowed, the ‘Method’ section should be more specific in the absence of source code. While the paper provides a detailed description of the implementation process, it lacks more intuitive mathematical formulations.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
a) The proposed method is novel, utilizing popoular foundation model and multi-modal concepts. Futhermore, the network is designed to address specific research challenges with promising experimental results. b) The comparison methods in experiments are not up-to-date. If possible, more state-of-the-art methods specifically for TMJ segmentation tasks should be included for comparison c) The downstream diagnostic task lacks interpretability. More state-of-the-art medical image classification methods should be taken into account.
- 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.
The authors have fully addressed all my core concerns. For all the issues, they have provided a complete explanation and committed to release the code to ensure reproducibility. I recommend acceptance.
Review #3
- Please describe the contribution of the paper
The main contribution of this paper is the proposal of an anatomically consistent TMJ disc segmentation framework that integrates semantic anchoring and clinical metadata. Specifically, the method introduces a Prototypical Semantic Anchoring module to improve localization using foundation model features, and a Clinical-Metadata Point Refinement module to enhance boundary accuracy by leveraging clinical priors such as Mouth Open Limitation.
- 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 addresses a clinically meaningful and challenging problem, namely TMJ disc segmentation, which is difficult due to small object size, low contrast, and high anatomical variability. The proposed framework is well-motivated and technically sound. The combination of semantic anchoring in foundation model feature space and metadata-guided refinement is a thoughtful design that directly targets both localization and boundary ambiguity. The method effectively incorporates clinical prior knowledge. This adds interpretability and domain relevance.
- 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 proposed method partially relies on empirical design choices and task-specific clinical priors. While these components are effective for the current dataset, it remains unclear whether the approach can generalize to external datasets. Therefore, validation on independent external datasets is necessary to demonstrate the robustness and generalizability of the method. The current evaluation mainly focuses on integrating the proposed method into baseline models, rather than comparing with recent dedicated TMJ disc segmentation methods. As a result, it is difficult to properly assess the relative performance level of the proposed approach within the current state of the field. We note that although the proposed method achieves notable improvements over the baselines, there is still room for further performance gains. Providing an analysis of failure cases or samples with large segmentation errors would be beneficial for better understanding the limitations of the method and guiding future improvements. In the ablation study, we observe that the performance gains mainly stem from the MOL and PSA components, while the contribution of C-MPR appears relatively limited. Moreover, MOL also plays an important role in supporting PSA. This raises a practical concern for clinical deployment: whether MOL can be consistently obtained for all patients. In cases where MOL is partially missing, it remains unclear whether the model can still maintain stable performance.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
Overall, this paper presents a well-designed framework for TMJ disc segmentation, addressing a clinically meaningful problem. The integration of semantic anchoring and clinical metadata is interesting and shows clear improvements over baseline models. However, several issues limit a stronger recommendation. The method relies on empirical designs and task-specific priors, and its generalizability remains unclear due to the lack of external validation. The comparison experiments are also not sufficiently comprehensive, making it difficult to assess the method’s standing relative to recent TMJ segmentation approaches. In addition, there is still room for performance improvement, and analysis of failure cases would help better understand the method’s limitations. The ablation study further suggests a strong dependence on MOL, raising concerns about robustness when such metadata is unavailable in real clinical settings. Overall, the work is promising but slightly below the acceptance threshold. I would be open to a higher score if these concerns are addressed in the rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank Reviewers 1,2,3 (R1,R2,R3) and the Meta-Reviewer (M) for the constructive feedback.
[M,R2,R3: Baseline rationale] We agree that recent TMJ methods need more discussion. For example, Liu et al. use a topology-preserving multi-branch decoder with joint loss, and Hu et al. extend TransUNet with edge enhancement and cross-slice attention. However, without a public TMJ benchmark, paper-level SOTA comparison is unreliable due to different datasets/settings, and fair reproduction is difficult without released code/protocols. We therefore used strong, widely validated, reproducible baselines for a fair test across CNN, self-configuring, Transformer, and foundation-model families.
[M,R3: External validity] We acknowledge the lack of external TMJ validation as a limitation. Suitable cohorts are scarce because no public TMJ benchmark exists. As a sanity check, our pipeline improved Dice from 0.82 to 0.89 on a structurally similar knee meniscus task, but this was not treated as external validation due to anatomical/task mismatch. Within available TMJ data, we used patient-level 5-fold CV on a relatively large cohort for fair evaluation and observed consistent gains across backbones. Future work will pursue multi-institutional validation.
[M,R3: C-MPR] C-MPR is a boundary refinement module, not the main localization component, so its effect is clearer in HD95 than Dice. Comparing Exp. 3 and Exp. 6 in Table 2, adding C-MPR lowers HD95 from 6.01 to 5.27 (↓12.3%), supporting its role in reducing boundary outliers after localization.
[M,R3: MOL dependency] MOL is routinely assessed in TMJ evaluation [23] and is naturally recognized during closed/open-mouth MRI acquisition, so missingness is unlikely. Nevertheless, if unavailable, the model can operate as the no-MOL variant (Table 2, Exp. 5), which still outperforms the vanilla backbone, showing that MOL improves performance but is not strictly required.
[R1,R2,R3: Code] Model code will be released for implementation clarity and reproducibility.
[R1: Empirical design choices] Most method-specific settings were chosen through qualitative inspection and empirical validation. The actual softmax temperatures are τ=0.05 for MOL=0 and τ=0.1 for MOL=1, yielding sharper and broader localization, respectively (the reversed description will be corrected). For point sampling, we follow PointRend [11] with 75% uncertainty-based samples and a 25% coverage pool, where 25% is drawn from PSA-defined salient regions and the rest remains random. We will clarify these details.
[R1: Topology metrics] Thank you for the suggestion. We agree and will reflect topology metrics, such as CC, FS, and Euler, in the released evaluation code for additional verification.
[R2: Related-methods] We agree that these methods need better context, as they serve different roles rather than a direct pros/cons comparison. Prior TMJ-specific studies are mostly CNN-based adaptations to TMJ MRI and were too briefly summarized. MedSAM is a strong recent medical segmentation baseline, while MedDINOv3 is both a representative foundation model and the semantic feature source for PSA. We will clarify this distinction and expand the TMJ-specific discussion.
[R2: Application settings] This pilot study tested the diagnostic value of predicted masks while limiting classifier effects. We used ImageNet-pretrained ResNet18 as a simple baseline, trained for 50 epochs with learning rate 6e-5 and batch size 16.Masks were concatenated with close/open MRI pairs as 4-channel inputs and improved AUROC. SOTA-oriented classifiers remain future work.
[R2: Typos] Thanks for catching these. We will fix them.
[R3: Failures] Thank you for the suggestion. During qualitative analysis, we observed failures, including over-segmentation into adjacent retrodiscal tissue when the posterior disc margin appeared intermixed and hard to distinguish. Given the tiny disc size, even small spill-over can affect metrics under severe anatomical ambiguity.
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.
There is clear consensus among reviewers that the work presents a well-designed and interpretable framework with clinical relevance. There is strong empirical performance demonstrated with improvements over baselines and clear presentation. However, all reviewers also acknowledge limitations in evaluation of the work, the fact that the evaluation is limited to a single dataset with no external validation making the generalization potential unclear and there is a lack of comparison with strong baselines.
In rebuttal, the authors should focus on strengthening the state of the art comparison by clarifying why the given baselines were chosen and comparing with more reject relevant approaches. If additional experimental results on external datasets exist, these should be reported and if not the authors should give a strong justification for limiting the work to the single dataset, acknowledging the limitations of this. The authors should also clarify the dependence on MOL and whether this can be consistently obtained given the apparent strong need for it. The contributions of the components should also be clarified given the apparently small contributions of C-MPR.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
N/A
- Please justify your recommendation.
N/A
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.
R1 kept the same recommendation of Accept, while R2 changed from Weak Accept to Accept. Both reviewers were satisfied with the rebuttal, in which the authors addressed initial concerns regarding the paper.
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.
TISC introduces an anatomically and clinically consistent framework for the challenging task of temporomandibular joint (TMJ) disc segmentation from MRI by coupling deep foundation feature anchoring with clinical metadata. The initial reviewer concerns regarding baseline rationale, missing external validation, and metadata dependency were comprehensively and convincingly resolved during the rebuttal phase. Specifically, the authors clarified that paper-level SOTA comparisons are structurally restricted due to the lack of a standardized public benchmark, while successfully demonstrating that their method degrades gracefully and maintains stable performance even if clinical metadata is completely absent.