Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Cervical vertebral maturation (CVM) staging guides orthodontic treatment and is conventionally formulated as a set of six discrete classes. In practice, skeletal development progresses gradually, and many lateral cephalometric radiographs exhibit transitional morphology between adjacent stages, resulting in well-documented inter-expert variability in CVM stage prediction. Prior deep learning work treats CVM as a standard multi-class classification problem despite its gradual nature, using exact-match accuracy that ignores clinically acceptable errors within one adjacent stage (Tol-1). We introduce the Ordinal Ranking of Cervical Vertebrae Maturation Using Deep Learning (ORACLE) framework and reformulate CVM assessment as a continuous learning task that models adjacent-stage relationships. A dataset comprising 1,248 lateral cephalograms was compiled with expert-assigned stage labels, of which 569 were additionally annotated with expert-verified polygon masks of the C2–C4 vertebrae. Across five stratified folds, the framework achieved 62.82% exact accuracy and 96.07% accuracy within Tol-1, confirming that adjacent-stage errors dominate, consistent with clinical expectations. On a separate set of 113 lateral cephalograms, we observed a similar trend of modest exact agreement but high adjacent-stage agreement with clinicians of varying experience. The ORACLE framework achieved quadratic weighted kappa scores of 0.88 and 0.79 against two senior experts, demonstrating performance comparable to inter-expert agreement (κ = 0.74). These findings demonstrate that CVM staging reflects a gradual developmental continuum not fully captured by rigid multi-class evaluation frameworks, supporting continuous-stage prediction with tolerance-based metrics as a clinically better-aligned framework for automated CVM assessment. The dataset is available via the data access form (https://forms.gle/C1VS7L6zeQo654UB7), and the source code is available at https://github.com/AI4Med-Lab/CVMI.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4854_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/AI4Med-Lab/CVMI

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MahKus_Cervical_MICCAI2026,
        author = { Mahajan, Kush AND Gupta, Dhruv AND Singh, Gagneet AND Soni, Vinay AND Jaiswal, Manojkumar AND Gupta, Sukrit},
        title = { { Cervical Vertebral Maturation Staging from a Continuous Perspective } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The main contribution of this paper is the introduction of the ORACLE framework (Ordinal RAnking of Cervical vertebrae maturation using deep LEarning) for cervical vertebral maturation (CVM) assessment. In addition, a dataset comprising 1,248 lateral cephalograms with expert-assigned stage labels is introduced.

  • 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 ORACLE framework reformulates CVM staging as a continuous learning problem that accounts for adjacent-stage relationships. This approach addresses the gradual nature of skeletal development and incorporates clinically acceptable tolerance (±1 stage, Tol-1) in the evaluation, which was previously ignored in traditional methods.

  • 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.Limited novelty in overall architecture composition：The paper adopts the existing U-Net++ framework to segment C2, C3, and C4, merely stacking different loss functions. For Stage Prediction, comparative experiments were conducted with different loss functions. From a continuous perspective, merely transforming the classification loss into a regression loss is insufficient. The approach does not fully address the inherent challenges and complexities associated with continuous prediction in cervical vertebral maturation. 2.The experiments lack comparisons with existing cervical vertebral maturation classification models[1,2,3,4,5]. [1] Zhang Y, Lu Z, Zhou J, et al. CDSNet: An automated method for assessing growth stages from various anatomical regions in lateral cephalograms based on deep learning[J]. Journal of the World Federation of Orthodontists, 2025, 14(3): 154-160. [2] Atici S, Pan H, Elnagar M H, et al. Classification of the cervical vertebrae maturation (cvm) stages using the tripod network[C]//ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023: 1-5. [3] Li H, Li H, Yuan L, et al. The psc-CVM assessment system: A three-stage type system for CVM assessment based on deep learning[J]. BMC Oral Health, 2023, 23(1): 557. [4] Wang P, Guan X, Wang A, et al. PCVMNet: Landmark-Aware Transformer Network for Pediatric Cervical Vertebral Maturation Assessment[C]//2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2025: 2879-2885. [5] Balaha H M, Alksas A, Fattal A, et al. Cervical vertebral maturation assessment using an innovative artificial intelligence-based imaging analysis system[J]. Biomedical Signal Processing and Control, 2025, 100: 107088. [6] Zhang F, Zheng L, Lin C, et al. Deeply learned cervical vertebrae maturation staging in ct images[C]//2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 2024: 1-8. 3.The paper does not include citations for key components such as U-Net++, ResNet-152, EfficientNet, ConvNeXt, and the AdamW optimizer. These are important references that should be included to properly acknowledge the foundational work. 4.Minor grammatical errors: Although the tolerances across models are similar, but the ConvNeXt backbone gives the highest accuracy and F1 scores. Keywords: The capitalization is inconsistent. 5.Section 4.2 uses U-Net++ with an EfficientNetB5 encoder pretrained on ImageNet, while the overview mentions a U-Net++ encoder–decoder to segment cervical vertebrae.

  • Please rate the clarity and organization of this paper

    Satisfactory

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

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

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

    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 manuscript addresses an important problem in cervical vertebral maturation assessment; however, it suffers from several significant limitations.The proposed approach relies on an existing U-Net++ framework with only incremental modifications through loss function combinations, which do not adequately tackle the challenges of continuous stage prediction. The manuscript also omits essential citations for widely used backbone networks and optimization techniques. Furthermore, there are minor grammatical issues and inconsistencies in methodological descriptions that affect clarity and presentation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    1.The claim that no prior CVM work has applied any continuous formulation appears too strong. QCVM and related works introduce formulations that can be interpreted as continuous representations, which provide a basis for subsequent continuous modeling. [1] Chen L L, Xu T M, Jiang J H, et al. Quantitative cervical vertebral maturation assessment in adolescents with normal occlusion: a mixed longitudinal study[J]. American Journal of Orthodontics and Dentofacial Orthopedics, 2008, 134(6): 720. [2] Fulin J, Yan Y, Fangyuan C, et al. Deep Learning Based Quantitative Cervical Vertebral Maturation Analysis[J]. 2024.2.From a continuous perspective, simply reformulating the classification objective as a regression loss is insufficient. Although the model achieves a high To1-1 accuracy of 95% under a tolerant evaluation criterion that treats predictions within one adjacent stage as correct, the strict overall accuracy remains around 62%, which indicates limited performance in exact stage prediction and may constrain its clinical applicability. 3.From a methodological perspective, the current approach lacks sufficient novelty and still falls short of the standards typically expected for MICCAI-level submissions.



Review #2

  • Please describe the contribution of the paper

    The paper proposes the ORACLE framework, which models CVM staging as a continuous/ordinal problem instead of standard multi-class classification. It combines vertebral segmentation with different learning formulations (EMD, CORAL, regression) and evaluates using tolerance-based metrics and inter-expert agreement

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper tackles a clinically meaningful problem (CVM staging) and highlights the limitation of treating it as strict classification. 2.Reformulating the task as ordinal/continuous learning is reasonable and better aligned with clinical practice. The inclusion of tolerance-based metrics and comparison with inter-expert variability is a nice touch and improves clinical relevance. 3.The inclusion of tolerance-based metrics and comparison with inter-expert variability is a nice touch and improves clinical 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.

    1.The introduction is not well structured, which makes it hard to clearly understand what the actual contribution is versus standard ordinal learning ideas. 2.The paper argues that continuous modeling is “clinically superior,” but the evidence is mostly indirect (e.g., Tol-1 and agreement metrics) and does not fully justify such a strong 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

    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?

    Overall, the paper addresses a relevant clinical problem and proposes a reasonable reformulation that better aligns with the nature of CVM staging. Although the technical contribution is somewhat incremental and the writing could be clearer, the idea is sound and the evaluation (especially with tolerance-based metrics and expert comparison) adds value. With improvements in clarity and stronger positioning, the work could be a useful contribution.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    N/A



Review #3

  • Please describe the contribution of the paper

    I think the main contribution is the clinical variability presented.

  • 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 analysis and comparison of the model performance with the raters is comprehensive and thorough.

  • 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.

    I would say the paper requires clarity and a stronger motivation for the biological representation of the workflow. They argue a lot that these methods better represent the biological framework (which I agree) but there is no acknowledgement about the fact that currently doctors also are confined to a discrete scale for classification. I think they can strengthen the wording to convey this message. They also should compare numerically to existing segmentation methods and clinician variability they reference in their work.

  • 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?

    Code was published in an anonymous link. Data was not in that link (I could not see it? maybe I missed it) so I cannot confirm if this is ethical.

  • 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

    1.Clarity of problem framing and clinical context What is the reference for the statement that agreement decreases between adjacent stages? How does this compare to previously reported variability (refs 3–6)? What level of disagreement (e.g., ±1 stage) is clinically acceptable? you need to reference this sentence. Why model the task as continuous when clinical assessment is categorical? What advantage does this provide in practice? Make this more convincing as it is a main improvement of your work to others. Also please numerical compare with other wrok. How is disease progression defined from the segmentation masks based on volume alone or shape/image features within or another metric? it is not clear in the intro. 2.Methodological justification What is the rationale for using Earth Mover’s Distance (EMD) loss? in the EMD equation: Why is there a squared term and a factor of 2? Is the loss computed per class and then summed?This is not clearly expressed. Why do you expect ordinal classification improve performance? What exactly is meant by “ordinal thresholds”? Why include a background channel in vertebral segmentation? how is it used downstream? 4.Results and reporting clarity What does “unannotated” mean in Table 1? Why focus on extraction details in Table 1 rather than agreement/performance metrics? Can you report and compare the exact values of competing methods more explicitly? (DSC, iou, h95) Could other methods segment better but classify worse - this is not discussed at all. Can you report learned model parameters of the weighted coefficients to see which loss its actually relying on. 5.Metrics and evaluation definitions What does “100% expert verification” mean? Full agreement between experts? How exactly are tol1 and tol2 calculated - are they accuracy within ±1 or ±2 stages? What is the precise definition of “proportion” here?

  • 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 think this paper does a good job at evaluating the current gaps in the literature and addressing them. I think there are improvements to be made in the communication of the work. I am not overly convinced about the novelty of the machine learning pipeline.

  • 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.

    Their responses made me less convinced of an accept.

    From their responses I don’t agree on the following: 2.Bench marking they say confidently could not be done. I think this is incorrect as they could have ran other models (ex for segmentation) 3.’‘clinically better aligned’’ I still think is incorrect. They are not replicated the clinical workflow, they are trying to simulate a biological growth which the clinicians do not do (from my understanding they bin patients into class) so they are make a regression problem to better simulate the biology NOT the clinical workflow. They would have to convince clinicians to change their practice to a regression workflow, which would make this a CAI paper. 5.Technical response says ‘‘clinical consultation was used’’ this is confusing and should be described in the main work. It also does not seem like a clinical standard to compare to, so this should have its own reported results so we can fully understand how these numbers were generated and how they may differ?



Author Feedback

We thank the reviewers for the constructive feedback. We address key concerns below. (1) Novelty: The primary contribution is not a new segmentation architecture or ordinal loss, but a clinically motivated reformulation with tolerance-aware evaluation and inter-expert contextualization. R1’s concern that loss substitution is incremental is valid: EMD, CORAL, and MSE are established tools. However, we are constrained since clinical settings generate discrete CVM stages which is an established common practice. Furthermore, no prior CVM work has applied any continuous formulation (Table 1, ‘Continuous?’ column, all ‘No’). We pair this with a clinically-grounded evaluation framework (Tol-1/Tol-2, inter-expert agreement) that exposes what exact-match accuracy conceals: adjacent-stage errors dominate and match clinician variability patterns. The proposed framework bridges the lack of metrics to assess the continuous modeling of CVM staging. (2) Benchmarking: No public benchmark exists, each prior method uses a private dataset, different annotation setup, and often a collapsed staging scheme (some reduce six stages to three, inflating accuracy), making cross-paper quantitative comparison hard. Specifically: Zhang et al. (CDSNet) and Zhang F. et al. use hand-wrist and 3D CT modalities incompatible with our 2D LC setting; Li et al. (PSC-CVM) requires object-detection masks and contour keypoints; Wang et al. (PCVMNet), Atici et al. (TRIPOD), and Balaha et al. depend on landmarks, age features, or handcrafted markers absent from our dataset. We will extend Table 1 with R1’s cited methods to make ORACLE’s positioning explicit. Further, Refs. 3–6 report clinician variability without automated models and serve as motivating evidence, not competing baselines. (3) Clinically superior (R2, R3, MR): We accept this phrasing is too strong and will revise it to ‘clinically better-aligned.’ Clinical reporting will remain categorical, but the underlying biological process is continuous, empirically reflected in expert disagreement concentrating at adjacent stages (Refs. 3–6; inter-senior κ = 0.74 in our study). A model trained with ordinal loss penalises distant-stage errors more heavily, whereas cross-entropy treats a 1-stage and a 3-stage error identically. Performance evidence supports the alignment claim: CORAL outperforms CE by +3.5% accuracy, with model–expert κ matching inter-senior κ. Prospective clinical trials remain future work, as we acknowledge. (4) R3 ethics query: Code was shared with reviewers; we will share the dataset via CMT upon request. (5) Minor Comments: We will (i) restructure the Introduction to highlight the three-gap motivation; (ii) add missing citations (U-Net++, ResNet-152, EfficientNet-B5, ConvNeXt-Base, AdamW); (iii) fix the grammatical error in 4.3 and align the EfficientNet-B5 description across 4.2 and Fig. 1; (iv) replace ‘clinically superior’ throughout. On R3’s technical questions: EMD computes squared ℓ₂ distance between predicted and target CDFs over K−1 cumulative bins, penalising larger ordinal gaps quadratically; CORAL outputs K−1 binary thresholds predicting stage > k, with class probabilities from consecutive sigmoid differences; the background segmentation class reduces false positives: only classes 1–3 feed downstream classification. Tol-k = |predicted − ground truth| ≤ k; “100% expert verification” means all 569 masks were reviewed and corrected by two senior orthodontists; “unannotated” means stage-labelled without pixel-level masks; the ±1 threshold was confirmed through clinician consultation; morphology is learned from pixel-level supervision with no handcrafted features. Learnt segmentation weights: Lovász 43.1%, FocalBoundary 28.6%, Tversky 27.1%. The final manuscript will present the first CVM framework grounded in continuous modeling, expert-verified annotation, and inter-expert-calibrated evaluation.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

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

    This paper proposes the ORACLE framework, which reformulates cervical vertebral maturation (CVM) staging as an ordinal/continuous learning problem, integrating vertebral segmentation with alternative loss formulations (e.g., EMD, CORAL, regression) and evaluating performance using tolerance-based metrics and inter-expert agreement. Reviewers generally agree that the problem is clinically relevant and that moving beyond strict categorical classification toward ordinal modeling is conceptually meaningful and better aligned with the gradual nature of skeletal development (R1,R2). The inclusion of tolerance-based evaluation and comparison with expert variability is also viewed as a strength that improves clinical relevance (R1,R2,R3).

    However, there are several notable concerns. A primary issue raised by multiple reviewers is the limited methodological novelty. The framework relies heavily on established components (U-Net++ for segmentation), with the main variation being different loss formulations, which are seen as incremental rather than fundamentally novel (R1,R2,R3). In addition, the experimental section lacks important comparisons to prior CVM-specific methods, making it difficult to contextualise the performance gains against the current state of the art (R1, R3). Clarity and positioning are also recurring weaknesses. The introduction and overall framing do not clearly distinguish the paper’s contributions from existing ordinal learning approaches (R2), and several methodological details (e.g., loss definitions, parameterisation, and metric formulations) are insufficiently explained (R3). There are also missing citations for standard architectures and minor inconsistencies in the description of the pipeline (R1). Furthermore, the claim that continuous modelling is “clinically superior” is not fully substantiated with direct evidence, and would benefit from a more careful and nuanced justification (R2, R3).

    Overall, while the paper addresses an important problem and proposes a reasonable reformulation with clinically motivated evaluation, the technical contribution is viewed as incremental and the experimental validation and clarity need strengthening. Based on the balance of reviews (R1: Weak Reject; R2: Weak Accept; R3: Accept), I would recommend a rebuttal.

  • 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 final review recommendations are two Accepts and one Reject. However, both R1 and R3 highlighted several important weaknesses that should be addressed. In particular, the main concern centres on the paper’s continuous formulation, where the classification problem is reframed using a regression loss. R1 questioned the novelty of this idea, noting that prior works had already explored continuous representations. Similarly, R3 was unconvinced by the claim of better clinical alignment, arguing that the method models biological growth rather than reflecting the actual categorical clinical workflow, while also raising concerns about benchmarking and clarity of the clinical evaluation process.



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.

    Author have addressed several of the issues in the rebuttal, and even though the contributions seems incremental, it tackles a relevant clinical problem, the concept is sound and the evaluation is well conceptualized. With improvements in clarity and stronger positioning, the work could be a useful contribution.



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 methodological contribution is incremental—combining established components without a novel technical insight. The claim of “clinical superiority” is not substantiated, and the manuscript’s framing and detail need strengthening. These fundamental gaps outweigh the clinically relevant motivation and the strength of tolerance-based evaluation.

    Deviation from reviewers’ recommendation: While the reviews were mixed (Weak Reject, Weak Accept, Accept (switched to nagative after rebuttal)), I am inclined to reject it in its current form



back to top