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

Abstract

Lumbar spine degeneration is a major contributor to chronic low back pain and is routinely assessed on MRI using ordinal grading systems, e.g. normal, mild, moderate, severe. Consequently, most approaches to train models to grade these MRIs formulate grading as a multi-class classification problem, treating ordinal grades as categorical, ignoring differences in misclassification severity, and imposing hard decision boundaries on a continuous disease process. This work explores modeling spinal degeneration as a continuous severity ranking problem. We introduce SpineNetRank, a framework that learns scalar severity scores from lumbar spinal MRI, and compare it against multi-class classification and ordinal regression. Using multiple degeneration measures from the Genodisc dataset, we show that a model trained using a ranking loss to produce a continuous score enables fine-grained ordering of MRI scans. Furthermore, the ordinal grading classes can be recovered from the score with comparable accuracy to those from a model trained directly for classification. The continuous score even improves discrimination between more distant classes.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5388_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/spinetools/spineranknet

Link to the Dataset(s)

N/A

BibTex

@InProceedings{MonMar_Be_MICCAI2026,
        author = { Monzon, Maria AND Zisserman, Andrew AND Park, Robin Y. AND Jutzeler, Catherine R. AND Jamaludin, Amir},
        title = { { Be Indiscrete: The Benefits of Learning Continuous Spine Degeneration Severity Scores } },
        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

    This paper proposes SpineRankNet, a multi-task framework for lumbar spine MRI grading that learns a continuous severity score for each pathology instead of treating ordinal grades as independent classes. The method formulates spine degeneration grading as a pairwise ranking problem with a severity-aware margin, distance-weighted pair loss, and same-grade similarity regularization, then recovers discrete grades from the learned continuous scores using task-specific thresholds. The paper evaluates classification, ordinal regression, and ranking objectives across 11 concurrent spinal grading tasks on the multicenter Genodisc dataset and shows that ranking-based continuous scores can preserve classification-level performance while improving ordinal agreement and reducing large-grade errors

  • 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 addresses a real limitation of current spine grading models: degenerative severity is biologically continuous, but is usually modeled as discrete categorical classes. This is especially relevant in radiology, where borderline cases and inter-rater ambiguity are common. 2) Reformulating spinal grading as a ranking problem rather than nominal classification is a sensible and clinically grounded idea. The proposed severity-aware pairwise loss, which explicitly weights larger grade differences more strongly, aligns well with the practical importance of avoiding large clinical errors. 3) The paper studies 11 grading tasks on a multicenter cohort with heterogeneous acquisition settings and severe class imbalance, which gives the work more breadth than many spine-grading papers. 4) Table 1 shows that SpineRankNet improves the most ordinally relevant metrics while maintaining ROC-AUC comparable to cross-entropy classification. 5) The model produces both a continuous score and a discrete radiological grade. This is a useful design because it preserves compatibility with current clinical grading workflows while adding finer-grained ordering information. 6) Figure 2 on page 8 clearly illustrates the intended benefit of continuous severity scoring by showing progressive score changes within and across ordinal grades for several pathologies.

  • 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 the loss description, the pair direction term r appears to be defined using predicted score differences, even though it is used as a supervision signal. This is likely just a notation error and presumably should come from the ground-truth grade ordering, but it affects the clarity of the core method and should be corrected. 2) The figure on page 3 suggests one data flow for the rank head, while the text describes the ranking head as a two-layer MLP operating on the shared embedding. This is probably resolvable, but the final model definition should be made fully consistent. 3) The paper includes CE, CORN, direct MAE regression, RankNet, and DeepRankSVM, which is a decent set, but it omits some stronger and more recent distance-aware ordinal losses and rank-consistent methods. 4) Since the paper’s main claim is about better severity ordering and fine-grained ranking, it would be stronger to include explicit ranking metrics such as Spearman correlation or pairwise ordering accuracy, rather than relying only on discretized classification-style metrics. 5) All experiments are performed on a single dataset, even though it is multicenter. The paper itself acknowledges this in the conclusion. For a paper making a general argument about continuous severity learning, external validation would strengthen the claim considerably. 6) The reported results are aggregated over tasks from what appears to be a single training run, without multiple random seeds or significance testing. This does not invalidate the paper, but it leaves some uncertainty about robustness. 7) The paper performs a fairly careful threshold search for the proposed continuous scores. It would be helpful to clarify whether comparable calibration or threshold optimization was also allowed for the classification/ordinal baselines, to avoid any concern about asymmetry in post-processing. 8) There are several typographical and formatting issues, such as inconsistent naming and a few awkward fragments, that should be cleaned up in revision.

  • 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

    This is a strong and clinically well-motivated paper. I particularly appreciate that the authors do not simply claim continuous severity is “better,” but actually compare classification, ordinal regression, and ranking objectives in a unified setting across many spine pathology tasks. Overall, I find the contribution meaningful and the empirical evidence fairly convincing.

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

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The proposed continuous-ranking view of spine degeneration is well-motivated, the experiments are broader than those in typical single-task grading papers, and the reported gains in ordinal metrics directly support the central claim that ranking reduces large-grade errors while preserving classification-level performance. My enthusiasm is tempered by several issues that should be addressed: a likely ambiguity in the definition of loss, some inconsistencies in the method description, the absence of stronger ordinal baselines, and the absence of external validation or ranking-specific evaluation metrics. These are important, but they feel more like correctable weaknesses than fundamental flaws

  • 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

    The authors propose a framework for continuous degeneration scoring of the lumbar spine from MRI that is done simultaneously for multiple conditions. The framework is compared against classification, ordinal regression, and pairwise ranking-based baselines, showing favorable overall performance and an arguably lower rate of distant grade misclassifications.

  • 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) Well-written paper, with extensive and up-to-date coverage of the related literature and a sufficient amount of technical detail. 2) Rigorous technical execution of the study, with strong aspects including the pair sampling procedure, test-time augmentation, selection of metrics, diverse set of baseline methods, reporting of confidence intervals, and experiments on ranking layer selection and thresholding approaches. 3) Learning continous grades from the existing ordinal grading schemes is a promising idea that can unlock new insights into chronic disease onset and progression (which in the chronic MSK conditions are predominantly heterogeneous). While a few prior articles exist related to this topic (e.g., in the osteoarthritis domain, as appropriately cited by the authors), the overall approach of learning a continuous grading scale from real-world ordinal labels is rather new. The extension of this approach to multiple conditions, proposed by the authors, is a clear novelty.

  • 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 study introduces several hyperparameters (lambda_dis, lambda_sim, m_base, m_scale, beta), yet does not provide any sensitivity analysis regarding the choice of their values. However, based on the meaning of the hyperparameters, the chosen values are sensible. 2) The choice of the backbone is suboptimal. Numerous studies have demonstrated that ResNet-34 or -50 often provide performance improvements over ResNet-18, particularly with volume sizes similar to those used (12825612). However, using a more powerful backbone would likely not change the conclusions. 3) The detailed analysis of the results (e.g., performance across different grading tasks) and the clinical interpretation of the learned continuous grades fall rather short. It is understandable, given the page limit, and the authors provide relevant directions for future 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 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

    1) While the related work section is comprehensive, please consider additionally discussing these two well-known papers that approach the presented idea - https://doi.org/10.1097/RLI.0000000000000709, https://doi.org/10.1002/jmri.27400 . 2) Abstract: “The continuous score even improves discrimination between more distant classes.”. This sentence reads as if the continuous nature of the score causes the reduction of such errors. Based on the results, however, it is more likely that the ranking property of the loss function itself is the driver of this behavior (see also the results of the CORN method). Consider rephrasing this sentence. 3) Page 7: “Notably, its lowest MAE demonstrates systematic reduction of large-grade prediction errors”. While lower MAE and higher QWK suggest such a reduction, they do not directly demonstrate it. Such a conclusion would be more appropriate if confusion matrices were presented. Please, rephrase. 4) Page 8: “Figure 2 further shows that the learned continuous severity scores capture intra-grade variability that categorical models cannot represent”. Currently, this reads more as speculation. Consider rephrasing the sentence and/or extending the interpretation of Figure 2.5) Page 4, last line: The reference is broken and shows as “[1?]”. Please, fix. 6) Page 6, 2nd paragraph: “distributionto” -> “distribution to”. 7) Page 8, 2nd paragraph (“Importantly, …”): This sentence duplicates the conclusions from Page 7 (2nd paragraph) and, moreover, contains several issues. Consider removing it.

  • 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?

    +methodological novelty ++excellent design and technical execution of the study; great writing -short interpretation of the results and discussion

  • 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.

    No changes to my original evaluation. The paper is very comprehensive and deep given the page limit. Regarding the statements of other Reviewers, I generally agree with Reviewer 1.However, I consider the evaluation of Reviewer 3 to be somewhat too negative, downplaying the original contributions of the work.



Review #3

  • Please describe the contribution of the paper

    The manuscript investigates spinal degeneration as a continuous severity ranking problem and introduces SpineNetRank, a framework that learns scalar severity scores from lumbar spinal MRI. The proposed approach is compared against multi-class classification and ordinal regression methods.

  • 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 key strength of the paper is the formulation of severity estimation as a continuous ranking problem rather than independent multi-class or ordinal classification, which offers an important perspective for modeling disease progression.

  • 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.

    However, the manuscript has several limitations:

    • The evaluation is limited to a single dataset. Validation on additional datasets (e.g., OAI or MOST for knee osteoarthritis, or other medical applications such as brain age prediction or cancer severity) would strengthen the generalizability and robustness of the proposed approach.
    • The paper lacks an ablation study. The authors should justify the choice of the ResNet-18 backbone and provide analysis to support this design decision.
    • Cross-entropy–based classification appears to outperform the proposed method in terms of ROC-AUC. This result should be discussed in detail, as it appears to conflict with the main motivation of the study.
  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents an interesting formulation of severity estimation as a continuous ranking problem. However, the contribution is limited by evaluation on a single dataset, lack of ablation studies, and insufficient justification of design choices. Additionally, the fact that standard classification methods outperform the proposed approach in ROC-AUC raises concerns about its practical advantage.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all reviewers for their thorough engagement and insightful review feedback. We would try to address each point directly below. R1.1 Loss notation Thank you for pointing it out. We will fix the notation error regarding rij to be defined from the ground-truth grade ordering, not predicted scores. R1.2 Figure/text inconsistency The Section 2 description is accurate, but for clarity Fig1 shows only one representative task head, rather than all 11 pathology-specific heads. We will clarify it in the caption and standardize SpineRankNet naming. R1.3 Missing ordinal losses Our selection spans three main objective families: categorical, ordinal , and pairwise ranking, using identical encoders for controlled comparison, representing state-of-the-art properties (CORN) and distance-aware margins[1]. Ordinal losses are usually not designed for continuous scores, making them unsuitable for our setting. R1.4 Missing ranking metrics. We agree explicit ranking metrics would be beneficial but most of the methods do not produce continuous scores and therefore cannot be evaluated fairly. However, QWK penalizes disagreements proportionally to their ordinal distance, and MAE quantifies the magnitude of grade errors. We chose these together with ROC-AUC and BA to enable fair comparison across all benchmarked methods. That said, pairwise ordering accuracy of our method could be reported. R1.5 & R3.1 Single dataset We appreciate this concern and acknowledge it as a limitation. The field is data-scarce at this annotation granularity. No other dataset provides expert grading of all 11 lumbar spine pathologies, which is precisely why prior works (SpineNet) also rely exclusively on Genodisc. Genodisc is still diverse, with ∼2K subjects from 5 centers with diverse MRI acquisitions. Future work could evaluate on RSNALumbarDISC, though it is limited to stenosis pathologies. External validation is important future work, although OAI/MOST are not directly comparable. R1.6 Single training run The mean ± std reported in Table 1 across tasks reflects task-level heterogeneity, but we acknowledge this does not replace seed-level variability. We will try to incorporate a more extensive multi-seed analysis in the camera-ready version.

R1.7 Threshold calibration asymmetry. The threshold search for SpineRankNet is purely a discretization step to recover discrete grades from continuous scores ONLY for evaluation, so there is no asymmetric post-processing advantage. R1.8 & R2 Typographical issues. Typographical issues. Thank you for catching these and we will carefully correct each issue. R2.1 Hyperparameters Their values were guided by literature[1] and parameter semantics mbase​ and mscale control distance-aware margins, and C2 weights within-grade clustering for calibration. A sensitivity analysis in the camera-ready will strengthen the paper. R2.2 and R3.2 Choice of backbone We agree that a dedicated backbone study would be useful. Due to space limits, we focused on ablations most relevant to our claim: ranking-head design and thresholding. ResNet-18 reflects a practical performance-memory tradeoff for 3D volumes across 11 tasks, and prior work* suggests limited gains from larger backbones in lumbar grading. As R2 noted, a stronger backbone would likely improve all methods similarly without changing the main conclusion. We noted further backbone study as future work. *10.1109/ISBI53787.2023.10230814 R2.3 Detailed Results Per Grading Task Due to space limits, results were aggregated across 11 tasks. We appreciate the reviewer’s understanding of the strict page limits. R3.3 ROC-AUC We respectfully clarify that this does not conflict with our motivation. ROC-AUC measures discrimination (optimized by CE). Our contribution is to show that ranking preserves comparable discrimination while simultaneously improving ordinal agreement and reducing large-grade errors, which are the clinically more consequential failures (as noted by Rev2).



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.

    Authors should meticulously address reviewers’ concerns, particularly the concerns about the experiment settings. Specifically, authours should address the issue about single dataset (even it is multicenter) and insufficient justification of ablations, backbones and hyperparameters, and missing ranking metrics, etc.

    Authors should also address other confusions about loss formulations, architecture descriptions, etc.

  • 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 reviewers acknowledge the strong motivation, solid methodology, and comprehensive evaluation, supporting the proposed continuous ranking framework as a meaningful contribution. After rebuttal, although concerns about limited validation remain, clarifaction issues have been addressed. Overall, the strengths and potential impact outweigh the weaknesses.



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.

    After considering the reviews and the rebuttal, I recommend acceptance. The paper presents a clinically meaningful formulation of spine degeneration grading as continuous severity learning, while still allowing recovery of discrete grades. The rebuttal clarifies the loss notation, model description, thresholding procedure, dataset scope, and the interpretation of ROC-AUC versus ordinal metrics. The overall contribution and empirical evidence are sufficient. I therefore recommend acceptance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Reviewers’ final decisions after the rebuttal are missing. However, the paper presents a novel idea with clear clinical motivation. The authors also promise to release the dataset and code, which is valuable for the research community. Therefore, despite the missing post-rebuttal decisions, I see potential value in the paper.



back to top