Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Be Indiscrete: The Benefits of Learning Continuous Spine Degeneration Severity Scores

T0 review · 2 major / 4 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Training MRI models to rank spine degeneration continuously recovers clinical grades while cutting large mistakes and capturing within-grade nuance.

desk verdict Solid multi-pathology ranking application that improves ordinal metrics on Genodisc without inventing a new paradigm; single-cohort transfer is the real soft spot. read the letter →

arxiv 2607.05090 v1 pith:6ILUBVVE submitted 2026-07-06 cs.CV

Maria Monzon , Andrew Zisserman , Robin Y. Park , Catherine R. Jutzeler , Amir Jamaludin This is my paper · ORCID
classification cs.CV
keywords pairwiserankingordinalregressionspinedegenerationlumbarMRIcontinuousseverityscoreRankNetPfirrmanngradingclassimbalance
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Spinal degeneration on MRI is a continuous biological process, yet clinical practice and most AI systems force it into a few ordered bins such as normal, mild, moderate, or severe. Treating those bins as unrelated categories penalizes every misclassification equally, discards subtle differences inside a grade, and struggles with rare severe cases. This paper shows that learning a continuous severity score by pairwise ranking recovers the original ordinal grades with accuracy comparable to direct multi-class classification, while improving agreement with the ordered scale and reducing clinically distant errors. The continuous score also separates borderline and within-grade cases that discrete labels cannot distinguish. A sympathetic reader cares because a finer, ordered measure can support earlier detection of progression, fairer handling of imbalance, and less brittle decisions at grade boundaries.

What carries the argument

SpineRankNet: a shared 3D ResNet-18 encoder plus task-specific ranking heads trained with a severity-weighted hinge loss on grade-difference pairs (larger margins and weights for distant grades) and a same-grade similarity penalty, followed by validation-set thresholds that recover discrete grades.

What would settle it

On a fully independent external lumbar MRI cohort with the same eleven grading tasks, check whether continuous scores from SpineRankNet still recover discrete grades at classification-level balanced accuracy while keeping higher quadratic-weighted kappa and lower mean absolute error than a matched cross-entropy classifier; failure on either metric would falsify the claimed benefit.

Watch

Extended reading notes

Core claim

Across eleven concurrent grading tasks on the multi-centre Genodisc lumbar MRI cohort, a model trained with a severity-aware pairwise ranking loss produces continuous scores in [0, 10] that, once thresholded, match direct classification accuracy while delivering higher quadratic-weighted kappa, lower mean absolute error, and better discrimination of distant grades than classification or ordinal-regression baselines.

Load-bearing premise

The relative orderings taken from existing discrete expert grades, plus thresholds tuned on the same cohort’s validation set, are a faithful enough proxy for the true continuous disease process that the ranking gains will hold outside this single dataset.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes SpineRankNet, a multi-task framework that learns continuous severity scores s∈[0,10] for 11 concurrent lumbar-spine MRI grading tasks on the multi-centre Genodisc cohort. A shared 3D ResNet-18 encoder is first pretrained with inverse-frequency-weighted cross-entropy; lightweight ranking heads are then fine-tuned with a severity-aware pairwise hinge loss (difference-specific weights and adaptive margins) plus a same-grade similarity term. At inference the continuous scores are discretized by validation-set grid-searched thresholds. On a held-out test set the ranking model recovers discrete grades with balanced accuracy comparable to direct multi-class CE (63.3 % vs 64.1 %) while improving ordinal concordance (QWK 0.76 vs 0.69, MAE 0.14 vs 0.18, MCC 0.59 vs 0.54) and outperforming CORN, MAE regression, RankNet and DeepRankSVM under an identical encoder and protocol. Qualitative examples illustrate within-grade ordering that categorical labels cannot capture.

Significance. If the reported gains hold under external validation, the work supplies a practical alternative to categorical cross-entropy for highly imbalanced ordinal medical grading: continuous scores that preserve severity ordering, reduce large-grade errors, and remain convertible to clinical grades. The multi-pathology evaluation (11 concurrent tasks), severity-adaptive ranking loss, and public code release are concrete strengths that distinguish the contribution from prior single-pathology ranking or pure ordinal-regression studies. The clinical utility of a continuous score for longitudinal monitoring and borderline-case annotation is clearly motivated, even though it remains untested here.

major comments (2)
  1. Table 1 reports mean±std over the 11 tasks from a single training run; no multi-seed statistics or paired significance tests are provided. Because the absolute BA gap versus CE is only 0.8 points and several ranking variants are close, it is unclear whether the ordinal improvements (QWK, MAE, MCC) are robust. At minimum, multi-seed means with confidence intervals or a non-parametric test across tasks should be added so that the central claim of consistent ordinal superiority can be evaluated.
  2. All quantitative claims rest on a single multi-centre cohort (Genodisc) with patient-level but not fully independent external splits. The Conclusion correctly flags the need for external validation and longitudinal evaluation; without at least one independent test set the transferability of the continuous scores (and of the validation-tuned thresholds τ) remains an untested assumption that is load-bearing for the claimed clinical benefit.
minor comments (4)
  1. Section 2.1, Eq. (3)–(6): the free parameters m_base, m_scale, λ_sim/λ_dis, β and the 70/30 hard-mining mix are stated without sensitivity analysis; a short ablation or justification would strengthen reproducibility.
  2. Figure 2 caption and surrounding text contain minor typos (“Pfirrmnan”, “Low.End. Mod.”) and the score ranges shown for identical grades would benefit from explicit numerical annotation.
  3. Table 1 footnote: ROC-AUC for CORN is marked unavailable; a brief explanation (or an alternative multi-class AUC) would avoid the impression of selective reporting.
  4. The two-phase training schedule (CE pre-training then ranking fine-tuning) is described clearly, yet the precise learning-rate schedule and early-stopping criteria for the ranking phase could be moved from the text into a short supplementary table for easier replication.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; mild self-citation of authors' prior SpineNet/Genodisc work for dataset and multi-pathology baselines, while ranking claims rest on independent held-out evaluation.

  1. self citation load bearing [Section 3 Dataset; Related Work; Acknowledgements]
    "The Genodisc dataset [7] includes T1w and T2w sagittal lumbar MRIs from ∼2,000 subjects … IVD volumes … are extracted via SpineNetv2 [25] vertebra localization … Among these, SpineNet [7,25] and the Context-Aware Spinal Transformer [24,26] are notable exceptions …"

    Genodisc and the SpineNet localization/grading pipeline are introduced via citations whose author lists overlap with the present paper (Jamaludin, Zisserman). This supplies the data and multi-task baseline architecture, yet the central ranking-versus-classification comparison is still evaluated independently on a held-out split; the self-citation is therefore present but not load-bearing for the claimed QWK/MAE improvements.

full rationale

This is a standard empirical supervised-learning paper. Continuous scores s(t) are produced by a severity-aware pairwise ranking loss whose relative labels r_ij are derived from independent expert ordinal grades y_t already present in Genodisc; the same discrete labels supervise the CE and CORN baselines. Discretization thresholds τ are grid-searched only on the validation split and then frozen for test-set metrics (BA, QWK, MAE, MCC). Nothing in the loss (Eqs. 1–3), soft-clamp parameterization (Eq. 6), or two-phase training protocol forces the reported QWK/MAE gains by construction; those gains are measured on a held-out patient-level test partition. Self-citations to SpineNet/SpineNetV2 supply the multi-pathology dataset, IVD localization, and concurrent-grading baselines, but do not supply a uniqueness theorem, ansatz, or fitted parameter that is later re-presented as a prediction. The paper itself flags the single-cohort limitation and calls for external validation. Consequently the derivation chain is self-contained against the reported benchmarks and exhibits only the mild, non-load-bearing self-citation typical of incremental medical-imaging work.

Assumptions & free parameters 7 free parameters · 4 assumptions · 1 invented entities

Central claims rest on standard supervised learning assumptions plus several hand-chosen ranking hyperparameters and the premise that discrete expert grades induce a usable continuous severity axis. No new physical entities are postulated; SpineRankNet is an architectural/loss design. Free parameters are the usual ML knobs that affect margins and score scale; axioms are domain and methodological assumptions, not ad-hoc physics.

free parameters (7)
  • m_base (severity-aware margin base) = 0.5
    Set to 0.5 by hand; controls minimum score separation for adjacent grades and directly shapes the ranking loss.
  • m_scale (severity-aware margin scale) = 1.5
    Set to 1.5 by hand; scales margin with grade distance and is load-bearing for the claim that distant pairs are more strongly separated.
  • λ_sim / λ_dis (loss trade-off) = 1.0 / 0.25
    λ_dis=1, λ_sim=0.25 chosen to balance margin enlargement vs same-grade compactness; affects within-grade score collapse.
  • S (score range upper bound) = 10
    Fixed to 10 for all tasks to align with clinical scales; defines the continuous score domain and weight normalization.
  • β (soft-clamp sharpness) = 5
    β=5 in the softplus score parameterization; controls boundary sharpness of s∈[0,S].
  • grade thresholds τ (per task) = validation grid search (coordinate descent)
    K−1 ordered thresholds grid-searched on validation to recover discrete grades from continuous scores; recovery accuracy depends on these fits.
  • pair mining mix (70% hard / 30% uniform) = 70/30
    Hand-chosen sampling of batch pairs by violation magnitude vs uniform; affects which ranking constraints dominate training.
assumptions (4)
  • domain assumption Expert ordinal grades on Genodisc are a valid partial order for constructing pairwise severity supervision r_ij.
    Entire ranking loss is built from signed grade differences; if labels are noisy or non-monotonic with biology, continuous scores inherit that structure (§2.1, Dataset).
  • domain assumption Spinal degeneration is a continuous progressive process that discrete bins under-resolve.
    Stated in Introduction as motivation for continuous scores; not independently measured in this paper.
  • standard math Standard pairwise ranking SVM / hinge and relative-attribute losses are appropriate objectives for ordinal medical severity.
    Method extends ranking SVM and relative attributes (cited [17,1,2]); mathematical form is conventional.
  • domain assumption Patient-level 80/10/10 split stratified by centre and grade yields an unbiased estimate of generalization within Genodisc.
    All metrics in Table 1 depend on this split; no external cohort is used.
invented entities (1)
  • SpineRankNet continuous severity score s_t ∈ [0,10]
    purpose: Scalar per-pathology output intended to preserve ordinal structure and within-grade variability for lumbar MRI.
    Learned representation, not a new biological quantity; independent clinical meaning is only suggested (longitudinal sensitivity) and not validated outside this training setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Be Indiscrete: The Benefits of Learning Continuous Spine Degeneration Severity Scores." pith.science (2026). https://pith.science/paper/6ILUBVVE

@misc{pith2026260705090,
  author       = {Pith},
  title        = {Pith review of: Be Indiscrete: The Benefits of Learning Continuous Spine Degeneration Severity Scores},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ILUBVVE}},
  note         = {Machine review of arXiv:2607.05090}
}
read the original 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 SpineRankNet, 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 score learned by ranking even improves discrimination between more distant classes. Source code is available at https://github.com/spinetools/spineranknet.

Figures

Figures reproduced from arXiv: 2607.05090 by the authors.

Figure 1
Figure 1. Multi-task architecture for joint classification and ranking. A shared 3D ResNet-18 encoder produces an embedding z, followed by task-specific classification and ranking heads. Each ranking head outputs a continuous sever￾ity score s (t) ∈ [0, 10]; for clarity, the figure shows one representative task head, though there are heads for each of the 11 spinal grading tasks. Pairwise su￾pervision is derived from ground-t… view at source ↗
Figure 2
Figure 2. Example visualization of severity ranking for the pathology scores. IVD [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Segmentation Pre-training for Label-Efficient Lumbar Spine Degeneration Grading

    cs.CV 2026-08 conditional novelty 6.0 of 10

    Segmentation pre-training on automatically generated MRI masks lets a spine-grading model reach near full-supervision performance with only 20% of manual grading labels.

Reference graph

Works this paper leans on

27 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    In: ICPR

    Ahmed, S.A.A., Yanikoglu, B.: Relative Attribute Classification with Deep- RankSVM. In: ICPR. pp. 659–671 (2021)

    2021

  2. [2]

    In: ICML

    Burges, C.J.C., Shaked, T., Renshaw, E., Lazier, A., Deeds, M., Hamilton, N., Hullender, G.N.: Learning to rank using gradient descent. In: ICML. pp. 89–96 (2005)

    2005

  3. [3]

    Brain and Spine3, 102401 (2023)

    Cina, A., Haschtmann, D., Damopoulos, D., Gerber, N., Loibl, M., Fekete, T., Mannion, A., Galbusera, F.: Modic change classification from sagittal MRI with 3D convolutional neural networks and comparative analysis of normalization tech- niques. Brain and Spine3, 102401 (2023)

    2023

  4. [4]

    The Lancet Rheumatology5(6), e316– e329 (2023)

    Ferreira, M.L., et al.: Global, regional, and national burden of low back pain, 1990– 2020, its attributable risk factors, and projections to 2050: a systematic analysis of the Global Burden of Disease Study 2021. The Lancet Rheumatology5(6), e316– e329 (2023)

    1990

  5. [5]

    The Spine Journal17(4), 554–561 (2017)

    Herzog, R., Elgort, D.R., Flanders, A.E., Moley, P.J.: Variability in diagnostic error rates of 10 MRI centers performing lumbar spine MRI examinations on the same patient within a 3-week period. The Spine Journal17(4), 554–561 (2017)

    2017

  6. [6]

    Artificial Intelligence172(16–17), 1897–1916 (2008)

    Hüllermeier, E., Fürnkranz, J., Cheng, W., Brinker, K.: Label ranking by learning pairwise preferences. Artificial Intelligence172(16–17), 1897–1916 (2008)

    1916

  7. [7]

    Medical Image Analysis41, 63–73 (2017)

    Jamaludin, A., Kadir, T., Zisserman, A.: SpineNet: Automated classification and evidence visualization in spinal MRIs. Medical Image Analysis41, 63–73 (2017)

    2017

  8. [8]

    In: ISBI

    Kowlagi, N., Nguyen, H.H., McSweeney, T., Saarakkala, S., Määttä, J., Karppinen, J., Tiulpin, A.: A stronger baseline for automatic Pfirrmann grading of lumbar spine MRI using deep learning. In: ISBI. pp. 1–5 (2023)

    2023

Show all 27 references
  1. [9]

    Skeletal Radiology40(8), 1033–1039 (2011)

    Lee, G.Y., Lee, J.W., Choi, H.S., Oh, K.J., Kang, H.S.: A new grading system of lumbar central canal stenosis on MRI: an easy and reliable method. Skeletal Radiology40(8), 1033–1039 (2011)

    2011

  2. [10]

    Journal of Spine Surgery 6(Suppl 1), S84–S99 (2020) 10 M

    Lewandrowski, K.U., de Carvalho, P.S.T., Calderaro, A.L., Dos Santos, T.S., de Lima E Silva, M.S., de Carvalho, P., J., Yeung, A.: Outcomes with transforami- nal endoscopic versus percutaneous laser decompression for contained lumbar her- niated disc: a survival analysis of tr...

    2020

  3. [11]

    npj Digital Medicine3(1) (2020)

    Li, M.D., Chang, K., Bearce, B., Chang, C.Y., Huang, A.J., Campbell, J.P., Brown, J.M., Singh, P., et al.: Siamese neural networks for continuous disease severity evaluation and change detection in medical imaging. npj Digital Medicine3(1) (2020)

    2020

  4. [12]

    In: MLHC

    Lu, J.T., Pedemonte, S., Bizzo, B., Doyle, S., Andriole, K.P., Michalski, M.H., Gonzalez, R.G., Pomerantz, S.R.: Deep spine: automated lumbar vertebral seg- mentation, disc-level designation, and spinal stenosis grading using deep learning. In: MLHC. pp. 403–419. PMLR (2018)

    2018

  5. [13]

    European Spine Journal (2025)

    Minin, A., Leonova, O., Krutko, A., Elgaeva, E., Antonets, D., Shtokalo, D., Tsepilov, Y.: SpineScan: a deep learning model for lumbar spine MRI annotation and Pfirrmann grading assessment. European Spine Journal (2025)

    2025

  6. [14]

    Radiology166(1), 193–199 (1988)

    Modic,M.T.,Steinberg,P.M.,Ross,J.S.,Masaryk,T.J.,Carter,J.R.:Degenerative disk disease: assessment of changes in vertebral body marrow with MR imaging. Radiology166(1), 193–199 (1988)

    1988

  7. [15]

    PLOS ONE15(11), e0241309 (2020)

    Natalia, F., Meidia, H., Afriliana, N., Young, J.C., Yunus, R.E., Al-Jumaily, M., Al-Kafri, A., Sudirman, S.: Automated measurement of anteroposterior diameter and foraminal widths in MRI images for lumbar spinal stenosis diagnosis. PLOS ONE15(11), e0241309 (2020)

    2020

  8. [16]

    Neubert, A., Fripp, J., Engstrom, C., Walker, D., Weber, M.A., Schwarz, R., Crozier, S.: Three-dimensional morphological and signal intensity features for de- tectionofintervertebraldiscdegenerationfrommagneticresonanceimages.Journal of the American Medical Informatics Associa...

    2013

  9. [17]

    In: ICCV

    Parikh, D., Grauman, K.: Relative attributes. In: ICCV. pp. 503–510 (2011)

    2011

  10. [18]

    Pedregosa, F., Bach, F.R., Gramfort, A.: On the consistency of ordinal regression methods. J. Mach. Learn. Res.18, 55:1–55:35 (2014)

    2014

  11. [19]

    In: International Workshop on Ma- chine Learning in Medical Imaging

    Pedregosa, F., Cauvet, E., Varoquaux, G., Pallier, C., Thirion, B., Gramfort, A.: Learning to rank from medical imaging data. In: International Workshop on Ma- chine Learning in Medical Imaging. pp. 234–241. Springer (2012)

    2012

  12. [20]

    Spine26(17), 1873– 1878 (2001)

    Pfirrmann, C.W., Metzdorf, A., Zanetti, M., Hodler, J., Boos, N.: Magnetic reso- nance classification of lumbar intervertebral disc degeneration. Spine26(17), 1873– 1878 (2001)

    2001

  13. [21]

    BMC Medical Imaging15(1) (2015)

    Saffari, S.E., Löve, Á., Fredrikson, M., Smedby, O.: Regression models for analyz- ing radiological visual grading studies – an empirical comparison. BMC Medical Imaging15(1) (2015)

    2015

  14. [22]

    Pattern Analysis and Applications 26(3), 941–955 (2023)

    Shi, X., Cao, W., Raschka, S.: Deep neural networks for rank-consistent ordinal regression based on conditional probabilities. Pattern Analysis and Applications 26(3), 941–955 (2023)

    2023

  15. [23]

    In: ACCV

    Souri, Y., Noury, E., Adeli, E., Mousavi, P., Tizhoosh, H.R.: Deep relative at- tributes. In: ACCV. pp. 118–133. Springer (2016)

    2016

  16. [24]

    In: MICCAI

    Windsor, R., Jamaludin, A., Kadir, T., Zisserman, A.: Context-aware transformers for spinal cancer detection and radiological grading. In: MICCAI. pp. 271–281 (2022)

    2022

  17. [25]

    Windsor, R., Jamaludin, A., Kadir, T., Zisserman, A.: SpineNetV2: Auto- mated detection, labelling and radiological grading of clinical MR scans (2022), https://arxiv.org/abs/2205.01683

    2022 arXiv

  18. [26]

    Scientific Reports14(1) (2024)

    Windsor, R., Jamaludin, A., Kadir, T., Zisserman, A.: Automated detection, la- belling and radiological grading of clinical spinal MRIs. Scientific Reports14(1) (2024)

    2024

  19. [27]

    Nature Communications13, 841 (2022)

    Zheng, H.D., Sun, Y.L., et al.: Deep learning-based high-accuracy quantitation for lumbar intervertebral disc degeneration from MRI. Nature Communications13, 841 (2022)

    2022

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.