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

Jamaludin Amir; Jutzeler Catherine R.; Monzon Maria; Zisserman Andrew

REVIEW 3 major objections 6 minor 20 references

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

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Pre-training on automatic anatomical masks lets spine grading match full supervision with one fifth of the labels.

desk verdict A well-executed label-efficiency study for spine MRI grading with a plausible and useful result, but the paper must clarify whether segmentation pre-training touched the test split before the headline number can be trusted. read the letter →

arxiv 2608.04810 v1 pith:QKDWBMBO submitted 2026-08-05 cs.CV

Monzon Maria , Zisserman Andrew , Jutzeler Catherine R. , Jamaludin Amir This is my paper

classification cs.CV

keywords segmentationpre-traininglabelefficiencylumbarspinedegenerationgradingMRIpseudo-labelstransferlearning3DResNetordinalclassification

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

Grading lumbar spine degeneration on MRI normally demands large datasets of expert radiological annotations, while the geometry that actually drives those grades — disc height, canal width, vertebral slippage — can be captured by automatic segmentation at negligible cost. This paper argues that pre-training a 3D encoder to segment the vertebrae, intervertebral discs, and spinal canal can replace most of that expensive manual labelling. On roughly 2,000 multicentre subjects across 11 pathologies, a segmentation-pretrained 3D ResNet-18 matched the task-averaged one-vs-rest ROC-AUC of a fully supervised from-scratch model using only 20% of the grading labels, and stayed ahead at every label fraction from 10% to 100%. The largest gains land on rare, spatially grounded pathologies such as spondylolisthesis and endplate defects, which supports the paper's claim that the benefit is anatomical structure rather than generic regularization. If correct, this substantially lowers the annotation budget for building reliable spine-grading systems.

What carries the argument

The load-bearing mechanism is a two-stage training loop. In the first stage, a 3D ResNet-18 encoder is attached to a symmetric U-Net decoder and optimized under the composite loss $L_{\text{seg}} = \tfrac{1}{2}(L_{\text{Dice}} + L_{\text{CE}})$ to reproduce four-class anatomical masks (background, vertebral body, intervertebral disc, spinal canal) generated automatically by the SpinePS tool. In the second stage, the trained encoder's 512-dimensional pooled embedding feeds lightweight per-task MLP grading heads, fine-tuned end-to-end under class-weighted cross-entropy with an optional auxiliary segmentation term $L = L_{\text{clf}} + \lambda_{\text{seg}} L_{\text{seg}}$, and a ranking-head variant supervises continuous severity scores with a margin-hinge pairwise loss. The working hypothesis is that the segmentation objective forces the encoder to internalize precisely the morphological geometry radiological grading scales are built on — disc-height loss, canal narrowing, listhesis offset — so that the learned representation transfers to grading with far fewer expert labels than a randomly initialized backbone.

What would settle it

Repeat the protocol with pre-training explicitly restricted to the 80% training-split subjects and check whether the 20%-label macro one-vs-rest ROC-AUC still sits near the from-scratch full-supervision value of 0.919; if excluding test subjects from pre-training erases the gap, the central claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that segmentation, viewed not as a preprocessing step but as a pre-training objective, is a direct substitute for a large share of expert grading labels in multi-pathology lumbar spine degeneration grading. The paper shows that a 3D ResNet-18 encoder trained under a combined Dice–cross-entropy loss to reproduce automatic masks of the vertebral bodies, intervertebral discs, and spinal canal — reaching a Dice of 0.94 against those pseudo-labels — transfers to eleven ordinal grading tasks. Fine-tuned with 20% of the grading labels, it reaches a macro one-vs-rest ROC-AUC of 0.915, above the from-scratch baseline's 0.869 at the same fraction and effectively equal to the from-scratch model's full-supervision 0.919; under full supervision it reaches 0.932 versus 0.919. The same label-efficiency pattern holds for a continuous severity-score variant, where 10% of labels suffices to lift the quadratic weighted kappa from 0.55 to 0.68, and the per-pathology gains concentrate in low-prevalence, morphology-driven conditions such as spondylolisthesis and upper and lower endplate defects.

Load-bearing premise

The load-bearing premise is that segmentation pre-training saw only the 80% training-split subjects: Section 3 never states that the pre-training cohort is restricted to that partition, so if validation and test subjects' MRI volumes entered pre-training, the reported 20%-label gains could reflect test-set exposure rather than a genuine geometric prior.

Editorial extensions

If this is right

Spine-degeneration datasets could be built with roughly one fifth of the radiologist annotation effort: the remaining labels are replaced by near-free automated segmentation, and the pretrained grader matches the from-scratch full-supervision macro ROC-AUC of 0.919 with 20% of labels (0.915).
The advantage persists at every label fraction and even under full supervision (0.932 vs 0.919), so the geometric prior helps beyond the scarce-label regime.
The benefit is concentrated where annotation-hungry training hurts most: rare, morphology-defined pathologies gain most at 20% labels (spondylolisthesis ROC-AUC 0.860 to 0.962), while common or signal-driven tasks such as Pfirrmann show small or even slightly negative differences at some fractions.
The same label-efficiency pattern holds for continuous severity ranking, so the segmentation prior is not tied to the categorical classification formulation.
Segmentation quality is essentially identical across ResNet-18/34/50 backbones (Dice 0.94), so the cheapest encoder suffices for the pretask, keeping the full pipeline lightweight.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the paper leaves implicit: the recipe should transfer to other anatomy-driven ordinal grading tasks such as knee osteoarthritis on MRI, with the gain expected to scale with how much of the grade is determined by measurable morphology rather than signal intensity.
The comparison is only against a randomly initialized baseline; a stronger test would pit segmentation pre-training against self-supervised alternatives such as masked-image reconstruction at matched label budgets, which would isolate how much of the gain is the geometric prior versus generic representation learning.
The paper itself notes that evaluation is on a single (multicentre) cohort with automatically generated masks; whether the savings survive transfer to other cohorts or to radiologist-refined masks is untested.
The geometric-prior explanation is directly testable by ablating individual mask classes: if dropping the spinal-canal or vertebral-body class from the pre-training loss erases the advantage on spondylolisthesis or stenosis, the claim that specific anatomy carries the gain is confirmed.

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.

Desk Editor's Note

read the letter

The core result is genuinely useful: pre-training a 3D ResNet to segment vertebrae, discs, and spinal canal from SpinePS pseudo-labels, then fine-tuning on a fraction of grading labels, consistently beats a from-scratch baseline across 11 pathologies and every label fraction from 10% to 100%. The gain is largest for rare, geometry-grounded tasks like spondylolisthesis and endplate defects, which is exactly where you'd expect a spatial prior to help. The systematic sweep of label fractions, three seeds, and per-pathology breakdowns make this a careful empirical study. The segmentation quality itself is credible, and the authors are appropriately cautious about single-cohort evaluation and pseudo-label noise.

The soft spots are real but mostly fixable. The biggest one is the reader's stress-test concern: Section 3 says patients are split 80/10/10, but the pre-training description never states which split supplied the MRI volumes for segmentation pre-training. Since SpinePS masks are generated for the whole cohort, the natural implementation is to pre-train on all subjects, including the test split. The from-scratch baseline never sees a test volume at any stage, so if pre-training used the full cohort, the seg-pretrained encoder gets an unfair advantage. This would directly inflate the headline 20%-labels result. The paper must state, and ideally verify with a split-confined pre-training run, that the test subjects' volumes were unseen during pre-training.

Second, Table 2 and Table B1 report different macro ROC-AUCs for what look like the same configurations (e.g., 20% labels scratch 0.869 vs 0.895). The difference is presumably the auxiliary segmentation loss during fine-tuning, but the paper doesn't reconcile this. Since the auxiliary loss also helps the from-scratch model, the paper needs to be explicit about which setting anchors the main comparison, otherwise the reader can't tell how much of the gain comes from pre-training versus the auxiliary loss itself.

Third, the per-pathology standard deviations for very rare classes (e.g., spondylolisthesis with 7 severe cases out of 12,000) seem implausibly low, and the paper doesn't explain how these were computed across seeds or whether the rare-class AUC is reliable at all. This is a minor issue for the aggregate claim but should be addressed.

The central argument holds up structurally, and the concerns are clarifications rather than refutations. This paper deserves a serious referee; the split question is exactly what a good referee would catch. If the authors can confirm training-split isolation and reconcile the tables, the result is a solid contribution to medical imaging label efficiency.

Referee Report

3 major / 6 minor

Summary. The paper proposes a two-stage framework for lumbar spine degeneration grading. In the first stage, a 3D ResNet encoder with a U-Net decoder is pre-trained to segment vertebral bodies, intervertebral discs, and the spinal canal from automatically generated SpinePS masks. In the second stage, the encoder is fine-tuned with lightweight per-pathology classification heads on 10-100% of the expert grading labels from the Genodisc dataset (roughly 2,000 subjects, 11 ordinal tasks). The authors report that segmentation pre-training improves the macro one-vs-rest ROC-AUC over a from-scratch baseline at every label fraction, reaches the from-scratch full-supervision performance with only 20% of labels, and gives the largest gains for rare or spatially grounded pathologies. A continuous severity-ranking variant and segmentation-derived morphometry correlations are also presented.

Significance. The paper addresses a practically important problem: reducing expert annotation cost for spine degeneration grading through cheap, automatically produced segmentation pseudo-labels. Its strengths are the systematic label-fraction sweep (10-100%), per-pathology reporting, three-seed evaluation, and the coherent per-pathology pattern that gains concentrate in low-prevalence, geometry-grounded tasks. The limitations of single-cohort evaluation and pseudo-label supervision are acknowledged. However, the central claim is empirical, and its validity depends critically on the isolation of the pre-training data and on the consistency and statistical robustness of the reported numbers.

major comments (3)

[Section 3 (Dataset and Training paragraphs)] The manuscript never states whether the segmentation pre-training phase is restricted to the 80% training split. Because SpinePS masks can be generated for the entire cohort, the natural implementation would pre-train on all subjects, which would give the pre-trained encoder access to test volumes and confound the comparison with the from-scratch baseline, which never sees test images. Please state explicitly which split was used for pre-training and verify the isolation (e.g., by reporting the number of subjects used for pre-training). If pre-training used the full cohort, the label-efficiency results, including the 20%-labels claim, must be recomputed with a pre-training set that excludes test subjects.
[Table 2 vs. Supplementary Table B1] The macro one-vs-rest ROC-AUC values in Table 2 cannot be reconciled with Supplementary Table B1 for the same configurations. For example, at 20% labels Table 2 reports Scratch=0.869 and Seg=0.915, while Table B1 reports lambda_seg=0 values of 0.875 and 0.911 and lambda_seg=0.5 values of 0.895 and 0.924. The manuscript should state the exact lambda_seg setting, task set, and seed protocol used in the main experiments and correct any discrepancy; otherwise the quantitative basis of the headline claim is not reproducible.
[Section 4.2 (Table 2)] The headline result that 20% of labels 'achieves near full-supervision performance' rests on a difference of 0.004 in macro ROC-AUC (Seg at 20%: 0.915 +/- 0.007; Scratch at 100%: 0.919 +/- 0.007, with standard deviations converted from the reported 10^-2 units), which is within one standard deviation across the three seeds. The paper should report confidence intervals, bootstrap over patients or pathology-level paired tests, or otherwise demonstrate that the differences are not seed noise, especially for the claim that pre-training helps 'at all proportions.'

minor comments (6)

[Figure 3 and Table 1] The 'ground-truth overlay' in Figure 3 and the Dice scores in Table 1 are computed against SpinePS pseudo-labels, not manual ground-truth segmentations; the text should use 'pseudo-label' rather than 'ground truth' to avoid overstating anatomical accuracy.
[Section 3 (Training paragraph)] The default value of lambda_seg used in the headline experiments is never stated; please specify it in the main text so that Table 2 can be interpreted in light of Supplementary Table B1.
[Abstract and Section 5] The phrase 'near full-supervision performance' is ambiguous; specify that the reference is the fully supervised from-scratch baseline (macro ROC-AUC 0.919) rather than the pre-trained model at 100% labels (macro ROC-AUC 0.932).
[Figure 4 (left panel)] The x-axis ticks (10, 30, 50, 70, 90) omit 20% and 100% even though these fractions are reported in Table 2; adding ticks for all evaluated fractions would make the plot easier to read.
[Section 3 (Preprocessing paragraph)] Please state whether SpineNetv2, used for disc localisation, is a frozen off-the-shelf model or was trained on the Genodisc cohort; this is relevant to the data-leakage assessment.
[Supplementary Tables B1 and B2] Report the number of seeds and the label-fraction sampling procedure used for the auxiliary-loss and continuous-ranking experiments, since the main text only gives selected values at 10% and 20%.

Circularity Check

0 steps flagged · score 2.0 of 10

No construction-level circularity: the label-efficiency claim is an empirical A/B comparison on independent expert grading labels; the only overlapping-author citation supplies a secondary loss and is not load-bearing. A separate, non-circular validity ambiguity is that the paper does not explicitly state whether segmentation pre-training was confined to the training split.

full rationale

The central claim is an empirical comparison, not a derivation: a 3D ResNet encoder is pre-trained to reproduce automatic SpinePS masks and then fine-tuned at label fractions from 10% to 100%, against an identically trained from-scratch baseline, evaluated on held-out expert grading labels. The grading labels are independent of the SpinePS pseudo-labels, so the result does not reduce by construction to the pre-training target. No equation equates a predicted quantity to a fitted input: Eq. (1) is the segmentation loss against pseudo-labels, Eq. (2) is the fine-tuning loss on grading labels, and Eqs. (3)-(4) are supervised ranking losses on grading labels; none renames a fitted parameter as a prediction. The one overlapping-author citation, [13] (Monzon et al., 2026), supplies the continuous severity-ranking loss used in a secondary variant; the loss is restated in the paper and the central segmentation-pre-training result does not depend on it, so this self-citation is non-load-bearing and does not raise the score above the minor category. A non-circular validity concern should be flagged: Section 3 states 'Patients are split 80/10/10 into train, validation, and test sets' but the Training paragraph says only that the encoder is 'first pretrained for up to 200 epochs ... to reproduce the SpinePS masks' without specifying that pre-training is restricted to the training split. If pre-training included validation/test subjects, the Seg model would have an unfair advantage over Scratch; this is a potential data-leakage confound, not a circularity, and the manuscript should state the split used for pre-training. Overall, the derivation chain is self-contained against the independent grading labels; score 2 reflects the minor self-citation and the split ambiguity, not construction-level circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim is an empirical method comparison. It depends on the standard inductive-bias assumption that segmentation supervision transfers to grading, on the adequacy of automatic masks, and on the unstated protocol detail that pre-training excludes the test split. The three tuning weights are minor and not fitted to the test labels in a way that could manufacture the result; the pre-training-split assumption is the hidden dependency that most affects the quantitative claim.

free parameters (3)

lambda_seg = 0.5 (auxiliary segmentation loss weight during fine-tuning)
Chosen by the authors; Table B1 shows it affects results, and the main text does not state clearly which value produced Table 2.
lambda_sim = 0.25
Similarity regularization weight in the continuous ranking loss, stated in Section 2.2 and supplementary.
ranking_margins_m0_m1 = m0=2.0, m1=1.5
Margins in the pairwise ranking hinge (Eq. 4); chosen constants, not fitted to the test set.

assumptions (3)

domain assumption The latent representation needed for anatomical segmentation is sufficient for multi-task pathology grading.
Core hypothesis of Section 2; the paper verifies it empirically on one cohort but does not prove it.
domain assumption SpinePS pseudo-labels with measured Dice 0.94 are accurate enough to serve as pre-training supervision without expert correction.
Section 4.1 evaluates segmentation quality against the same pseudo-labels; the effect of pseudo-label noise on grading transfer is not analyzed.
domain assumption Segmentation pre-training is performed only on the training-split subjects, so the test set is never seen by the encoder before grading evaluation.
Section 3 specifies the split for the grading data but not for the pre-training phase; the label-efficiency comparison is only valid under this assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Segmentation Pre-training for Label-Efficient Lumbar Spine Degeneration Grading." pith.science (2026). https://pith.science/paper/QKDWBMBO

@misc{pith2026260804810,
  author       = {Pith},
  title        = {Pith review of: Segmentation Pre-training for Label-Efficient Lumbar Spine Degeneration Grading},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QKDWBMBO}},
  note         = {Machine review of arXiv:2608.04810}
}

abstract

Automated assessment of degenerative pathology in the lumbar spine on magnetic resonance imaging (MRI) requires access to large-scale datasets of expert-annotated radiological gradings. In contrast, segmentation pseudo-labels can be generated by automated tools at negligible radiologist cost. We examine whether pre-training on segmentation can effectively replace a fraction of the manual grading annotations required for downstream supervision. We pre-train a 3D ResNet encoder to segment the vertebrae, intervertebral discs (IVDs), and the spinal canal, then fine-tune lightweight task-specific grading heads using different proportions of the available training data, ranging from $10\%$ to $100\%$. On a multicentre dataset of ${\sim}2{,}000$ subjects across 11 pathologies, segmentation pre-training, achieving a Dice score of $0.94$ against pseudo-labels, improved the task-averaged (macro) one-vs-rest ROC-AUC at all proportions. With only 20\% of grading labels after pre-training, the method achieved near full-supervision performance, with the largest gains observed for either low-prevalence or spatially grounded pathologies.

Figures

Figures reproduced from arXiv: 2608.04810 by the authors.

**Figure 2.** Segmentation-pretrained multi-task grading model. Segmentation is used [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Segmentation across backbones, spondylolisthesis and upper-endplate de [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Label efficiency of segmentation pre-training comparison. (left) Macro one-vs-rest ROC-AUC vs. labeled-data fraction (mean ± s.d., n=3 seeds), pretrained (petrol) vs. scratch (grey). (middle, right) Per-pathology Matthews correlation coefficient (MCC) at 20% and 60% la…

**Figure 5.** Figure 5: Continuous severity ranking with predicted segmentations. Test IVDs ordered by predicted severity score (top-left; ground-truth bottom-left). The score increases monotonically, correctly ordering severity even when adjacent clinical grades overlap. Disc narrowing exam…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

[1]

In: Pattern Recognition

Ahmed, S.A.A., Yanikoglu, B.: Relative attribute classification with deep-ranksvm. In: Pattern Recognition. ICPR International Workshops and Challenges. p. 659–671. Springer International Publishing (2021). https://doi.org/10.1007/978-3-030-68790-8_51

work page doi:10.1007/978-3-030-68790-8_51 2021
[2]

Frontiers in Surgery 10(2023)

Bassani, T., Cina, A., Galbusera, F., Sconfienza, L.M., Albano, D., Barcel- lona, F., et al.: Automatic classification of the vertebral endplate lesions in magnetic resonance imaging by deep learning model. Frontiers in Surgery 10(2023). https://doi.org/10.3389/fsurg.2023.1172313

work page arXiv 2023
[3]

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)

work page 2023
[4]

In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

Hooper, S., Chen, M., Saab, K., Bhatia, K., Langlotz, C., Ré, C.: A case for reframing automated medical image classification as segmentation. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 55415– 55441. Curran Associates, Inc. (2023)

work page 2023
[5]

Medical Image Analysis41, 63– 73 (2017)

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

work page 2017
[6]

Koslosky, E., Gendelberg, D.: Classification in brief: The Meyerding classifi- cation system of spondylolisthesis. Clin. Orthop. Relat. Res.478(5), 1125– 1130 (May 2020). https://doi.org/10.1097/CORR.0000000000001153 10 M. Monzon et al

work page doi:10.1097/corr.0000000000001153 2020
[7]

In: ISBI

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

work page 2023
[8]

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

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

work page 2011

Show all 20 references

[9]

AJR American Journal of Roentgenology194(4), 1095–1098 (2010)

Lee, S., Lee, J.W., Yeom, J.S., Kim, K.J., Kim, H.J., Chung, S.K., et al.: A practical mri grading system for lumbar foraminal steno- sis. AJR American Journal of Roentgenology194(4), 1095–1098 (2010). https://doi.org/10.2214/AJR.09.2772

2010 doi
[10]

Li, W., Yuille, A., Zhou, Z.: How well do supervised 3d models transfer to medical imaging tasks? In: International Conference on Learning Represen- tations (ICLR) (2024), https://arxiv.org/abs/2501.11253

2024 arXiv
[11]

In: MLHC

Lu, J.T., Pedemonte, S., Bizzo, B., Doyle, S., Andriole, K.P., Michal- ski, M.H., et al.: Deep spine: automated lumbar vertebral segmentation, disc-level designation, and spinal stenosis grading using deep learning. In: MLHC. pp. 403–419. PMLR (2018)

2018
[12]

European Radiology35(3), 1178–1189 (Oct 2024)

Möller, H., Graf, R., Schmitt, J., Keinert, B., Schön, H., Atad, M., et al.: Spineps—automatic whole spine segmentation of t2-weighted mr images using a two-phase approach to multi-class semantic and in- stance segmentation. European Radiology35(3), 1178–1189 (Oct 2024). https...

2024 doi
[13]

Monzon, M., Zisserman, A., Park, R.Y., Jutzeler, C.R., Jamaludin, A.: Be indiscrete: The benefits of learning continuous spine degeneration severity scores (2026), https://arxiv.org/abs/2607.05090

2026 arXiv
[14]

PLOS ONE 15(11), e0241309 (2020)

Natalia, F., Meidia, H., Afriliana, N., Young, J.C., Yunus, R.E., Al-Jumaily, M., et al.: Automated measurement of anteroposterior diameter and foram- inal widths in MRI images for lumbar spinal stenosis diagnosis. PLOS ONE 15(11), e0241309 (2020)

2020
[15]

Journal of the American Medical Informatics Association20(6), 1082–1090 (2013)

Neubert, A., Fripp, J., Engstrom, C., Walker, D., Weber, M.A., Schwarz, R., et al.: Three-dimensional morphological and signal intensity features for detection of intervertebral disc degeneration from magnetic resonance images. Journal of the American Medical Informatics Assoc...

2013
[16]

Spine26(17), 1873–1878 (2001)

Pfirrmann, C.W.A., Metzdorf, A., Zanetti, M., Hodler, J., Boos, N.: Magnetic resonance classification of lumbar intervertebral disc degenera- tion. Spine26(17), 1873–1878 (2001). https://doi.org/10.1097/00007632- 200109010-00011

2001 doi
[17]

In: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)

Snaauw, G., Gong, D., Maicas, G., Hengel, A.v.d., Niessen, W.J., Ver- jans, J., et al.: End-to-end diagnosis and segmentation learning from cardiac magnetic resonance imaging. In: 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). pp. 802–805 (2019). htt...

2019
[18]

ResearchGate preprint (2025) Seg-Pretrain for Data-Efficient Spine Grading 11

Warszawer, Y., Molinier, N., Valosek, J., Benveniste, P.L., Bédard, S., Shirbint, E., et al.: TotalSpineSeg: Robust Spine Segmentation with Landmark-Based Labeling in MRI. ResearchGate preprint (2025) Seg-Pretrain for Data-Efficient Spine Grading 11

2025
[19]

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

2022 arXiv
[20]

Nature Com- munications13, 841 (2022) 12 M

Zheng, H.D., Sun, Y.L., et al.: Deep learning-based high-accuracy quanti- tation for lumbar intervertebral disc degeneration from MRI. Nature Com- munications13, 841 (2022) 12 M. Monzon et al. Supplementary Material A Dataset The Genodisc grading distribution is summarized in ...

2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

In: Pattern Recognition

Ahmed, S.A.A., Yanikoglu, B.: Relative attribute classification with deep-ranksvm. In: Pattern Recognition. ICPR International Workshops and Challenges. p. 659–671. Springer International Publishing (2021). https://doi.org/10.1007/978-3-030-68790-8_51

work page doi:10.1007/978-3-030-68790-8_51 2021

[2] [2]

Frontiers in Surgery 10(2023)

Bassani, T., Cina, A., Galbusera, F., Sconfienza, L.M., Albano, D., Barcel- lona, F., et al.: Automatic classification of the vertebral endplate lesions in magnetic resonance imaging by deep learning model. Frontiers in Surgery 10(2023). https://doi.org/10.3389/fsurg.2023.1172313

work page arXiv 2023

[3] [3]

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)

work page 2023

[4] [4]

In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S

Hooper, S., Chen, M., Saab, K., Bhatia, K., Langlotz, C., Ré, C.: A case for reframing automated medical image classification as segmentation. In: Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., Levine, S. (eds.) Advances in Neural Information Processing Systems. vol. 36, pp. 55415– 55441. Curran Associates, Inc. (2023)

work page 2023

[5] [5]

Medical Image Analysis41, 63– 73 (2017)

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

work page 2017

[6] [6]

Koslosky, E., Gendelberg, D.: Classification in brief: The Meyerding classifi- cation system of spondylolisthesis. Clin. Orthop. Relat. Res.478(5), 1125– 1130 (May 2020). https://doi.org/10.1097/CORR.0000000000001153 10 M. Monzon et al

work page doi:10.1097/corr.0000000000001153 2020

[7] [7]

In: ISBI

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

work page 2023

[8] [8]

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

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

work page 2011