Pith. sign in

REVIEW 5 major objections 6 minor 17 references

Diffusion-Based Semantic Segmentation of Lumbar Spine MRI Scans of Lower Back Pain Patients

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read SpineSegDiff, a diffusion model seeded by nnU-Net, reports slightly higher Dice than nnU-Net and IISDM for lumbar spine MRI segmentation, with the biggest gain on intervertebral discs.

desk verdict A reproducible diffusion-based lumbar spine MRI segmentation benchmark with a small IVD Dice gain, undercut by a selected-slice evaluation and unsupported abstract claims about degenerated discs. read the letter →

arxiv 2411.10755 v2 pith:KXC7RKEO submitted 2024-11-16 eess.IV cs.CVcs.LG

Maria Monzon , Thomas Iff , Ender Konukoglu , Catherine R. Jutzeler This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords diffusionmodelssemanticsegmentationlumbarspineMRIintervertebraldisclowbackpainnnU-NetbaselineuncertaintyestimationSPIDERdataset
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

This paper is trying to establish that a diffusion-based segmentation model, SpineSegDiff, can segment vertebrae, intervertebral discs, and the spinal canal in lumbar spine MRI scans of low back pain patients using a single framework that handles T1w, T2w, and combined contrasts. On 5-fold cross-validation over 218 patients, it reports SpineSegDiff slightly outperforming nnU-Net and IISDM on every modality, with the clearest gain on intervertebral discs (Dice 0.89–0.90 versus 0.84–0.85 for nnU-Net). This matters because disc degeneration is a frequent source of low back pain, and diffusion models additionally produce per-pixel uncertainty maps that could guide clinical review. The paper also claims that starting the diffusion from an nnU-Net pre-segmentation cuts the required diffusion time steps to about 30 without losing accuracy.

What carries the argument

The load-bearing machinery is the conditional denoising loop with an nnU-Net prior. A pretrained nnU-Net produces an initial segmentation, that mask is partially noised, and SpineSegDiff learns a short denoising trajectory back to the ground-truth mask. The denoising UNet combines features from a dedicated multi-scale image encoder with sinusoidal time-step embeddings, and the composite loss adds Dice and cross-entropy terms to the standard MSE denoising loss. Inference averages several stochastic samples and entropy-weights the last timesteps, so the same mechanism that produces the mask also produces an uncertainty map.

What would settle it

Score SpineSegDiff and nnU-Net on the 18 excluded oblique scans and on off-center slices of the same 218 subjects under the same 5-fold split; if IVD Dice on those slices is no better than nnU-Net's, the claimed superiority is an artifact of evaluating only well-aligned central slices.

Watch

Extended reading notes

Core claim

The central claim is that treating lumbar spine segmentation as conditional denoising—generating the mask from a partially noised starting point with the MRI as condition—improves accuracy over one-shot segmentation baselines, particularly for intervertebral discs. SpineSegDiff directly predicts the mask $x_0$ rather than the noise, using a pretrained nnU-Net mask $\hat{x}_{\text{pre}}$ as a partially noised prior and a two-encoder denoising UNet with time embeddings to refine it. Reported mean Dice is 0.908–0.917 for SpineSegDiff versus 0.887–0.903 for IISDM and 0.887–0.893 for nnU-Net, with IVD Dice 0.89–0.90 versus 0.84–0.85 for nnU-Net. At inference, an uncertainty ensemble averages $S$ stochastic samples and weights predictions across timesteps by entropy, yielding both the final mask and uncertainty estimates.

Load-bearing premise

The load-bearing premise is that the evaluation subset represents the cohort: only the central slice of non-oblique scans is scored, and the 18 oblique scans are excluded from testing, so the reported Dice values do not cover the full variety of scans the model would see in practice.

Editorial extensions

If this is right

  • Intervertebral disc segmentation, the clinically sensitive structure, improves by roughly five Dice points over nnU-Net while spinal canal and vertebrae scores stay comparable.
  • The nnU-Net pre-segmentation strategy makes the diffusion component practical: $T=30$ timesteps already matches $T=1000$ performance and lifts IVD Dice from 0.84 at $T=0$ to 0.89.
  • One model can serve T1w, T2w, and combined inputs without contrast-specific retraining, which simplifies deployment across imaging protocols.
  • Segmentation accuracy varies with pathology: spondylolisthesis and disc narrowing significantly degrade scores, and disc degeneration specifically lowers IVD Dice, so pathology burden must be accounted for when benchmarking.

Reading between the lines

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

  • Beyond the paper, the reported Dice is computed on the central slice only, so the headline accuracy likely overstates whole-volume performance; the same comparison on off-center slices is a natural next experiment.
  • Beyond the paper, the 18 oblique scans kept out of evaluation but used in training are a built-in stress test; scoring them separately would show whether the model's edge survives acquisition-angle variation.
  • Beyond the paper, because entropy-weighted uncertainty comes free with the ensemble, a clinical workflow could route low-confidence disc regions to manual review rather than accepting the mask blindly.
  • Beyond the paper, the pathology-dependent accuracy drops hint that per-patient segmentation error could be used as a crude indicator of disc degeneration, though the paper does not test that use.
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

5 major / 6 minor

Summary. The paper presents SpineSegDiff, a 2D diffusion-based segmentation model for lumbar spine MRI that uses a pre-segmentation from nnU-Net as a conditioning prior and an uncertainty ensemble for inference. The model is evaluated on the public SPIDER dataset (T1w, T2w, and combined) with 5-fold cross-validation, against nnU-Net and IISDM baselines, and on an ablation of diffusion timesteps. The authors report slight Dice improvements over the baselines, especially for intervertebral disc (IVD) segmentation, and analyze how various degenerative pathologies affect segmentation performance. Code and reproducibility instructions are provided.

Significance. If the performance claims hold, the paper offers a practical diffusion-based segmentation approach with uncertainty quantification and a reduced-step sampling strategy, on a public dataset with released code. Strengths include the use of a public benchmark (SPIDER), patient-wise 5-fold cross-validation, an explicit timestep ablation, and a pathology-stratified analysis that attempts to connect segmentation accuracy to clinically relevant conditions. The main limitation is that the reported superiority over baselines rests on small mean differences that are not tested for statistical significance, and the evaluation is confined to a selected subset of scans and slices, so the generalizability of the headline numbers is not established as stated.

major comments (5)
  1. [Section 4.3, Table 1] The central claim that SpineSegDiff 'slightly outperformed' the baselines is not supported by any statistical test. For example, IVD Dice is 0.89 vs. 0.84 (T1w) and 0.90 vs. 0.85 (T2w), but these are reported only as cross-validation means and standard deviations. Since the same folds are used for all models, a paired significance test (e.g., Wilcoxon signed-rank or paired t-test on fold-level Dice) is required to establish that the differences are not due to chance. Without this, the headline comparison is not verified.
  2. [Section 4.3 and Appendix B.2] The evaluation subset is load-bearing and its representativeness is unquantified. The paper states that 18 oblique MRI scans were excluded from evaluation but kept for training, and that 2D models were evaluated only on the central slice. Yet Table 1 presents these results as the model's performance on the dataset, and the abstract/conclusion generalize to 'MRI scans' without this scope limitation. The authors should justify the exclusion, report results on the excluded scans and on non-central slices (e.g., using the 3D nnU-Net for comparison), and provide a sensitivity analysis. If SpineSegDiff performs worse on oblique scans or off-center slices, the reported improvements could shrink or reverse.
  3. [Abstract and Figure 2] The abstract's claim about 'identification of degenerated IVDs' is not supported by the paper's own analysis. The paper performs semantic segmentation, not detection or classification of degenerated discs. Moreover, Figure 2 (Disc Degeneration panel) shows that IVD Dice is significantly lower when disc degeneration is present (p = 3.43e-04), which contradicts the implication that the model is particularly effective for degenerated IVDs. This claim should be removed or replaced with a precise statement about segmentation performance on degenerated discs.
  4. [Appendix B.2, Table 1] It is unclear which nnU-Net configuration is reported in Table 1. The appendix describes both 2D and 3D nnU-Net training, and states that 2D models are evaluated on the central slice while 3D models use the entire volume. If Table 1 compares SpineSegDiff (2D, central slice) against a 3D nnU-Net (full volume), the comparison is inherently unfair. The authors must specify the configuration for each row and ensure that all compared models are evaluated under the same protocol.
  5. [Section 4.4, Table 2] The pre-segmentation ablation shows that T=30 yields IVD Dice of 0.89 versus 0.84 for T=0 (the nnU-Net prior), which is the key practical advantage of the method. As with the main table, no significance testing is reported for this improvement. Given the small absolute difference and the fact that T=0 is the same nnU-Net baseline, a paired test is needed to confirm that the diffusion refinement is statistically meaningful.
minor comments (6)
  1. [Abstract] The phrase 'comparable outperformed' is grammatically unclear and should be rewritten to state precisely whether the model is comparable or superior to the baselines.
  2. [Introduction] There are numerous typographical errors, including 'segmenton', 'due toltifaceted', 'strucures', and 'mantaining'. A thorough language edit is needed.
  3. [Appendix B.3] The heading 'Implicit Image Segmentation Diffusion Model (IISMD)' should be 'IISDM' to match the abbreviation used elsewhere in the paper.
  4. [Equation (3)] The notation is confusing: the sum runs from t=1 to Ts, but the text says 'across Ts last timesteps'. Please clarify the indexing (e.g., whether t indexes the last Ts timesteps in reverse order).
  5. [Figure 2] The figure caption says 'T-test statistic (p-value)' but the subplot labels show only the statistic and p-value; consider clarifying that these are unpaired t-tests and note whether the Benjamini-Hochberg correction was applied to all shown p-values.
  6. [Section 5] The conclusion states that the statistical analysis reveals spondylolisthesis and disc narrowing 'exhibit the highest t-statistics and the lowest p-values,' but according to Figure 2 the lowest p-value is for disc degeneration (4.80e-03 for spondylolisthesis SC, and 8.04e-05 for lower endplate IVD are lower). Please verify the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical benchmark; the nnU-Net prior/baseline overlap is not a derivation-level circularity because evaluation uses held-out ground truth.

full rationale

The paper reports a comparative segmentation evaluation and does not present a derivation chain in which a predicted quantity is defined in terms of itself. The only potentially self-referential element is that the nnU-Net model serves both as the pre-segmentation prior for SpineSegDiff and as one of the compared baselines (Section 3.1: 'SpineSegDiff training is significantly accelerated through the implementation of a pre-segmentation strategy... The initial segmentation x̂_pre is predicted with the pre-trained baseline nnU-Net model'). However, this is not circular: SpineSegDiff is trained to recover the ground-truth mask x0 from the partially noised pre-segmentation, and the reported Dice scores are computed against manual ground truth in a patient-wise 5-fold cross-validation (Sections 4.1-4.3). The comparison nnU-Net versus SpineSegDiff is therefore an empirical benchmark, not a fitted quantity renamed as a prediction. The uncertainty-weighted ensemble (Eq. 3) is adopted from prior work (Xing et al., 2023), but no external source is used as the sole justification of a conclusion that reduces to its own input. Concerns about the evaluation subset—18 oblique scans excluded from testing and only the central slice scored (Sections 4.3 and B.2)—are threats to generalizability and statistical validity, not circularity. No self-citation chain carries a load-bearing argument, and no known result is merely renamed. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper's central claim rests on the quality of SPIDER labels, the representativeness of the central slice, and the exclusions made in the evaluation protocol. These are empirical assumptions, not mathematical axioms, but they are load-bearing for the reported Dice scores.

free parameters (4)
  • Diffusion timesteps T = 1000 for training; ablation over 30, 100, 300, 500, 1000 for inference
    Chosen for training and then ablated; the T=0 baseline is the nnU-Net prior.
  • Number of samples S in uncertainty ensemble = not reported
    Appendix A.2 defines S samples but never gives S; the final mask depends on it.
  • Uncertainty ensemble window Ts = not reported
    Equation (3) sums over Ts last timesteps, but Ts is never specified.
  • Image size and normalization percentile = 320x320, 98th percentile normalization
    Standard preprocessing choices; affect the exact Dice numbers.
assumptions (3)
  • domain assumption SPIDER ground-truth segmentation labels are accurate for the three structures (SC, VB, IVD).
    All Dice scores are computed against these labels; no independent verification is reported.
  • domain assumption A single central 2D slice is sufficient to evaluate lumbar spine segmentation performance.
    Appendix B.2 states 2D models are trained and evaluated on the central slice only, yet the paper suggests clinical relevance for general MRI scans.
  • domain assumption The 18 excluded oblique scans do not affect the model's reported performance.
    Section 4.3 excludes them from evaluation but retains for training, and no sensitivity analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion-Based Semantic Segmentation of Lumbar Spine MRI Scans of Lower Back Pain Patients." pith.science (2026). https://pith.science/paper/KXC7RKEO

@misc{pith2026241110755,
  author       = {Pith},
  title        = {Pith review of: Diffusion-Based Semantic Segmentation of Lumbar Spine MRI Scans of Lower Back Pain Patients},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXC7RKEO}},
  note         = {Machine review of arXiv:2411.10755}
}
read the original abstract

This study introduces a diffusion-based framework for robust and accurate segmenton of vertebrae, intervertebral discs (IVDs), and spinal canal from Magnetic Resonance Imaging~(MRI) scans of patients with low back pain (LBP), regardless of whether the scans are T1w or T2-weighted. The results showed that SpineSegDiff achieved comparable outperformed non-diffusion state-of-the-art models in the identification of degenerated IVDs. Our findings highlight the potential of diffusion models to improve LBP diagnosis and management through precise spine MRI analysis.

Figures

Figures reproduced from arXiv: 2411.10755 by the authors.

Figure 1
Figure 1. SpineSegDiff architecture overview: the 2D MRI scan (y) is concatenated with the partially noised mask to generate the segmentation image xT . The architecture is composed by a multi-scale image encoder and a UNet-based model for denoising at each diffusion. The final mask is inferred using a uncertainty-step ensemble that creates uncertainty maps across multiple samples and time steps. 2. Related Work Recently, Den… view at source ↗
Figure 2
Figure 2. Qualitative and quantitative analysis of SpineSegDiff performance. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 7 canonical work pages

  1. [5]

    Accelerating Diffusion Models via Pre-segmentation Diffusion Sampling for Medical Image Segmentation

    URL http://arxiv.org/abs/2210.17408. Jan Hartvigsen, Mark J Hancock, Alice Kongsted, Quinette Louw, Manuela L Ferreira, St´ ephane Genevay, et al. What low back pain is and why we need to pay attention. Lancet, 391(10137): 2356–2367, 6

    work page Pith review arXiv

  2. [8]

    doi: 10.1007/ 978-3-658-25326-4 {\ }7

    ISSN 1431472x. doi: 10.1007/ 978-3-658-25326-4 {\ }7. URL https://arxiv. org/abs/1809.10486v1. Fabian Isensee, Paul F. Jaeger, Simon A.A. Kohl, Jens Petersen, and Klaus H. Maier- Hein. nnU-Net: a self-configuring method for deep learning-based biomedical image segmenta- tion. Nature Methods 2020 18:2 , 18(2):203– 211, 12

    arXiv

  3. [10]

    Zengxin Liu, Caiwen Ma, Wenji She, and Meilin Xie

    URL http://arxiv.org/ abs/2311.01797. Zengxin Liu, Caiwen Ma, Wenji She, and Meilin Xie. Biomedical image segmentation using denoising diffusion probabilistic models: A comprehensive re- view and analysis. Applied Sciences, 14(2),

    arXiv

  4. [13]

    Dexu Wang, Zhikai Yang, Ziyan Huang, and Lixu Gu

    URL https://zenodo.org/records/8009680. Dexu Wang, Zhikai Yang, Ziyan Huang, and Lixu Gu. Spine segmentation with multi-view gcn and boundary constraint. In 2022 44th Annual Inter- national Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), pages 2136– 2139,

    work page arXiv 2022

  5. [14]

    Julia Wolleb, Robin Sandk¨ uhler, Florentin Bieder, Philippe Valmaggia, and Philippe C

    doi: 10.1109/embc48229.2022.9871222. Julia Wolleb, Robin Sandk¨ uhler, Florentin Bieder, Philippe Valmaggia, and Philippe C. Cattin. Dif- fusion Models for Implicit Image Segmentation En- sembles. 12

    arXiv 2022

  6. [15]

    Diffusion Models for Implicit Image Segmentation Ensembles

    URL https://arxiv.org/abs/ 2112.03145v2. Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, et al. MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model. 11

    work page Pith review arXiv

  7. [16]

    Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu

    URL https://arxiv.org/abs/ 2211.00611v5. Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu. Diff-UNet: A Diffusion Embedded Network for Volumetric Segmentation. 3

    arXiv

  8. [17]

    6 Diffusion-Based Segmentation of Lumbar MRI of Lower Back Pain Patients Appendix A

    URL http://arxiv.org/abs/2303.10326. 6 Diffusion-Based Segmentation of Lumbar MRI of Lower Back Pain Patients Appendix A. SpineSegDiff A.1. Architecure Details SpineSegDiff employs a dual-encoder architecture op- timized for lumbar spine MRI segmentation. The model integrates a U-shaped backbone, adapted from the optimal nnUNet configuration, with an addi...

    arXiv 2023

Show all 17 references
  1. [2006]

    doi: 10.1093/ageing/afj055

    ISSN 1468-2834. doi: 10.1093/ageing/afj055. Daryl R. Fourney, Gunnar Andersson, Paul M. Arnold, Joseph Dettori, Alex Cahana, Michael G. Fehlings, et al. Chronic low back pain: A heterogeneous condition with challenges for an evidence-based approach. Spine, 36(21 Suppl.), 10

    doi

  2. [2011]

    doi: 10.1097/brs.0b013e31822f0a0d

    ISSN 03622436. doi: 10.1097/brs.0b013e31822f0a0d. URL https://journals.lww.com/spinejournal/ fulltext/2011/10011/chronic_low_back_ pain__a_heterogeneous_condition.1.aspx. Xutao Guo, Yanwu Yang, Chenfei Ye, Shang Lu, Yang Xiang, and Ting Ma. Accelerating Diffusion Models via Pr...

    2011 doi

  3. [2018]

    URL https://pubmed.ncbi.nlm.nih

    doi: 10.1016/s0140-6736(18) 30480-x. URL https://pubmed.ncbi.nlm.nih. gov/29573870/. 5 Diffusion-Based Segmentation of Lumbar MRI of Lower Back Pain Patients Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denois- ing Diffusion Probabilistic Models. 6

  4. [2019]

    org/abs/1911.01685

    URL http://arxiv. org/abs/1911.01685. M. Jorge Cardoso et al. Monai: An open-source framework for deep learning in healthcare

    1911 arXiv

  5. [2020]

    Fabian Isensee, Jens Petersen, Andre Klein, David Zimmerer, Paul F

    URL http://arxiv.org/abs/2006.11239. Fabian Isensee, Jens Petersen, Andre Klein, David Zimmerer, Paul F. Jaeger, Simon Kohl, et al. nnU- Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation. Informatik aktuell, page 22, 9

    2006 arXiv

  6. [2021]

    doi: https://doi.org/10.1016/j.media

    ISSN 1361-8415. doi: https://doi.org/10.1016/j.media. 2021.102166. URL https://www.sciencedirect. com/science/article/pii/S1361841521002127. Jasper Willem van der Graaf, Miranda L van Hooff, Constantinus F M Buckens, Matthieu Rutten, Job L C van Susante, Robert Jan Kroeze, et ...

    2021

  7. [2022]

    Hyungjin Chung, Eun Sun Lee, and Jong Chul Ye

    URL https://arxiv.org/abs/2211.02701. Hyungjin Chung, Eun Sun Lee, and Jong Chul Ye. Mr image denoising and super-resolution using reg- ularized reverse diffusion. IEEE Transactions on Medical Imaging, 42(4):922–934,

    arXiv

  8. [2023]

    doi: 10.1016/j.media.2023.102846

    ISSN 1361-8423. doi: 10.1016/j.media.2023.102846. URL http: //www.ncbi.nlm.nih.gov/pubmed/37295311. Boah Kim, Yujin Oh, and Jong Chul Ye. Diffu- sion adversarial representation learning for self- supervised vessel segmentation. In The Eleventh International Conference on Learn...

    2023

  9. [2024]

    doi: 10.3390/app14020632

    ISSN 2076-3417. doi: 10.3390/app14020632. URL https://www.mdpi.com/2076-3417/14/2/632. S. Pang, C. Pang, et al. Spineparsenet: Spine parsing for volumetric mr image by a two-stage segmen- tation framework with semantic image representa- tion. IEEE Transactions on Medical Imagi...

    doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.