Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Flatten Wisely: How Patch Order Shapes Mamba-Powered Vision for MRI Segmentation

T0 review · 4 major / 4 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 This paper shows that the order in which image patches are flattened into a 1D sequence shifts Mamba-based MRI segmentation Dice by up to 27 points, and that simple raster scans win.

desk verdict Well-scoped benchmark with a plausible central claim, but single-run results and a three-block Friedman test don't yet support the strong quantitative conclusions. read the letter →

arxiv 2507.13384 v1 pith:QMKHVTLG submitted 2025-07-15 eess.IV cs.CVcs.LG

Osama Hardan , Omar Elshenhabi , Tamer Khattab , Mohamed Mabrok This is my paper
classification eess.IVcs.CVcs.LG
keywords VisionMambapatchscanorderMRIsegmentationstate-spacemodelsspatiallocalityMS2DFriedmantestbrain
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 argues that the order in which a 2D image is flattened into the 1D patch sequence fed to a Mamba-based vision model is a critical hyperparameter for brain-MRI segmentation, not a neutral implementation detail. Across 21 scan strategies on three public datasets covering more than 70,000 slices, test Dice scores varied by up to 27 points depending only on the scan path, with a Friedman test giving $\chi^2_{20}=43.9$ and $p=0.0016$. Spatially contiguous raster scans, especially a single horizontal pass and a four-direction mix of horizontal and vertical rasters, consistently ranked highest, while disjointed diagonal scans ranked lowest. The paper introduces MS2D, a parameter-free module that exposes twelve primitive scan paths with no change in FLOPs or parameter count, and concludes that scan order should be treated as a free hyperparameter in state-space vision models.

What carries the argument

The load-bearing object is MS2D (Multi-Scan 2D), a replacement for the standard four-direction scan unit in a VMamba-based UNet backbone. MS2D defines a bank of twelve primitive permutations of the $32\times32$ patch grid—two row rasters, two column rasters, four main diagonals, and four serpentine traversals—and always runs exactly four parallel Mamba streams, duplicating the user-chosen scan IDs so that the FLOPs and parameter count are identical to the baseline. This makes scan order an isolated experimental variable. The statistical machinery is the Friedman test applied to the Dice-score matrix of three datasets by twenty-one scan strategies.

What would settle it

Run the 21 scan strategies with, say, five random seeds on the same three datasets and compare seed-averaged Dice scores; if the best-minus-worst gap shrinks below the reported range or the top-ranked scans (the horizontal raster and the orthogonal mix) are no longer consistently on top, the claim that scan order is the decisive factor would be refuted. A second check is to apply the identical protocol to a natural-image segmentation dataset: if contiguous rasters do not outperform diagonals there, the brain-MRI specificity of the claim is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that the serialization order of image patches into a 1D sequence is a statistically significant and practically consequential design choice for Mamba-based MRI segmentation, and that spatially coherent rasters systematically beat disjointed diagonals. The author's evidence is the 21-experiment benchmark: the right-to-left horizontal scan (Experiment 3) and the orthogonal four-direction mix (Experiment 19) obtain the best Dice scores across datasets, with Experiment 19 reaching 0.820 on ISLES 2022 and 0.746 on LGG, while diagonal scans such as Experiment 7 fall to 0.551 on ISLES 2022. The Friedman test on the three-dataset-by-21-strategy matrix yields $\chi^2_{20}=43.9$ with $p=0.0016$, rejecting the null that all scan orders perform equally. The paper further claims that because MS2D keeps parameter count and runtime fixed, any observed difference can be attributed solely to the scan order.

Load-bearing premise

The load-bearing assumption is that the single Dice score measured per scan per dataset is a reliable estimate of that scan's true performance, so that the Friedman test on the three-dataset-by-21-strategy matrix, computed with only three blocks, and the reported 27-point gap are not artefacts of training noise.

Editorial extensions

If this is right

  • Practitioners deploying Mamba-based vision models for brain MRI should screen simple horizontal and vertical raster scans before any other configuration, since they deliver near-optimal accuracy at zero extra compute.
  • Scan order should be added to the standard hyperparameter search for state-space vision models, on the same footing as learning rate or patch size.
  • Diagonal and disjointed serpentine scans should be avoided for segmentation of contiguous anatomical lesions, where they can degrade Dice by tens of points.
  • The MS2D module allows future work to explore adaptive or learned scan selection without changing the model's cost profile.
  • The preference for contiguous scans may extend to other anisotropic medical modalities and high-resolution imaging tasks, though this paper only tests brain MRI.

Reading between the lines

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

  • If the ranking is robust to random-seed variation, the 27-point gap implies that scan order can be cheaply exploited as a data-dependent prior, and suggests that future Mamba variants could learn per-image scan schedules.
  • The paper's contrast with a prior remote-sensing study, where a single left-to-right scan sufficed, implies that scan-order sensitivity is tied to directional anatomical structure; testing the same 21 strategies on natural images or on CT would directly test this domain-dependence.
  • A testable extension is to measure the effective receptive field or sequence-level coherence of each scan to predict which scan will win before training, rather than benchmarking all 21.
  • Because the paper uses a single run per configuration, the exact ranking of middle-tier scans is uncertain; cross-validation with repeated seeds could shrink the gap and reorder scans below the top tier.
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

4 major / 4 minor

Summary. The paper argues that the order in which a Vision Mamba model serializes image patches into a 1D sequence is a critical, often overlooked hyperparameter for brain MRI segmentation. The authors introduce MS2D, a drop-in replacement for VMamba's SS2D module that exposes twelve primitive scan paths while keeping the parameter count and FLOPs fixed, and use it to benchmark 21 scan strategies (twelve single-direction, six bidirectional, three four-direction mixes) on BraTS 2020, ISLES 2022, and LGG. They report Dice-score spreads of up to 27 points, a Friedman test that is claimed to show statistical significance (chi-square(20)=43.9, p=0.0016), and a qualitative finding that spatially contiguous raster scans outperform disjointed diagonal scans. The paper concludes with practical guidance to treat scan order as a free hyperparameter and to prefer contiguous scans.

Significance. If the central claim were fully supported, the paper would offer a simple, zero-cost design rule for Mamba-based medical image segmentation and a reusable module for exploring scan orders. The benchmark design has real strengths: a fixed architecture and fixed computational budget across all comparisons, three public datasets covering different pathologies and imaging modalities, more than 70,000 slices, and an attempt at nonparametric significance testing. The recommendation of a shortlist of scan orders is directly actionable for practitioners. However, the current evidence is not yet sufficient to establish the headline claims, because the entire comparison rests on a single run per configuration and a statistical test computed on only three blocks. The work is a useful empirical contribution, but it needs additional experimental and statistical support before the central claims can be accepted.

major comments (4)
  1. [§III-D and Table II] Every number in Table II, including the headline 27-point ISLES spread (Exp 19 at 0.820 versus Exp 7 at 0.551), comes from a single training run per configuration. There are no repeated seeds, no confidence intervals, and no measure of run-to-run variability. Since deep segmentation training is stochastic, seed noise could easily be on the order of several Dice points, which would change the ranking of the 21 strategies and the reported gaps. The manuscript needs at least three to five seeds per configuration, with mean and standard deviation reported, and paired statistical comparisons across seeds. This is load-bearing because the central claim that scan order drives performance differences depends on the ranking being reproducible.
  2. [§V-D] The Friedman test is computed on a 3-by-21 matrix (three datasets as blocks, 21 strategies as treatments). With only three blocks, the asymptotic chi-square approximation used to obtain p=0.0016 is not reliable, and the test has no within-cell replication with which to separate scan-order effects from training noise. An exact Friedman test or a permutation-based version should be used, and ideally the analysis should be repeated across multiple seeds so that within-condition variance can be estimated. The current p-value alone does not establish that scan order is a statistically significant factor.
  3. [§III-D] The paragraph at the end of Section III-D states that 'any performance variation reported in Section V can be attributed solely to the ordering of patches.' This is not supported by the design. While the training hyperparameters are held fixed, random weight initialization, data shuffling, and other sources of stochasticity mean that observed differences conflate scan order with seed effects. The word 'solely' overstates what a single-run comparison can establish; the limitation is acknowledged only in Section VI-C as a future protocol improvement, but it should be addressed before the main claim is made.
  4. [§V-B and §V-C] The claim that spatially contiguous rasters 'consistently outperform' disjointed diagonal scans is too strong given Table II. Experiment 20, which uses all four diagonal scans, achieves 0.754 Dice on ISLES 2022, beating Experiment 1 (0.643) and several raster-based experiments, and its LGG result (0.666) is close to Experiment 1 (0.674). The qualitative pattern is plausible, but the paper should state it as a tendency with exceptions rather than a consistent ordering, or it should test the specific pairwise contrast statistically with repeated runs.
minor comments (4)
  1. [§V-B] Experiment 13 is described as a 'bidirectional horizontal scan,' but the experiment combines S1 (left-to-right) and S2 (top-to-bottom), which are a horizontal and a vertical raster, not two horizontal directions. This mislabeling should be corrected.
  2. [Abstract and §II-B] The paper calls itself the 'first systematic study' of scan order, but it cites Zhu et al. [10], who ablated 22 scan orders on aerial imagery. The novelty claim should be qualified as the first systematic study for medical imaging or for brain MRI specifically.
  3. [§IV and §V-D] The chi-square value appears as 43.9 in the abstract and as 43.86 in Section V-D. This is not a substantive issue, but the numbers should be made consistent, and the degrees of freedom and exact test variant should be stated clearly.
  4. [Overall] The contribution list promises 'enabling reproducible study,' but no code repository, configuration files, or seed values are provided. Releasing the implementation and the exact training configurations would substantially strengthen the reproducibility claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning: the scan-order benchmark is an empirical comparison with neither fitted parameters nor predictions derived from them.

full rationale

The paper's central claim is an empirical effect: twenty-one scan strategies are trained on three datasets and compared by Dice, with a Friedman test on the resulting matrix. There is no fitted parameter later renamed a prediction, and no definitional relation between the scan catalog and the measured outcome. The recommended shortlist (Exp. 3 and Exp. 19) is simply the top performers in Table II; selecting a best strategy from the same benchmark is standard practice and does not reduce the claim to its inputs. Section III-D's statement that variation is 'solely' due to scan order is a controlled-design assumption, not a circular derivation. Section VI-C candidly lists the absence of k-fold cross-validation and confidence intervals as a limitation: 'employing a more rigorous k-fold cross-validation protocol in future studies would provide tighter confidence intervals on performance metrics and further verify the robustness of the scan order rankings.' That is a threat to statistical robustness, not circularity. No load-bearing self-citation or imported uniqueness theorem appears; references to VMamba and Vision Mamba are external prior work. The paper is therefore self-contained as an empirical study and receives a circularity score of 0.

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

The central claim is entirely empirical, so the ledger captures the assumptions behind the experiments. There are no fitted parameters or invented theoretical entities. The key assumptions are the stability of single-run Dice scores, the validity of the Friedman test with three blocks, and the representativeness of the datasets and architecture.

assumptions (4)
  • domain assumption Single-run Dice scores are stable estimates of each scan order's true performance.
    All 21 experiments are run once (Section V); rankings and the 27-point gap treat these point estimates as ground truth with no variance.
  • domain assumption The Friedman test with three datasets as blocks yields a reliable p-value.
    Invoked in Section V-D; the asymptotic chi-square approximation is known to be inaccurate for as few as three blocks.
  • domain assumption The three public datasets (BraTS 2020, ISLES 2022, LGG) are representative of brain-MRI segmentation.
    Used as the evidence base; the authors acknowledge in Section VI-C that generalization to other tasks and modalities is untested.
  • domain assumption VM-UNet with MS2D is a representative state-space vision backbone for the conclusions.
    Only one architecture is tested; the authors note in Section VI-C that other Mamba backbones remain untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Flatten Wisely: How Patch Order Shapes Mamba-Powered Vision for MRI Segmentation." pith.science (2026). https://pith.science/paper/QMKHVTLG

@misc{pith2026250713384,
  author       = {Pith},
  title        = {Pith review of: Flatten Wisely: How Patch Order Shapes Mamba-Powered Vision for MRI Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QMKHVTLG}},
  note         = {Machine review of arXiv:2507.13384}
}
abstract

Vision Mamba models promise transformer-level performance at linear computational cost, but their reliance on serializing 2D images into 1D sequences introduces a critical, yet overlooked, design choice: the patch scan order. In medical imaging, where modalities like brain MRI contain strong anatomical priors, this choice is non-trivial. This paper presents the first systematic study of how scan order impacts MRI segmentation. We introduce Multi-Scan 2D (MS2D), a parameter-free module for Mamba-based architectures that facilitates exploring diverse scan paths without additional computational cost. We conduct a large-scale benchmark of 21 scan strategies on three public datasets (BraTS 2020, ISLES 2022, LGG), covering over 70,000 slices. Our analysis shows conclusively that scan order is a statistically significant factor (Friedman test: $\chi^{2}_{20}=43.9, p=0.0016$), with performance varying by as much as 27 Dice points. Spatially contiguous paths -- simple horizontal and vertical rasters -- consistently outperform disjointed diagonal scans. We conclude that scan order is a powerful, cost-free hyperparameter, and provide an evidence-based shortlist of optimal paths to maximize the performance of Mamba models in medical imaging.

Figures

Figures reproduced from arXiv: 2507.13384 by the authors.

Figure 1
Figure 1. presents an overview of the VM -UNet [19] back￾bone and the proposed MS2D integration. VM–UNet is the first pure state–space U-shaped model for medical segmen￾tation. After preprocessing, every axial slice has spatial size 128 × 128 pixels and Cin modality channels, so its raw shape is 128 × 128 × Cin. We divide the slice into non￾overlapping 4 × 4 patches, yielding a 32 × 32 grid—i.e. 32 × 32 = 1024 patch tokens—ea… view at source ↗
Figure 2
Figure 2. Overview of scan permutations and experimental design. (a) An axial ISLES slice is patched into a 3 × 3 grid; numbers mark the visitation order for the chosen scan. (b) Example sequence produced by the single left-to-right raster (S1); the remaining primitive scans S2–S12 generate analogous sequences and are omitted for space. (c) Experiment matrix. The top row depicts the 12 primitive scans (S1–S12). Experiments 1–… view at source ↗
Figure 3
Figure 3. Mean Dice scores of the 21 scan strategies on BraTS 2020, ISLES 2022, and LGG-MRI. Top: line plot; bottom: bar chart. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in 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. Partial Ring Scan: Revisiting Scan Order in Vision State Space Models

    cs.CV 2026-02 conditional novelty 6.0 of 10

    Ring-based scanning with selective channel routing improves accuracy, speed, and rotation robustness of vision state-space models.

Reference graph

Works this paper leans on

22 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Proc. MICCAI , ser. Lecture Notes in Computer Science, vol. 9351. Springer, 2015, pp. 234–241

    work page 2015

  2. [2]

    Understanding the effective receptive field in deep convolutional neural networks,

    W. Luo, Y . Li, R. Urtasun, and R. Zemel, “Understanding the effective receptive field in deep convolutional neural networks,” arXiv preprint arXiv:1701.04128, 2017

    arXiv 2017

  3. [3]

    An image is worth 16×16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16×16 words: Transformers for image recognition at scale,” in Proc. ICLR, 2021

    work page 2021

  4. [4]

    Deep learning for medical image segmentation: A comprehensive review of current techniques and future trends,

    P. Kainz, M. Pfeiffer, and M. Urschler, “Deep learning for medical image segmentation: A comprehensive review of current techniques and future trends,” Medical Image Analysis , p. 103034, 2024

    work page 2024

  5. [5]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,” arXiv preprint arXiv:2111.00396 , 2021

    arXiv 2021

  6. [6]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    arXiv 2023

  7. [7]

    Vision mamba: Efficient visual representation learning with bidirectional state space model,

    L. Zhu, B. Liao, Q. Zhang, X. Wang, W. Liu, and X. Wang, “Vision mamba: Efficient visual representation learning with bidirectional state space model,” arXiv preprint arXiv:2401.09417 , 2024

    arXiv 2024

  8. [8]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,” arXiv preprint arXiv:2401.10166, 2024

    arXiv 2024

Show all 22 references
  1. [9]

    Plainmamba: Improving non-hierarchical mamba in visual recognition,

    C. Yang, S. Yang, M. Zhang, K. Wang, J. Li, X. Bai, Y . Qiao, W. Wang, and P. Luo, “Plainmamba: Improving non-hierarchical mamba in visual recognition,” arXiv preprint arXiv:2403.17695 , 2024

    2024 arXiv

  2. [10]

    Rethinking scanning strategies with vision mamba in semantic segmentation of remote sens- ing imagery: An experimental study,

    Q. Zhu, Y . Fang, Y . Cai, C. Chen, and L. Fan, “Rethinking scanning strategies with vision mamba in semantic segmentation of remote sens- ing imagery: An experimental study,” arXiv preprint arXiv:2405.08493, 2024

    2024 arXiv

  3. [11]

    Diffusion tensor imaging of the brain,

    A. L. Alexander, J. E. Lee, M. Lazar, and A. S. Field, “Diffusion tensor imaging of the brain,” Neurotherapeutics, vol. 4, no. 3, pp. 316–329, 2007

    2007

  4. [12]

    Anisotropic diffusion tensor imaging for tracking white- matter fiber tracts,

    S. K. Lee, D.-H. Kim, J.-H. Kim, S. S. Yeo, S.-H. Kim, B.-S. Han, and K.-H. Chang, “Anisotropic diffusion tensor imaging for tracking white- matter fiber tracts,” Journal of Korean Neurosurgical Society , vol. 45, no. 5, p. 267, 2009

    2009

  5. [13]

    ISLES 2022: A multi-center mri stroke-lesion segmentation dataset,

    M. R. H. Petzsche, E. de la Rosa, U. Hanning, R. Wiest, W. E. V . Pinilla, M. Reyes, M. I. Meyer, S.-L. Liew, F. Kofler, I. Ezhov et al. , “ISLES 2022: A multi-center mri stroke-lesion segmentation dataset,” Scientific Data, vol. 9, no. 1, p. 762, 2022

    2022

  6. [14]

    The multimodal brain tumor image segmentation benchmark (brats),

    B. H. Menze, A. Jakab, S. Bauer, J. Kalpathy-Cramer, K. Farahani, J. Kirby et al. , “The multimodal brain tumor image segmentation benchmark (brats),” IEEE Transactions on Medical Imaging , vol. 34, no. 10, pp. 1993–2024, 2015

    1993

  7. [15]

    Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features,

    S. Bakas, H. Akbari, A. Sotiras, M. Bilello, M. Rozycki, J. S. Kirby et al., “Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features,”Scientific Data, vol. 4, p. 170117, 2017

    2017

  8. [16]

    Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge,

    S. Bakas, M. Reyes, A. Jakab, S. Bauer, M. Rempfler, A. Crimi et al., “Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge,” arXiv preprint arXiv:1811.02629 , 2018

    2018 arXiv

  9. [17]

    The cancer genome atlas low grade glioma collection (tcga-lgg) (version 3) [data set],

    N. Pedano, A. E. Flanders, L. Scarpace, T. Mikkelsen, J. M. Eschbacher, B. Hermes, V . Sisneros, J. Barnholtz-Sloan, and Q. Ostrom, “The cancer genome atlas low grade glioma collection (tcga-lgg) (version 3) [data set],” The Cancer Imaging Archive , 2016

    2016

  10. [18]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation,

    J. Ma, F. Li, and B. Wang, “U-mamba: Enhancing long-range dependency for biomedical image segmentation,” arXiv preprint arXiv:2401.04722, 2024

    2024 arXiv

  11. [19]

    Vm-unet: Vision mamba unet for medical image segmentation,

    J. Ruan and S. Xiang, “Vm-unet: Vision mamba unet for medical image segmentation,” arXiv preprint arXiv:2402.02491 , 2024

    2024 arXiv

  12. [20]

    Msvm-unet: Multi-scale vision mamba unet for medical image segmentation,

    J. Chen, J. Wang, J. Wu, and D. Z. Chen, “Msvm-unet: Multi-scale vision mamba unet for medical image segmentation,” arXiv preprint arXiv:2408.13735, 2024

    2024 arXiv

  13. [21]

    Lkm-unet: Large kernel vision mamba unet for medical image segmentation,

    J. Wang, J. Chen, D. Z. Chen, and J. Wu, “Lkm-unet: Large kernel vision mamba unet for medical image segmentation,” arXiv preprint arXiv:2403.07332, 2024

    2024 arXiv

  14. [22]

    Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,

    Z. Xing, T. Ye, Y . Yang, G. Liu, and L. Zhu, “Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation,” arXiv preprint arXiv:2401.13560, 2024

    2024 arXiv

Pith tools

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