Pith. sign in

REVIEW 4 major objections 4 minor 19 references

This paper claims that learned, data-dependent scan ordering — Dynamic Adaptive Scan in a 3D Mamba U-Net — matches a fixed tri-orientated Mamba segmenter at 13× lower parameter count and beats it when scaled up.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 12:52 UTC pith:XIVKE2N5

load-bearing objection A cleanly reported architecture paper whose central DAS-vs-ToM claim is not actually isolated—but the 13x parameter reduction at near-parity is real and worth a referee's time. the 4 major comments →

arxiv 2607.22718 v1 pith:XIVKE2N5 submitted 2026-07-21 cs.CV cs.AI

DAMamba-UNet3D: A Parameter-Efficient Mamba State Space U-Net with Dynamic Adaptive Scan for 3D Medical Image Segmentation

Mohammad Arafat Hussain , Ellen Grant , Yangming Ou This is my paper
classification cs.CV cs.AI
keywords dynamic adaptive scanstate space modelMamba3D medical image segmentationbrain tumor segmentationparameter efficiencyU-NetBraTS 2020
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that a U-Net for 3D medical image segmentation can get the long-range context benefits of Mamba-style state-space models without paying for enormous parameter budgets, by letting the model learn the order in which it scans the volume. The authors build a hybrid U-Net where only the encoder's middle three stages contain Mamba blocks that reorder voxels using learned, data-dependent offsets (Dynamic Adaptive Scan) applied across three anatomical planes, while the rest of the network stays convolutional. On BraTS 2020 brain tumor segmentation, the compact version (5.3M parameters) nearly matches a fixed-scan Mamba segmenter of about 70M parameters, and a wider version surpasses it. The central claim is that adaptive scan ordering is a viable, parameter-efficient alternative to hand-designed fixed scan patterns for volumetric segmentation.

Core claim

The paper's central discovery is that Dynamic Adaptive Scan — a learning-based reordering of the voxel sequence fed to a selective state-space model — can be lifted to 3D by averaging scans over axial, coronal, and sagittal planes, and that this tri-plane formulation works well enough to make a 5.3M-parameter U-Net competitive with a 70M-parameter fixed-scan Mamba segmenter (0.815 vs 0.824 mean Dice on BraTS 2020). With a wider variant matched to the baseline's scale, the adaptive-scan model exceeds the fixed-scan model (0.829 vs 0.824). The authors present this as the first application of DAS to medical segmentation and the first 3D extension of DAS, and they argue that scan order, not just

What carries the argument

The central mechanism is the tri-plane Dynamic Adaptive Scan (DAS3D) block. DAS predicts 2D offsets per position in each of the three anatomical planes (axial, coronal, sagittal), samples features with bilinear interpolation, averages the three reordered maps, and then runs the standard selective scan of the state-space model along the resulting 1D sequence. This turns the scan trajectory from a fixed design choice into a data-dependent operation that costs O(N) time per block, preserving the efficiency of Mamba while adding adaptability. An important secondary finding is that placing these blocks only in the encoder stages E2–E4, with convolutions at the bottleneck and decoder, is critical

Load-bearing premise

The load-bearing premise is that averaging three 2D adaptive scans produces a meaningful 3D scan order; the paper offers no evidence about the learned offsets, so if they collapse to identity or random noise the adaptive mechanism is inert and the reported differences come from other architectural choices.

What would settle it

Train the identical DAMamba-UNet3D architecture but replace the tri-plane DAS with a fixed tri-orientated flattening (same placement, same parameter count, same pipeline). If mean Dice stays at 0.815 or the gap to the baseline disappears, the adaptive mechanism is not responsible for the result. Alternatively, dump the learned offsets on a few test volumes and check whether they systematically deviate from the identity grid and focus on tumor boundaries.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Near-parity with a 13× larger fixed-scan model shows that global context can be bought with learned scan order rather than parameter count.
  • The encoder-only placement principle — no SSM blocks at bottleneck or decoder — gives a concrete design rule for hybrid U-Nets.
  • DAMamba-L's +0.5 pt Dice gain over the fixed-scan baseline suggests learned scanning can outperform fixed multi-orientation flattening when capacity is matched.
  • The O(N) cost of the adaptive scan preserves Mamba's linear-time advantage, making the approach usable on high-resolution 3D volumes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the paper never inspects the learned offsets, the 'adaptive' part may be contributing less than the architectural differences (block placement, width, stacking); a direct swap of DAS for a fixed scan inside the identical model would isolate the effect.
  • The tri-plane averaging is a simple 2D-per-plane lifting; a genuinely 3D DAS (e.g., volumetric deformable sampling) might show larger gains, especially for non-axis-aligned tumor geometry.
  • If the adaptive ordering is truly learning meaningful tumor-first scanning, the learned offsets should be visualizable as concentration near lesion boundaries; that is a testable prediction of the paper's mechanism.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes DAMamba-UNet3D, a hybrid 3D U-Net in which encoder stages E2–E4 contain tri-plane Dynamic Adaptive Scan (DAS) Mamba blocks and convolutions are retained elsewhere. On BraTS 2020 five-fold full-volume evaluation, the compact model (~5.3M parameters) attains 0.815±0.013 mean Dice versus 0.824±0.014 for retrained SegMamba (~70M); a large variant DAMamba-L (~70M) reaches 0.829±0.012, 0.5 pt higher than SegMamba. The authors claim the first medical-segmentation DAS architecture and the first 3D extension of DAS, and argue that learned tri-plane DAS is competitive with and may improve upon fixed Tri-orientated Mamba (ToM) scanning.

Significance. If the central claim were established, this work would provide a parameter-efficient alternative to fixed-scan Mamba segmenters and a meaningful 3D extension of an adaptive scan operator. The paper has notable strengths: it uses a standard public benchmark (BraTS 2020), a clearly described five-fold full-volume evaluation protocol, and transparent reporting of training settings. However, the evidence for the central DAS-versus-ToM claim is confounded by the scan-pattern ablation design, and the paper itself concedes that scan protocol and architecture are not isolated. As presented, the significance of the claimed adaptive-scan advantage is not yet demonstrated; what is demonstrated is that a particular hybrid U-Net with encoder-only DAS blocks is competitive at low parameter count.

major comments (4)
  1. [§3.2, Table 4] The scan-pattern ablation conflates scan type with bottleneck removal. The caption states that fixed orders use 'auxiliary E2–E4 + bottleneck layout,' while the final learned-DAS row removes the bottleneck. The only matched comparison — learned DAS with bottleneck (0.807) versus tri-orientated with bottleneck (0.808) — shows no gain from DAS. The +0.7 pt difference between learned DAS (final, 0.815) and tri-orientated (0.808) therefore cannot be attributed to DAS; it is attributable to removing the bottleneck block. The Discussion admits this ('We do not isolate scan protocol from architecture in a single swap experiment'), but the Abstract and Conclusion assert a stronger claim. Please supply a same-layout fixed-versus-DAS comparison, e.g., fixed scans with E2–E4 only, and report paired significance tests.
  2. [§3.2 text] The sentence 'learned DAS ahead of fixed orders when encoder placement is matched' is directly contradicted by Table 4's matched row: learned DAS with bottleneck is 0.807, tri-orientated with bottleneck is 0.808. Either the claim must be corrected to state that DAS is ahead only when the bottleneck is removed, or new matched experiments are needed. This is a report-of-results error, not merely an interpretation issue.
  3. [§4, Discussion and Table 2] The claim that DAMamba-L 'surpasses SegMamba ... suggesting that learned DAS in an encoder-only hybrid design can outperform fixed ToM' is unsupported as stated. DAMamba-L differs from SegMamba in U-Net depth, block stacking, and bottleneck design, as the authors acknowledge. The improvement, even if real, could come entirely from those architectural differences. To support the attribution to DAS, the paper needs either a scale-matched ablation that varies only the scan mechanism, or a much more cautious conclusion limited to the full DAS-native hybrid architecture.
  4. [§2.6, §3, Table 3/4] The key numerical comparisons (0.815 vs 0.824, 0.829 vs 0.824, 0.807 vs 0.808, 0.808 vs 0.815) are all within one standard deviation or less, and no significance testing is reported. Since the protocol uses five subject-level folds, paired tests (e.g., Wilcoxon or paired t on fold Dice) are straightforward and would materially strengthen the claims. The Limitations section correctly notes that statistical testing is needed, but the abstract and conclusion nevertheless draw positive conclusions without it.
minor comments (4)
  1. [Abstract vs. Conclusion] The abstract provides a GitHub code URL, while the Conclusion states 'We will publish our code upon acceptance of this paper.' Please reconcile these statements and make the availability status unambiguous.
  2. [Eq. (4)] The tri-plane DAS operation is defined only as an average of 2D DAS on axial, coronal, and sagittal planes. It would clarify whether the offsets are shared or separately learned for each plane, and how the 2D operator is adapted when the plane dimensions differ (e.g., D×W for coronal). This affects the reader's ability to reproduce the method.
  3. [Eq. (5)] The loss equation uses per-class weights w_i^(c) but does not define the class index c or the range of i. A short definition would help reproducibility.
  4. [References] Reference [7] lists an incomplete author list ('Zhou, Y., Zhang, Y., et al.'), and reference [8] contains a broken URL fragment ('https–www'). Please fix these formatting issues.

Circularity Check

0 steps flagged

No circularity: empirical architecture evaluation against external public benchmark; acknowledged confounds are experimental limitations, not circular reasoning.

full rationale

DAMamba-UNet3D is an empirical architecture paper, not a derivation-from-inputs result. Its method is built by adopting the externally published DAS operator (Zhou et al., DAMamba, ref [7]) and extending it to tri-plane 3D scanning; the authors explicitly disclaim redesign of the core DAS mechanism ('we adopt the 2D DAS operator from [7]'). The central claims are evaluated against the public BraTS 2020 benchmark and against a retrained external baseline (SegMamba) under identical splits, augmentation, loss, optimizer, and epoch budget, so the numbers are self-contained empirical evidence rather than consequences of definitions. No parameter is fitted to a subset and then reported as a prediction; no uniqueness theorem or load-bearing self-citation is imported; no known result is renamed. The paper's own caveats are experimental limitations, not circularity: 'We do not isolate scan protocol from architecture in a single swap experiment' (Sec. 4), and Table 4's fixed-scan rows use an auxiliary E2-E4 + bottleneck layout, so the comparison of the final E2-E4-only DAS model to fixed scans is confounded. This weakens the causal claim about DAS versus ToM, but it does not make the claim true by construction or reduce any output to an input. The limitations paragraph also notes that multi-dataset validation and statistical testing are needed. Those are validity concerns, not circularity; the correct circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No physical constants are fitted and no new physical entities are introduced. The central claim relies on domain assumptions about the validity of tri-plane DAS as a 3D reordering mechanism, about BraTS as a benchmark, and about the transferability of placement ablations across model scales.

free parameters (4)
  • DAS offset scale/regularization = not reported
    The paper does not state the magnitude bound or regularization for the learned offsets; presumably inherited from DAMamba [7]. If offsets degenerate, the adaptive mechanism may be inert.
  • SSM state dimension = 16
    Set to the standard setting in Mamba [5] and DAMamba [7]; chosen by hand, not fitted to BraTS.
  • Loss modulation exponent gamma = 2
    Chosen in the loss definition (Eq. 5); hand-selected.
  • Base width C = 16 (compact), SegMamba-aligned widths (DAMamba-L)
    Architecture choice that controls the parameter count; the compact width is chosen for efficiency.
axioms (4)
  • domain assumption BraTS 2020 five-fold cross-validation is representative enough of 3D medical segmentation quality to support the comparative claims.
    All claims rest on a single publicly available brain-tumor dataset; generalization to other organs/modalities is not tested.
  • ad hoc to paper Tri-plane averaging of 2D DAS produces a valid and useful 3D adaptive scan order.
    This is the core methodological invention (Eq. 4), but no evidence is given that the averaged 2D offsets approximate true 3D reordering.
  • standard math The bilinear interpolation used in DAS is differentiable and stable enough for end-to-end training.
    Standard practice in deformable/adaptive sampling literature, but not analyzed here.
  • domain assumption Encoder-only placement, validated on the compact model, transfers to DAMamba-L.
    DAMamba-L uses encoder-only DAS based on the compact-model ablation; no direct large-scale placement ablation is run.

pith-pipeline@v1.3.0-alltime-deepseek · 6491 in / 10728 out tokens · 100726 ms · 2026-08-01T12:52:35.514610+00:00 · methodology

0 comments
read the original abstract

We propose parameter-efficient SSM-based U-Net architectures for 3D medical image segmentation. Convolutional U-Nets afford O(n) local mixing per layer but lack explicit global context; transformers provide global reasoning at O(n^2) cost in sequence length $n$. State-space models (SSMs), such as Mamba, offer $O(n)$ global propagation per block. Yet, existing medical SSM segmenters rely on fixed scan patterns and large parameter budgets. Dynamic Adaptive Scan (DAS), which learns data-dependent reordering before selective scan, has not been applied to medical imaging or extended to 3D volumes. We propose DAMamba-UNet3D, a hybrid encoder-decoder that integrates tri-plane 3D-DAS blocks at encoder stages E2-E4 while retaining convolutions elsewhere (~5.3M parameters). On BraTS 2020 five-fold cross-validation, DAMamba-UNet3D achieves mean Dice 0.815+/-0.013 (full-volume per-case evaluation) at ~13x lower parameter cost than SegMamba (0.824+\-0.014, ~70M). At comparable scale, DAMamba-L (~70M), a wide DAS-native variant with encoder-only DAMamba and a convolutional bottleneck, reaches 0.829+\-0.012, surpassing retrained SegMamba by 0.5pt. Component ablations show that encoder-only DAS placement is critical as bottleneck and decoder SSM blocks lower Dice. Together, the results suggest that learned tri-plane DAS in a hybrid U-Net is competitive with, and under our large-scale design may improve upon, SegMamba's fixed Tri-orientated Mamba (ToM) scanning on BraTS 2020. Code: https://github.com/marafathussain/DAMamba-UNet3D.

Figures

Figures reproduced from arXiv: 2607.22718 by Ellen Grant, Mohammad Arafat Hussain, Yangming Ou.

Figure 1
Figure 1. Figure 1: DAMamba-UNet3D block diagram with stage-wise feature dimensions (D×H×W×C; example shown for 1283 training patches). Blue: convolutional blocks (E1, decoder); orange: encoder stages E2–E4, each convolutional block followed by a tri-plane DAMamba block. 2. Tri-plane DAS. The 2D DAS operator is applied on axial (H×W), coronal (D×W), and sagittal (D×H) planes and averaged: DAS3D(X) = 1 3 [PITH_FULL_IMAGE:figu… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative full-volume predictions on BraTS 2020 validation subjects. Broader impact. Regardless of the BraTS scores above, extending DAS to tri-plane 3D volumes and applying it to medical segmentation for the first time is a substantive contribution in its own right. It offers a learnable alternative to fixed scan trajectories that can be reused in other volumetric tasks without redesigning the core oper… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 5 linked inside Pith

  1. [1]

    IEEE transactions on medical imaging , volume=

    The multimodal brain tumor image segmentation benchmark (BRATS) , author=. IEEE transactions on medical imaging , volume=

  2. [2]

    International MICCAI brainlesion workshop , pages=

    Multidimensional and multiresolution ensemble networks for brain tumor segmentation , author=. International MICCAI brainlesion workshop , pages=. 2019 , organization=

    2019

  3. [3]

    International MICCAI Brainlesion Workshop , pages=

    3D brain tumor segmentation and survival prediction using ensembles of convolutional neural networks , author=. International MICCAI Brainlesion Workshop , pages=. 2020 , organization=

    2020

  4. [4]

    International MICCAI Brainlesion Workshop , pages=

    Cerberus: A multi-headed network for brain tumor segmentation , author=. International MICCAI Brainlesion Workshop , pages=. 2020 , organization=

    2020

  5. [5]

    International MICCAI brainlesion workshop , pages=

    Uncertainty-driven refinement of tumor-core segmentation using 3d-to-2d networks with label uncertainty , author=. International MICCAI brainlesion workshop , pages=. 2020 , organization=

    2020

  6. [6]

    International MICCAI Brainlesion Workshop , pages=

    The BraTS 2020 challenge , author=. International MICCAI Brainlesion Workshop , pages=. 2020 , organization=

    2020

  7. [7]

    The journal of machine learning for biomedical imaging , volume=

    QU-BraTS: MICCAI BraTS 2020 challenge on quantifying uncertainty in brain tumor segmentation-analysis of ranking scores and benchmarking results , author=. The journal of machine learning for biomedical imaging , volume=

    2020

  8. [8]

    Nature methods , volume=

    nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation , author=. Nature methods , volume=

  9. [9]

    European conference on computer vision , pages=

    Swin-Unet: Unet-like pure transformer for medical image segmentation , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  10. [10]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=

    TransBTS: Multimodal brain tumor segmentation using transformer , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , pages=. 2021 , organization=

    2021

  11. [11]

    arXiv preprint arXiv:2312.00752 , year=

    Mamba: Linear-time sequence modeling with selective state spaces , author=. arXiv preprint arXiv:2312.00752 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2401.10166 , year=

    VMamba: Visual state space model , author=. arXiv preprint arXiv:2401.10166 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2401.09417 , year=

    Vision mamba: Efficient visual representation learning with bidirectional state space model , author=. arXiv preprint arXiv:2401.09417 , year=

    Pith/arXiv arXiv

  14. [14]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , year=

    SegMamba: Long-range sequential modeling mamba for 3D medical image segmentation , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , year=

  15. [15]

    arXiv preprint arXiv:2502.12627 , year=

    DAMamba: Vision state space model with dynamic adaptive scan , author=. arXiv preprint arXiv:2502.12627 , year=

    Pith/arXiv arXiv

  16. [16]

    International conference on medical image computing and computer-assisted intervention , pages=

    3D U-Net: learning dense volumetric segmentation from sparse annotation , author=. International conference on medical image computing and computer-assisted intervention , pages=. 2016 , organization=

    2016

  17. [17]

    International Conference on Medical image computing and computer-assisted intervention , pages=

    U-net: Convolutional networks for biomedical image segmentation , author=. International Conference on Medical image computing and computer-assisted intervention , pages=. 2015 , organization=

    2015

  18. [18]

    Computerized Medical Imaging and Graphics , volume=

    Combo loss: Handling input and output imbalance in multi-organ segmentation , author=. Computerized Medical Imaging and Graphics , volume=. 2019 , publisher=

    2019

  19. [19]

    arXiv preprint arXiv:2506.15744 , year=

    Pixel-wise modulated Dice loss for medical image segmentation , author=. arXiv preprint arXiv:2506.15744 , year=

    Pith/arXiv arXiv