Pith. sign in

REVIEW 4 major objections 5 minor 24 references

Johnson-Lindenstrauss Lemma Guided Network for Efficient 3D Medical Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A network derives convolution group sizes from the Johnson–Lindenstrauss lemma, not from pruning.

desk verdict A solid lightweight 3D segmentation architecture with strong results, but the JL-based 'theoretical' group-size bound is an empirical fit, and several reported numbers need reconciliation. read the letter →

arxiv 2509.22307 v2 pith:T4SKF4RG submitted 2025-09-26 cs.CV

Jinpeng Lu , Linghan Cai , Yinda Chen , Guo Tang , Songhan Jiang , Haoyuan Shi , Zhiwei Xiong This is my paper · ORCID
classification cs.CV
keywords 3DmedicalimagesegmentationJohnson–LindenstrausslemmalightweightnetworkgroupedconvolutionpairedwindowattentionmultimodalfusionGrammatrixknowledgetransferefficiencyvsrobustness
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

VeloxSeg is an attempt to dissolve the “efficiency/robustness conflict” in lightweight 3D medical segmentation. The central claim is that the Johnson–Lindenstrauss lemma, which says finitely many high-dimensional points can be embedded in $O(\log N)$ dimensions while roughly preserving pairwise distances, also fixes the minimum group size of grouped convolutions: $C_{\mathrm{group}} \ge c_{\mathrm{JL}}\,\varepsilon^{-2}\log N(\mathcal{M},v)$. Because the exact manifold coverage count $N$ is unknown, the paper substitutes the empirical approximation $(\mathcal{M}\cdot v)^\alpha$, calibrates $\alpha$ on AutoPET-II, and uses the resulting group sizes $\{n,2n,2n,4n\}$ at the four network stages. If this holds, spatial adjacency in feature space survives at very low parameter counts, so the model does not need data-specific pruning to stay robust. On the reported benchmarks the recipe yields 1.66 million parameters, 1.79 GFLOPs, 62.51% Dice on AutoPET-II, and 91.44% Dice on BraTS2021.

What carries the argument

The load-bearing object is Johnson–Lindenstrauss-guided convolution (JLC): a grouped convolution whose per-stage group size is set by $C_{\mathrm{group}} = d' \ge c_{\mathrm{JL}}\varepsilon^{-2}\log N(\mathcal{M},v)$, with $N(\mathcal{M},v)$ approximated as $(\mathcal{M}\cdot v)^\alpha$. This formula plays the role that pruning-importance metrics play in other lightweight designs: it decides how many channels share each weight tensor. The lemma’s logarithmic dependence on the point count is what converts a crude coverage estimate into a small integer group size, and the volume-ratio exponents $\{4/3,8/3,16/3,32/3\}$ are what turn that bound into the concrete pattern $\{n,2n,2n,4n\}$. The same formula also fixes the minimum head size in the Paired Window Attention, tying the two streams together.

What would settle it

Train the same dual-stream backbone on a new dataset with a very different volume-ratio profile (for example, small-lesion whole-body PET/CT with more modalities), and compare the JL-derived $\{n,2n,2n,4n\}$ schedule against a grid of alternative group ratios at matched parameter count; if a non-JL ratio wins consistently, the lower-bound derivation is not determining the result. A more direct check is to estimate $N(\mathcal{M},v)$ empirically on AutoPET-II features by covering the feature manifold with $\varepsilon$-balls at each stage and testing whether $(\mathcal{M}\cdot v)^\alpha$ lies anywhere near the true $\log N(\mathcal{M},v)$ that the lemma requires.

Watch

Extended reading notes

Core claim

The paper’s discovery claim is that grouped-convolution width can be chosen by theory rather than by pruning or grid search. Starting from the JL lemma, the authors argue that each voxel in the feature map must preserve distances between points on the segmentation-related manifold $\mathcal{M}$ of the input patch; covering that manifold with $N(\mathcal{M},v)$ points forces the embedding dimension $d'$ to be at least $c_{\mathrm{JL}}\varepsilon^{-2}\log N(\mathcal{M},v)$, and they identify $d'$ with the number of channels per convolution group, $C_{\mathrm{group}}$. With $N$ replaced by the empirical $(\mathcal{M}\cdot v)^\alpha$, the volume ratios $\{4/3,8/3,16/3,32/3\}$ across stages produce the stage-wise group sizes $\{n,2n,2n,4n\}$. The paper then assembles VeloxSeg around this JLC convolution, a paired-window attention that handles multi-scale and cross-modal context at near-linear cost, and a Gram-matrix knowledge-transfer loss that adds texture detail at no inference cost. The intended upshot is that robustness to complex anatomy and heterogeneous modalities does not require large models.

Load-bearing premise

The argument's load-bearing premise is that the empirical coverage count $\hat{N}(\mathcal{M},v)=(\mathcal{M}\cdot v)^\alpha$, with a single task-difficulty exponent $\alpha$, faithfully replaces the unknown manifold coverage count $N(\mathcal{M},v)$, and that a grouped convolution's channel count per group can be identified with the JL embedding dimension $d'$.

Editorial extensions

If this is right

  • If JLC’s group sizes are doing the claimed work, lightweight segmentation can skip pruning cycles entirely: the same $\{n,2n,2n,4n\}$ pattern transfers across PET/CT and MRI benchmarks instead of being re-tuned per dataset.
  • The reported 62.51% Dice on AutoPET-II at 1.66 million parameters and 1.79 GFLOPs, with a 14.2% Dice gain over a self-configuring baseline at 1.87% of its parameters, implies the efficiency/robustness conflict is not a fixed trade-off.
  • PWA’s multimodal interaction adds only 0.27 million parameters and 0.09 GFLOPs while raising Dice by 5.75% on AutoPET-II, so modal cooperation can be cheap when attention operates on paired windows.
  • SDKT’s Gram-matrix matching is the only transfer path in the comparison with positive gain (62.51% versus 57.15% for shared-ROI guidance), meaning texture priors can be injected without any inference-time cost.

Reading between the lines

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

  • If the empirical coverage count $(\mathcal{M}\cdot v)^\alpha$ is what actually determines the winning group sizes, then the JL lemma is functioning as a justification for a scaling heuristic; a direct estimate of $N(\mathcal{M},v)$ by epsilon-covering real feature manifolds would tell whether the logarithmic bound is the mechanism or a post-hoc fit.
  • The same lower-bound logic could be applied to any layer with a discrete width that must be fixed before seeing data—attention heads, expert routing, or feature-map channels in other modalities—giving a parameter-free initialization principle that does not require pruning.
  • A stronger test of the theory would fix $\alpha$ from one dataset’s manifold statistics and predict group sizes on another dataset with very different lesion sizes; the current paper calibrates $n$ on AutoPET-II and then transfers it, which conflates theory with a tuned hyperparameter.
  • The BraTS2021 result (91.44% Dice with 1.46 million parameters) suggests that when lesions are large and centralized, even a weaker local feature extractor suffices, so the robustness gain of JL-guided groups should show up mainly in small-lesion, heterogeneous datasets.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The manuscript proposes VeloxSeg, a lightweight dual-stream CNN-Transformer architecture for 3D medical image segmentation. Its three claimed contributions are Paired Window Attention (PWA) for multi-scale cross-modal attention, Johnson-Lindenstrauss lemma-guided convolution (JLC) to set group sizes that preserve spatial adjacency, and Spatially Decoupled Knowledge Transfer (SDKT) via Gram matrices to distill texture priors from a self-supervised teacher. The authors report strong efficiency (1.66M parameters, 1.79 GFLOPs on AutoPET-II) and competitive Dice scores on AutoPET-II, Hecktor2022, BraTS2021, and MSD2019, and they release code.

Significance. If the empirical results hold, VeloxSeg is a useful efficient baseline that compares favorably with many existing methods at a fraction of the compute, and the PWA and SDKT ideas are plausible and worth further study. The release of code and the breadth of experiments across four public datasets are clear strengths. However, the central 'theoretical' JL contribution is not actually derived: the group sizes are fitted to the target dataset, and several headline quantitative claims in the abstract are either undefined or unsupported by the experimental section. The architecture's empirical value does not depend on the JL lemma, and the paper would be significantly strengthened by honestly re-framing JLC as a heuristic design choice and by fully documenting all claimed efficiency numbers.

major comments (4)
  1. [Section 2.3, Eq. (2), Appendix E] The claimed 'theoretical lower bound' on group size is not derived. Equation (2) requires a finite set X of size N in the JL lemma, but the manuscript never constructs the segmentation-related manifold or gives a covering bound for N(M,v). Section 2.3 itself states: 'Due to the lack of N in the vision domain, we empirically approximate N(M,v) using Nhat(M,v)=(M·v)^α, where α is related to the difficulty of the segmentation task at hand.' Since α (or n) is selected by ablation on AutoPET-II (Table 2), the final group sizes {n,2n,2n,4n} are fitted values, not independent predictions. Moreover, Appendix E's Eq. (13) for M=2 predicts group sizes {4.9α, 6.9α, 9.0α, 11.1α}, yet the paper adopts {4α, 8α, 8α, 16α}; this is a change of shape, not merely rounding. The paper should either supply a genuine covering argument and verify the JL bound, or re-frame JLC as a JL-inspired heuristic and remove the wording 'theoretically determines a minimum group size' from the contributions.
  2. [Abstract; Section 3.3; Appendix H] Several headline quantitative claims are unsupported or undefined. The abstract states a '26% Dice improvement' with no stated baseline; in Table 1 the margin over the best basic model is 0.05 percentage points (62.51 vs 62.46 for VSmTrans), and the margin over the best lightweight model is about 5.3 points, while Appendix H reports a 14.2-point improvement over nnUNet (70.05 vs 55.85). The abstract also claims reductions in training peak GPU memory by 1/20 and inference memory by 1/24, but no table or text in the paper reports any memory measurements. These claims need to be either precisely defined with baselines or removed.
  3. [Table 1] The DINOv3-L rows contain physically impossible negative HD95 values (-6.85 and -7.50). A Hausdorff distance cannot be negative. The table footnote '—' does not explain these entries. Since Table 1 is the central quantitative comparison, these entries must be corrected or the DINOv3-L results removed from the comparison; otherwise readers cannot trust the evaluation.
  4. [Table 2; Appendix K.1; Section 3.4] The group-size ablation does not support the claim that the JL configuration is optimal. In the conv-only rows of Table 2, the configuration {4,8,8,16} achieves 55.84% Dice, while the larger {16,32,64,128} achieves 56.20% Dice, so the selected setting is not the best even in that ablation. In addition, Appendix K.1 states that the 'Optimal group channel setting ⟨4,8,8,16⟩ achieved a 55.14% Dice', but Table 2 reports 55.84% for {4,8,8,16} and 55.14% for {8,16,16,32}. The text in Section 3.4 also reports a best Dice of 54.14% for n=4 in the JLC robustness experiment, which is inconsistent with Table 2. These discrepancies must be reconciled, and the robustness argument needs to be formulated in terms of the accuracy/efficiency trade-off rather than claiming dominance.
minor comments (5)
  1. [Appendix F] One paragraph refers to 'the effectiveness of SlimMSCT', which appears to be a leftover from a different paper; this should be corrected to VeloxSeg.
  2. [Figure 5 caption] The caption contains a typo: '(c (d))' should read '(c) (d)'.
  3. [References] The reference entry 'Determine Filters’Importance. Pruning filters for efficient convnets' is malformed; the author names are missing and it should be attributed to Li et al.
  4. [Section 2.1] The abbreviation 'SDTK' is used once for Spatially Decoupled Knowledge Transfer; elsewhere it is SDKT. Please make the abbreviation consistent.
  5. [Table 6] The CPU throughput values for HCMA-UNet and U-RWKV are listed as '−' because they lack CPU support; this should be stated explicitly in the table caption or main text to avoid confusion.

Circularity Check

3 steps flagged · score 6.0 of 10

The JL-guided group-size 'lower bound' is fixed by a fitted power law and an AutoPET-II ablation; the {4,8,8,16} configuration is a tuned hyperparameter, not a JL prediction.

  1. fitted input called prediction [Section 2.3, Eq. (2)]
    "Due to the lack of N in the vision domain, we empirically approximate N(M,v) using N̂(M,v) = (M·v)^α, where α is related to the difficulty of the segmentation task at hand. We conduct ablation studies on datasets with the richest modality heterogeneity and data distribution to identify the most generalizable scaling factor, which we use to obtain a lower bound on the group size of the convolution layers in each network stage."

    Eq. (2) is the paper's claimed lower bound for the convolution group size. Its right-hand side depends on N(M,v), which is not computed anywhere. The paper substitutes N̂(M,v)=(M·v)^α, with α 'related to the difficulty of the segmentation task at hand' and selected by ablations on the benchmarks. With that substitution the bound becomes C_group ≥ c_JL α ε^-2 (log M + log v); only the logarithmic-in-(M,v) functional form remains from the JL lemma, while the value of α (hence the recommended group size) is fitted. Presenting this as a 'theoretical lower bound' is a fitted parameter renamed as a prediction.

  2. fitted input called prediction [Appendix E and Table 2 / Appendix K.1]
    "Considering that the group size needs to be divisible by the total number of channels and that nonlinear networks have stronger compression capabilities than linear networks, we use {C^k_group}^4_{k=1} = {4α,8α,8α,16α} for each stage, where α is determined by the most difficult AutoPET-II dataset to ensure universality. For convenience, we replace α with n=⌈α/4⌉∈N, and the final convolution group size of each stage of the network is set to {C^k_group}^4_{k=1} = {n,2n,2n,4n}."

    Even after the power-law approximation, Appendix E's own computation for M=2 (Eq. 13) gives {4.9,6.9,9.0,11.1}α, not the adopted {4,8,8,16}α. The paper changes the shape for channel divisibility and then determines n on AutoPET-II; in the group-size sweep of Table 2, ⟨4,8,8,16⟩ is highlighted as the chosen 'optimal group channel setting' (Appendix K.1). The final four-stage configuration is therefore selected by measured Dice and efficiency on the target dataset, not forced by Eq. (2).

1 more flagged steps
  1. self definitional [Section 2.3, Eq. (2)]
    "Cgroup = d′ ≥ c_JL ε−2 log N(M,v), (2) where C_group is the number of channels per group."

    This is the definitional bridge of the JL claim. The JL lemma bounds d′, the dimension of a linear embedding that approximately preserves pairwise distances among N points. Eq. (2) simply labels the number of channels per grouped convolution as d′. No argument connects learned convolution-group channels to a JL embedding of the 'segmentation-related manifold'; the identification C_group = d′ is asserted, not derived. Combined with the fitted N(M,v), the group-size 'prediction' reduces to a definition plus a data fit.

full rationale

VeloxSeg's empirical segmentation evaluation is not circular: Dice is reported on held-out AutoPET-II and Hecktor2022 test splits, on BraTS2021, and on the external MSD2019 set, so the measured performance of the architecture is an independent result. The PWA and SDKT components are also validated by module ablations rather than by construction. The circularity is confined to the JL-guided convolution claim. In Eq. (2), C_group ≥ c_JL ε^-2 log N(M,v) cannot be evaluated because N(M,v) is never specified; the paper replaces it with N̂=(M·v)^α and fits α to the benchmark datasets. Appendix E then replaces the shape implied by Eq. (13) with {4,8,8,16}α for channel divisibility, and n is selected from the AutoPET-II group-size sweep in Table 2. Thus the 'theoretically determined minimum group size' is, by the paper's own account, an ablated hyperparameter: the JL lemma supplies only a logarithmic scaling form, with all constants and stage ratios coming from data. The only apparent self-citation, Liu et al. (2025) for the DINOv3 evaluation, is not load-bearing. Because the central theoretical claim reduces to a fit while the segmentation results stand independently, the appropriate score is 6 rather than 0 or 8.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The signature theoretical contribution rests on an empirical approximation: the manifold coverage count N is replaced by (M·v)^α with α tuned on AutoPET-II, so the JL lower bound becomes a fitted group-size schedule rather than a parameter-free derivation. The remaining components rely on standard assumptions: finite coverage of the image manifold, an analogy between convolution groups and JL target dimensions, and the established equivalence of Gram-matrix matching to MMD. No new physical entities are introduced.

free parameters (4)
  • α (JL group-size exponent) = not stated; implied by n=4 via α≈4n
    The manifold coverage count is approximated as Nhat=(M·v)^α; α is tuned via ablation on AutoPET-II (Table 2), making the claimed JL lower bound empirical rather than derived.
  • n (group-size base) = 4
    The stage group sizes {n,2n,2n,4n} become {4,8,8,16}; n=4 is selected as best on AutoPET-II in Table 2, then applied to other datasets.
  • λ_rc (reconstruction loss weight) = 0.5
    Reconstruction loss weight in SDKT, tuned on AutoPET-II (Table 11).
  • λ_style (Gram-matrix loss weight) = 2.0
    Gram-matrix distillation loss weight, tuned on AutoPET-II (Table 11).
assumptions (5)
  • standard math Johnson-Lindenstrauss lemma
    Invoked in Lemma 1 and Section 2.3 as the basis for the lower bound on group size.
  • domain assumption The segmentation-related manifold of a 3D image patch can be covered by a finite number N(M,v) of samples
    Assumed in Section 2.3 to connect volume ratio v to a point count N; no construction or bound is given.
  • domain assumption The mapping implemented by a grouped convolution layer behaves like the JL linear embedding, so the target dimension d' corresponds to the channel group size
    This analogy is the bridge between JL and C_group; the paper does not prove that convolution features preserve pairwise distances in the JL sense.
  • ad hoc to paper The coverage count N is well approximated by Nhat(M,v)=(M·v)^α with a single α across datasets
    This is an empirical ansatz fitted on AutoPET-II, stated in Section 2.3, and it determines the claimed lower bound.
  • domain assumption Gram-matrix matching approximates MMD with a second-order polynomial kernel and thus transfers texture knowledge
    Used in Section 2.4; follows from prior work (Li et al. 2017) but applied here to reconstruction-to-segmentation transfer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Johnson-Lindenstrauss Lemma Guided Network for Efficient 3D Medical Segmentation." pith.science (2026). https://pith.science/paper/T4SKF4RG

@misc{pith2026250922307,
  author       = {Pith},
  title        = {Pith review of: Johnson-Lindenstrauss Lemma Guided Network for Efficient 3D Medical Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4SKF4RG}},
  note         = {Machine review of arXiv:2509.22307}
}
abstract

Lightweight 3D medical image segmentation remains constrained by a fundamental \textit{``efficiency / robustness conflict''}, particularly when processing complex anatomical structures and heterogeneous modalities. In this paper, we study how to redesign the framework based on the characteristics of high-dimensional 3D images, and explore data synergy to overcome the fragile representation of lightweight methods. Our approach, VeloxSeg, begins with a deployable and extensible dual-stream CNN-Transformer architecture composed of Paired Window Attention (PWA) and Johnson-Lindenstrauss lemma-guided convolution (JLC). For each 3D image, we invoke a ``glance-and-focus'' principle, where PWA rapidly retrieves multi-scale information, and JLC ensures robust local feature extraction with minimal parameters, significantly enhancing the model's ability to operate with low computational budget. Followed by an extension of the dual-stream architecture that incorporates modal interaction into the multi-scale image-retrieval process, VeloxSeg efficiently models heterogeneous modalities. Finally, Spatially Decoupled Knowledge Transfer (SDKT) via Gram matrices injects the texture prior extracted by a self-supervised network into the segmentation network, yielding stronger representations than baselines at no extra inference cost. Experimental results on multimodal benchmarks show that VeloxSeg achieves a 26\% Dice improvement, alongside increasing GPU throughput by 11$\times$, CPU by 48$\times$, and reducing training peak GPU memory usage by $1/20$, inference by $1/24$. Code is available at https://github.com/JinPLu/VeloxSeg.

Figures

Figures reproduced from arXiv: 2509.22307 by the authors.

Figure 1
Figure 1. Comparison of our proposed VeloxSeg with recent methods on the AutoPET-II test set. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of VeloxSeg. VeloxSeg employs an encoder-decoder architecture with Paired [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Overview of Paired Window Attention (PWA). (b) Intuitive difference between depth￾wise (DW) convolution and Johnson-Lindenstrauss (JL) guided Convolution in the feature space. 2 METHODOLOGY 2.1 OVERVIEW OF VELOXSEG As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Visualization of model decoding and t-SNE plot. (a)-(b) show results without atten￾tion; (c)-(d) with attention. In (a) and (c), “Low Met.” and “High Met.” represent low/high￾metabolism PET regions. In (b) and (d), CT background with red tumor outline. The visualizatio…
Figure 6
Figure 6. Figure 6: Dice performance comparison between differ [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Box plots of Dices at different lesion vol [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Detailed architecture of Paired Window Attention (PWA). This figure focuses on visually [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: 3D qualitative visualization of different methods on the AutoPET-II, Hecktor2022 and [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 11 canonical work pages

  1. [5]

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

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752,

    arXiv

  2. [7]

    On the expressive power of self-attention matrices.arXiv preprint arXiv:2106.03764,

    Valerii Likhosherstov, Krzysztof Choromanski, and Adrian Weller. On the expressive power of self-attention matrices.arXiv preprint arXiv:2106.03764,

    arXiv

  3. [9]

    Our model improves the Dice by 1.72% compared to the state-of-the-art SuperLightNet

    Since brain tumors are large and centralized, and the slices processed are relatively fixed, almost all models achieved good results. Our model improves the Dice by 1.72% compared to the state-of-the-art SuperLightNet. 23 Hecktor 2022 AutoPET-II CT+GT PET Slim UNETR VSmTrans UNet H-DenseFormer Ours HCMA-UNet SuperLightNet U-RWKV A2FSeg CT+GT Nestedformer ...

    work page arXiv 2022

  4. [10]

    Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756, 2024b

    Ziming Liu et al. Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756, 2024b. Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    arXiv

  5. [13]

    Rwkv: Reinventing rnns for the transformer era.arXiv preprint arXiv:2305.13048,

    Bo Peng et al. Rwkv: Reinventing rnns for the transformer era.arXiv preprint arXiv:2305.13048,

    arXiv

  6. [14]

    Super-resolution and infection edge detection co-guided learning for covid-19 ct segmentation

    Yu Sang, Jinguang Sun, Simiao Wang, Heng Qi, and Keqiu Li. Super-resolution and infection edge detection co-guided learning for covid-19 ct segmentation. InICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1665–1669,

    work page 2021

  7. [15]

    Abdelrahman M Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan

    doi: 10.1109/ICASSP39728.2021.9414327. Abdelrahman M Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Unetr++: delving into efficient and accurate 3d medical image seg- mentation.IEEE Transactions on Medical Imaging,

    arXiv 2021

  8. [17]

    A large annotated medical image dataset for the development and evaluation of segmentation algorithms.arXiv preprint arXiv:1902.09063,

    Amber L Simpson et al. A large annotated medical image dataset for the development and evaluation of segmentation algorithms.arXiv preprint arXiv:1902.09063,

    arXiv 1902

Show all 24 references
  1. [18]

    Instance normalization: The missing in- gredient for fast stylization.arXiv preprint arXiv:1607.08022,

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing in- gredient for fast stylization.arXiv preprint arXiv:1607.08022,

    arXiv

  2. [19]

    Sam-med3d: A vision foundation model for general-purpose segmentation on volumetric medical images.IEEE Transactions on Neural Net- works and Learning Systems, 2025a

    Haoyu Wang, Sizheng Guo, Jin Ye, Zhongying Deng, Junlong Cheng, Tianbin Li, Jianpin Chen, Yanzhou Su, Ziyan Huang, Yiqing Shen, et al. Sam-med3d: A vision foundation model for general-purpose segmentation on volumetric medical images.IEEE Transactions on Neural Net- works and ...

    2021

  3. [20]

    U-rwkv: Lightweight medical image segmentation with direction-adaptive rwkv

    Hongbo Ye et al. U-rwkv: Lightweight medical image segmentation with direction-adaptive rwkv. arXiv preprint arXiv:2507.11415,

    arXiv

  4. [21]

    README.md

    15 APPENDIX Table of content: • §A: Reproducibility Statement • §B: Use of Large Language Models (LLMs) • §C: Necessity of Multi-Scale Attention • §D: Details of PW A • §E: Details of JL-guided Group Size • §F: Dataset Details • §G: Details of computational performance • §H: R...

    1989

  5. [22]

    Notably,r,B k 1 , andS k 1 are closely related to the computational cost, and the specific settings for different datasets are given in Appendix D.3

    Paired Window Mixer */ //A:[M, C, H, W, D] A←E+Dropout(PWC(A)) returnA 18 As shown in Figure 8, given them-th modal feature of thek-th encoder stage,E k m ∈ RCk×H k×W k×Dk , we need to first compute a set of ordered paired window sizes Wink i N k win i=1 , whereN k win is the ...

  6. [23]

    MP.”: Million Parameters; “GF

    G DETAILS OF COMPUTATIONAL PERFORMANCE We evaluate the computational performance of VeloxSeg against other leading models on the AutoPET-II, Hecktor2022, and BraTS2021 datasets. Our analysis focus on four key metrics: the number of model parameters in millions, GFLOPs, GPU thr...

  7. [1984]

    Does dinov3 set a new medical vision standard?arXiv preprint arXiv:2509.06467,

    Che Liu, Yinda Chen, Haoyuan Shi, Jinpeng Lu, Bailiang Jian, Jiazhen Pan, Linghan Cai, Jiayi Wang, Yundi Zhang, Jun Li, et al. Does dinov3 set a new medical vision standard?arXiv preprint arXiv:2509.06467,

  8. [2016]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:...

    2010 arXiv

  9. [2017]

    3d u-net: learning dense volumetric segmentation from sparse annotation

    ¨Ozg¨un C ¸ ic ¸ek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. InMedical Image Comput- ing and Computer-Assisted Intervention–MICCAI 2016: 19th International Conference, Athen...

    2016

  10. [2019]

    The multimodal brain tumor image segmentation benchmark (brats).IEEE transactions on medical imaging, 34(10):1993–2024,

    Bjoern H Menze et al. The multimodal brain tumor image segmentation benchmark (brats).IEEE transactions on medical imaging, 34(10):1993–2024,

    1993

  11. [2020]

    Towards lightweight hyper- spectral image super-resolution with depthwise separable dilated convolutional network.arXiv preprint arXiv:2505.00374,

    Usman Muhammad, Jorma Laaksonen, and Lyudmila Mihaylova. Towards lightweight hyper- spectral image super-resolution with depthwise separable dilated convolutional network.arXiv preprint arXiv:2505.00374,

    arXiv

  12. [2021]

    Super convergence cosine annealing with warm-up learning rate

    12 Zhao Liu. Super convergence cosine annealing with warm-up learning rate. InCAIBDA 2022; 2nd International Conference on Artificial Intelligence, Big Data and Algorithms, pp. 1–7,

    2022

  13. [2022]

    The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification.arXiv preprint arXiv:2107.02314,

    Ujjwal Baid et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification.arXiv preprint arXiv:2107.02314,

    2021 arXiv

  14. [2023]

    Pruning filters for efficient convnets.arXiv preprint arXiv:1608.08710,

    Determine Filters’Importance. Pruning filters for efficient convnets.arXiv preprint arXiv:1608.08710,

    arXiv

  15. [2024]

    Dinov3.arXiv preprint arXiv:2508.10104,

    Oriane Sim´eoni, Huy V V o, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Micha ¨el Ramamonjisoa, et al. Dinov3.arXiv preprint arXiv:2508.10104,

    arXiv

  16. [2025]

    Hcma-unet: A hybrid cnn-mamba unet with axial self-attention for efficient breast cancer segmentation.arXiv preprint arXiv:2501.00751,

    Haoxuan Li, Peiwu Qin, Xi Yuan, and Zhenglin Chen. Hcma-unet: A hybrid cnn-mamba unet with axial self-attention for efficient breast cancer segmentation.arXiv preprint arXiv:2501.00751,

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.