pith. sign in

arxiv: 2605.18835 · v1 · pith:65ODMHV6new · submitted 2026-05-13 · 💻 cs.LG

StampFormer: A Physics-Guided Material-Geometry-Coupled Multimodal Model for Rapid Prediction of Physical Fields in Sheet Metal Stamping

Jiajie Luo , Mohamed Mohamed , Osama Hassan , Haosu Zhou , Yingxue Zhao , Haoran Li , Xinrun Li , Zhutao Shao
show 3 more authors
Yang Long Nan Li Jichun Li
This is my paper

Pith reviewed 2026-05-20 20:42 UTC · model grok-4.3

classification 💻 cs.LG
keywords sheet metal stampingsurrogate modeldeep learningphysical field predictionmultimodal learningfinite element analysismaterial propertiesgeometry coupling
0
0 comments X

The pith

StampFormer fuses geometry and material stress-strain data to predict stamping physical fields in under a second.

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

The paper develops StampFormer to replace slow finite element simulations with fast predictions of detailed physical fields during sheet metal stamping design. It processes both the part geometry and the material's stress-strain response together rather than treating them separately. This joint processing aims to deliver accurate maps of thinning, strains, and displacement that designers can use for quick checks on whether a part can be formed without defects. A reader would care because current design loops are limited by the hours or days needed for each simulation run.

Core claim

StampFormer is a physics-guided multimodal framework that takes component geometry and material stress-strain responses as inputs to predict FEA outcomes. It first fuses the two data types in a Material-Augmented Geometric Network, then injects the combined information at multiple scales through a Hierarchical Material Embedding Injection Unit before feeding it to an adapted Swin-UNet backbone. On two simulation datasets for a crossmember panel in steel and aluminium, the model produces thinning, major strain, minor strain, plastic strain, and displacement fields in under a second with average relative error below 8.5 percent on the 2D fields and mean squared error below 1.2 mm squared on 3D

What carries the argument

Material-Augmented Geometric Network that fuses geometry with material stress-strain curves before hierarchical injection into the network backbone

If this is right

  • Designers can run many more geometry variants in the same time previously used for one full simulation.
  • The model supplies complete field maps instead of single scalar quality metrics.
  • The same architecture handles both steel and aluminium without separate models.
  • Real-time manufacturability feedback becomes possible inside CAD tools.

Where Pith is reading between the lines

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

  • If accuracy holds for new shapes, the approach could be retrained on a broader library of parts to cover entire vehicle programs.
  • The same fusion pattern might apply to other manufacturing simulations such as forging or injection molding.
  • Running the model inside an optimization loop could automatically suggest geometry changes that reduce forming problems.

Load-bearing premise

Data from simulations of one crossmember panel shape in steel and aluminium is enough for the model to work accurately on other part shapes and materials.

What would settle it

Test the trained model on finite element results for a different part geometry such as an automotive door or hood and check whether the relative errors on the physical fields remain below 8.5 percent.

Figures

Figures reproduced from arXiv: 2605.18835 by Haoran Li, Haosu Zhou, Jiajie Luo, Jichun Li, Mohamed Mohamed, Nan Li, Osama Hassan, Xinrun Li, Yang Long, Yingxue Zhao, Zhutao Shao.

Figure 1
Figure 1. Figure 1: Workflow of traditional sheet metal forming design and the proposed surrogate [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the steel stress-strain dataset evolution. The dataset is expanded [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the aluminium stress-strain dataset before and after augmentation. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The key geometric features for the proposed panel. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the 3D-to-2D conversion process for model inputs and targets. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The FEA model setup [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FE simulation stages used in the forming workflow: gravity loading, blank [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Simulation results showing the effective plastic strain distribution for one of the [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 6
Figure 6. Figure 6: Simulation results showing the thickness distribution for one of the matrix cases. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overall architecture of the StampFormer model. The framework consists of three main components: the MAGN for [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The structure of a Swin Transformer block, which consists of Layer Normaliza [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Histograms showing the distribution of RE computed from the representative [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Histograms showing the distribution of MSE (mm [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Visualization for thinning field prediction on five representative cases from the [PITH_FULL_IMAGE:figures/full_fig_p023_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Visualization for thinning field prediction on five representative cases from [PITH_FULL_IMAGE:figures/full_fig_p023_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of FLDs from the steel dataset. The top two rows display the [PITH_FULL_IMAGE:figures/full_fig_p024_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Comparison of FLDs from the aluminium dataset. The top two rows display [PITH_FULL_IMAGE:figures/full_fig_p024_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Visualization of displacement-reconstructed surfaces colored by plastic strain [PITH_FULL_IMAGE:figures/full_fig_p025_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Visualization of displacement-reconstructed surfaces colored by plastic strain [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: The thinning field distribution of a random sample (where negative values [PITH_FULL_IMAGE:figures/full_fig_p026_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Comparison of surfaces reconstructed from displacement and colored by plastic [PITH_FULL_IMAGE:figures/full_fig_p027_20.png] view at source ↗
read the original abstract

Traditional sheet metal forming relies on time-consuming and expensive Finite Element Analysis (FEA) for design validation, a process that significantly prolongs design cycles. While surrogate models offer faster iteration, current approaches have limitations: scalar-based methods cannot capture comprehensive field-based FEA results, while existing image-based models often ignore the critical role of material properties by focusing solely on geometry. To address this gap, we develop a physics-guided deep learning framework, namely StampFormer, which simultaneously uses component geometry and material stress-strain responses to predict FEA outcomes. The StampFormer framework uses three core components to process data. A Material-Augmented Geometric Network (MAGN) first fuses geometric and material data. This information is then integrated at various levels by a Hierarchical Material Embedding Injection Unit (HMEIU) before being processed by the primary network backbone, an adapted Swin-UNet. We evaluated our model on the stamping of a crossmember panel with two simulation datasets for steel and aluminium panels, and results demonstrate that StampFormer provides high-fidelity predictions of critical physical fields - including thinning, major strain, minor strain, plastic strain, and displacement - in under a second. Compared with ground truth FEA, our model achieved an average relative error of less than 8.5% on the four 2D fields and a mean squared error of less than 1.2 mm2 for the 3D displacement field. In summary, we introduce a practical and efficient framework that integrates multimodal information, namely geometry and material properties, to provide fast and accurate predictions, enabling designers to perform real-time manufacturability assessments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The manuscript introduces StampFormer, a multimodal deep learning model for rapid prediction of physical fields (thinning, major strain, minor strain, plastic strain, and 3D displacement) in sheet metal stamping. It fuses geometry and material stress-strain responses via a Material-Augmented Geometric Network (MAGN), Hierarchical Material Embedding Injection Unit (HMEIU), and adapted Swin-UNet backbone, reporting average relative error below 8.5% on 2D fields and MSE below 1.2 mm² on displacement for FEA simulations of a single crossmember panel using steel and aluminum, with inference under one second.

Significance. If the reported accuracy generalizes beyond the evaluated geometry and materials, the framework could meaningfully accelerate design validation cycles in manufacturing by replacing slow FEA with fast surrogate predictions, supporting real-time manufacturability checks.

major comments (3)
  1. [Evaluation / Results] Evaluation is confined to two simulation datasets for one crossmember panel geometry (steel and aluminum). This does not test the model's ability to maintain the claimed error bounds (<8.5% relative error, <1.2 mm² MSE) on unseen part shapes or boundary conditions, which is load-bearing for the central claim of enabling real-time assessment across sheet metal stamping.
  2. [Methods / Experiments] No baseline comparisons to existing scalar-based or image-based surrogate models, nor ablation studies isolating the contributions of MAGN and HMEIU, are provided. This makes it impossible to quantify the benefit of the proposed multimodal physics-guided fusion over standard supervised training on FEA labels.
  3. [Abstract / Methods] The abstract and methods describe the model as 'physics-guided' yet provide no explicit enforcement mechanism (e.g., physics-informed loss terms, constraints, or residual penalties) beyond end-to-end supervised training on FEA-generated labels; the approach appears purely data-driven.
minor comments (1)
  1. [Abstract] The abstract states 'average relative error of less than 8.5%' without specifying per-field breakdowns or confidence intervals; adding these would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive comments on our manuscript. We address each of the major comments below and outline the revisions we will make to improve the paper.

read point-by-point responses

  1. Referee: [Evaluation / Results] Evaluation is confined to two simulation datasets for one crossmember panel geometry (steel and aluminum). This does not test the model's ability to maintain the claimed error bounds (<8.5% relative error, <1.2 mm² MSE) on unseen part shapes or boundary conditions, which is load-bearing for the central claim of enabling real-time assessment across sheet metal stamping.

    Authors: We recognize the importance of evaluating generalization to unseen geometries and boundary conditions for the broader applicability of the model. The current study focuses on a representative industrial crossmember panel, which includes complex features such as varying thicknesses and curvatures encountered in automotive stamping. The multimodal architecture is designed to handle diverse inputs through the fusion of geometry and material properties. However, we agree that additional validation on different part shapes would strengthen the claims. In the revised manuscript, we will expand the discussion section to explicitly address the limitations regarding generalization and outline plans for future work on multi-geometry datasets. We will also attempt to include preliminary results on a simpler benchmark geometry if feasible with available simulation resources. revision: partial

  2. Referee: [Methods / Experiments] No baseline comparisons to existing scalar-based or image-based surrogate models, nor ablation studies isolating the contributions of MAGN and HMEIU, are provided. This makes it impossible to quantify the benefit of the proposed multimodal physics-guided fusion over standard supervised training on FEA labels.

    Authors: We appreciate this suggestion and agree that quantitative comparisons and ablations are essential to demonstrate the advantages of our approach. In the revised version, we will add baseline comparisons against a standard Swin-UNet trained solely on geometric inputs and against simpler convolutional models. Additionally, we will perform ablation studies by removing the MAGN and HMEIU components to isolate their impact on prediction accuracy. These additions will help quantify the benefits of the material-geometry coupling. revision: yes

  3. Referee: [Abstract / Methods] The abstract and methods describe the model as 'physics-guided' yet provide no explicit enforcement mechanism (e.g., physics-informed loss terms, constraints, or residual penalties) beyond end-to-end supervised training on FEA-generated labels; the approach appears purely data-driven.

    Authors: We thank the referee for pointing out this potential ambiguity in terminology. In our work, 'physics-guided' refers to the integration of physical material constitutive behavior (via stress-strain curves) as multimodal inputs to inform the geometric processing, thereby embedding domain-specific physical knowledge into the model. This is distinct from physics-informed neural networks that incorporate PDE residuals into the loss function. We will revise the abstract, introduction, and methods sections to clarify this usage and better distinguish our data-driven multimodal approach from explicit physics-constrained methods. revision: yes

Circularity Check

0 steps flagged

Supervised ML surrogate reports FEA-matched errors on single-geometry data without definitional reduction

full rationale

The paper describes an end-to-end trained neural architecture (MAGN + HMEIU + Swin-UNet) whose outputs are compared to FEA labels generated for one crossmember panel. Performance numbers (relative error <8.5 %, MSE <1.2 mm²) are standard test-set metrics from supervised fitting; no equation or claimed first-principles result is shown to equal its own inputs by construction, and no load-bearing self-citation chain is invoked. The work is therefore self-contained against its external FEA benchmark, yielding only minor circularity from the usual ML evaluation loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on supervised learning from FEA labels; the model weights are free parameters fitted to the two panel datasets. No new physical axioms or invented entities are introduced beyond standard neural network assumptions.

free parameters (1)
  • model weights and hyperparameters
    All network parameters are fitted to the steel and aluminium stamping simulation data to achieve the reported error levels.
axioms (1)
  • domain assumption FEA simulations provide accurate ground-truth labels for training
    The paper treats FEA outputs as reliable targets without discussing numerical convergence or modeling assumptions in the simulations.

pith-pipeline@v0.9.0 · 5867 in / 1333 out tokens · 27444 ms · 2026-05-20T20:42:01.552274+00:00 · methodology

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

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    Blala, L

    H. Blala, L. Lang, S. Khan, L. Li, S. Sijia, A. Guelailia, S. A. Slimane, S. Alexandrov, Forming challenges of small and complex fiber metal laminate parts in aerospace applications—a review, Int. J. Adv. Manuf. Technol. 126 (5) (2023) 2509–2543

    work page 2023

  2. [2]

    Więckowski, M

    W. Więckowski, M. Motyka, J. Adamus, P. Lacki, M. Dyner, Numer- ical and experimental analysis of titanium sheet forming for medical instrument parts, Materials 15 (5) (2022) 1735

    work page 2022

  3. [3]

    H. R. Attar, H. Zhou, A. Foster, N. Li, Rapid feasibility assessment of components to be formed through hot stamping: A deep learning approach, J. Manuf. Process. 68 (2021) 1650–1671. 29

    work page 2021

  4. [4]

    A. K. Perka, M. John, U. B. Kuruveri, P. L. Menezes, Advanced high- strength steels for automotive applications: Arc and laser welding pro- cess, properties, and challenges, Metals 12 (6) (2022) 1051

    work page 2022

  5. [5]

    Chandel, N

    R. Chandel, N. Sharma, S. A. Bansal, A review on recent develop- ments of aluminum-based hybrid composites for automotive applica- tions, Emerg. Mater. 4 (5) (2021) 1243–1257

    work page 2021

  6. [6]

    L. Hua, W. Zhang, H. Ma, Z. Hu, Investigation of formability, mi- crostructures and post-forming mechanical properties of heat-treatable aluminum alloys subjected to pre-aged hardening warm forming, Int. J. Mach. Tools Manuf. 169 (2021) 103799

    work page 2021

  7. [7]

    H. R. Attar, N. Li, A. Foster, A new design guideline development strat- egy for aluminium alloy corners formed through cold and hot stamping processes, Mater. Des. 207 (2021) 109856

    work page 2021

  8. [8]

    Multidiscip

    S.Li, D.Zhou, A.Pan, Integratedlightweightoptimizationdesignofwall thickness, material, and performance of automobile body side structure, Struct. Multidiscip. Optim. 67 (6) (2024) 95

    work page 2024

  9. [9]

    Alawadhi, S

    I. Alawadhi, S. Ramnath, A. Bolar, Y. Fu, J. J. Shah, N. Zurbrugg, D. Detwiler, Structural Design Vs Manufacturability Costs of Complex Stamped Components, in: Proc. Int. Des. Eng. Tech. Conf. Comput. Inf. Eng. Conf. (IDETC/CIE), Vol. 88391, 2024, p. V005T05A008

    work page 2024

  10. [10]

    H. Li, H. Zhou, N. Li, An integrated convolutional neural network- based surrogate model for crashworthiness performance prediction of hot-stamped vehicle panel components, in: MATEC Web Conf., Vol. 401, 2024, p. 03013

    work page 2024

  11. [11]

    H. Zhou, Y. Zhao, H. Li, T. Pfaff, N. Li, A multi-level graph-based surrogate model for real-time high-fidelity sheet forming simulations, Adv. Eng. Inform. 66 (2025) 103458

    work page 2025

  12. [12]

    Cantamessa, F

    M. Cantamessa, F. Montagna, G. D’Agnese, Data-driven innovation: challenges and insights of engineering design, Des. Sci. 10 (2024) e11

    work page 2024

  13. [13]

    I.K.Nti, A.F.Adekoya, B.A.Weyori, O.Nyarko-Boateng, Applications of artificial intelligence in engineering and manufacturing: a systematic review, J. Intell. Manuf. 33 (6) (2022) 1581–1601. 30

    work page 2022

  14. [14]

    A. T. G. Tapeh, M. Z. Naser, Artificial intelligence, machine learning, and deep learning in structural engineering: a scientometrics review of trends and best practices, Arch. Comput. Methods Eng. 30 (1) (2023) 115–159

    work page 2023

  15. [15]

    Taheri, A

    H. Taheri, A. S. Beni, Artificial Intelligence, Machine Learning, and Smart Technologies for Nondestructive Evaluation, in: Handbook of Nondestructive Evaluation 4.0, Springer, 2025, pp. 853–881

    work page 2025

  16. [16]

    A. Ucar, M. Karakose, N. Kırımça, Artificial intelligence for predictive maintenance applications: key components, trustworthiness, and future trends, Appl. Sci. 14 (2) (2024) 898

    work page 2024

  17. [17]

    Grover, AI-Enabled Supply Chain Optimization, Int

    N. Grover, AI-Enabled Supply Chain Optimization, Int. J. Adv. Res. Sci. Commun. Technol. (2025) 28–44

    work page 2025

  18. [18]

    Kardovskyi, S

    Y. Kardovskyi, S. Moon, Artificial intelligence quality inspection of steel bars installation by integrating mask R-CNN and stereo vision, Autom. Constr. 130 (2021) 103850

    work page 2021

  19. [19]

    Senoner, T

    J. Senoner, T. Netland, S. Feuerriegel, Using explainable artificial intel- ligence to improve process quality: evidence from semiconductor manu- facturing, Manage. Sci. 68 (8) (2022) 5704–5723

    work page 2022

  20. [20]

    S. J. Plathottam, A. Rzonca, R. Lakhnori, C. O. Iloeje, A review of artificial intelligence applications in manufacturing operations, J. Adv. Manuf. Process. 5 (3) (2023) e10159

    work page 2023

  21. [21]

    H. Zhou, Q. Xu, Z. Nie, N. Li, A study on using image-based machine learning methods to develop surrogate models of stamp forming simula- tions, J. Manuf. Sci. Eng. 144 (2) (2022) 021012

    work page 2022

  22. [22]

    Attar, H

    H. Attar, H. Zhou, N. Li, Deformation and thinning field prediction for HFQ formed panel components using convolutional neural networks, in: IOP Conf. Ser.: Mater. Sci. Eng., Vol. 1157, IOP Publishing, 2021, p. 012079

    work page 2021

  23. [23]

    M.Cheng, X.Zhao, M.Dhimish, W.Qiu, S.Niu, Areviewofdata-driven surrogate models for design optimization of electric motors, IEEE Trans. Transp. Electrific. 10 (4) (2024) 8413–8431. 31

    work page 2024

  24. [24]

    Alizadeh, J

    R. Alizadeh, J. K. Allen, F. Mistree, Managing computational com- plexity using surrogate models: a critical review, Res. Eng. Des. 31 (3) (2020) 275–298

    work page 2020

  25. [25]

    C. Ling, W. Kuo, M. Xie, An overview of adaptive-surrogate-model- assisted methods for reliability-based design optimization, IEEE Trans. Reliab. 72 (3) (2022) 1243–1264

    work page 2022

  26. [26]

    Hashemi, J

    A. Hashemi, J. Jang, J. Beheshti, A machine learning-based surrogate finite element model for estimating dynamic response of mechanical sys- tems, IEEE Access 11 (2023) 54509–54525

    work page 2023

  27. [27]

    Y. Shi, Z. Lu, J. Zhou, E. Zio, A novel time-dependent system constraint boundary sampling technique for solving time-dependent reliability- based design optimization problems, Comput. Methods Appl. Mech. Eng. 372 (2020) 113342

    work page 2020

  28. [28]

    Hart-Rawung, J

    T. Hart-Rawung, J. Buhl, M. Bambach, A fast approach for optimiza- tion of hot stamping based on machine learning of phase transformation kinetics, Procedia Manuf. 47 (2020) 707–712

    work page 2020

  29. [29]

    Jankovič, M

    D. Jankovič, M. Šimic, N. Herakovič, A data-driven simulation and Gaussian process regression model for hydraulic press condition diag- nosis, Adv. Eng. Inform. 59 (2024) 102276

    work page 2024

  30. [30]

    J. Kim, C. Lee, Prediction of turbulent heat transfer using convolutional neural networks, J. Fluid Mech. 882 (2020) A18

    work page 2020

  31. [31]

    K. Ren, Y. Chew, Y. F. Zhang, J. Y. H. Fuh, G. J. Bi, Thermal field prediction for laser scanning paths in laser aided additive manufacturing by physics-based machine learning, Comput. Methods Appl. Mech. Eng. 362 (2020) 112734

    work page 2020

  32. [32]

    Q. Tian, S. Guo, E. Melder, L. Bian, W. G. Guo, Deep learning-based data fusion method for in situ porosity detection in laser-based additive manufacturing, J. Manuf. Sci. Eng. 143 (4) (2021) 041011

    work page 2021

  33. [33]

    R. Azad, E. K. Aghdam, A. Rauland, Y. Jia, A. H. Avval, A. Bozorg- pour, S. Karimijafarbigloo, J. P. Cohen, E. Adeli, D. Merhof, Medical image segmentation review: The success of u-net, IEEE Trans. Pattern Anal. Mach. Intell. (2024). 32

    work page 2024

  34. [34]

    Z. Luo, W. Yang, Y. Yuan, R. Gou, X. Li, Semantic segmentation of agriculturalimages: Asurvey, Inf.Process.Agric.11(2)(2024)172–186

    work page 2024

  35. [35]

    J. W. Kim, A. U. Khan, I. Banerjee, Systematic review of hybrid vision transformer architectures for radiological image analysis, J. Imaging In- form. Med. (2025) 1–15

    work page 2025

  36. [36]

    X. Liu, C. Zhang, L. Zhang, Vision mamba: A comprehensive survey and taxonomy, arXiv:2405.04404 (2024)

    work page arXiv 2024

  37. [37]

    Kumar, S

    A. Kumar, S. P. Yadav, A. Kumar, An improved feature extraction al- gorithm for robust Swin Transformer model in high-dimensional medical image analysis, Comput. Biol. Med. 188 (2025) 109822

    work page 2025

  38. [38]

    Z. Zhu, M. Sun, G. Qi, Y. Li, X. Gao, Y. Liu, Sparse dynamic volume TransUNet with multi-level edge fusion for brain tumor segmentation, Comput. Biol. Med. 172 (2024) 108284

    work page 2024

  39. [39]

    D. Jain, R. Modzelewski, R. Hérault, C. Chatelain, E. Torfeh, S. Thureau, Differential-UMamba: Rethinking tumor segmentation un- der limited data scenarios, arXiv:2507.18177 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  40. [40]

    H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, M. Wang, Swin-unet: Unet-like pure transformer for medical image segmentation, in: Proc. Eur. Conf. Comput. Vis. (ECCV), 2022, pp. 205–218. 33

    work page 2022