pith. sign in

arxiv: 2605.16393 · v1 · pith:MKS7EZY3new · submitted 2026-05-12 · 💻 cs.CV · cs.AI

Vision Transformer-Conditioned UNet for Domain-Adaptive Semantic Segmentation

Joel Valdivia Ortega , Tingying Peng , Marion Jasnin This is my paper

Pith reviewed 2026-05-20 22:38 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords semantic segmentationvision transformerUNetdomain adaptationbiomedical imagingMRICTmedical image analysis
0
0 comments X

The pith

ViTC-UNet conditions a UNet on frozen Vision Transformer features through learnable tokens and two-way attention to improve biomedical semantic segmentation without retraining the transformer.

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

The paper presents ViTC-UNet to close the performance gap that Vision Transformers show on biomedical semantic segmentation of sparse, fine-structured targets in MRI and CT data. It does so by feeding representations from a frozen pre-trained ViT into a UNet decoder via learnable tokens and a two-way attention mechanism. The design keeps the ViT fixed while letting the UNet supply the local inductive bias and high-resolution output that lightweight ViT decoders often miss. This hybrid setup transfers large-scale visual priors to medical images even when the source and target domains differ, and it reports better accuracy than standard baselines on the tested modalities.

Core claim

ViTC-UNet conditions a UNet on frozen pre-trained ViT representations through learnable tokens and a two-way attention decoder. This combines ViT global visual priors with the local inductive bias and high-resolution decoding capacity of UNets, while avoiding end-to-end ViT fine-tuning even in cross-domain settings.

What carries the argument

ViTC-UNet, a decoder that routes frozen ViT features into a UNet via learnable tokens and two-way attention to generate high-precision biomedical masks.

If this is right

  • The method yields higher segmentation accuracy than baseline UNet and ViT decoders on MRI and CT data.
  • Large-scale visual priors from ViTs can be reused across imaging modalities without retraining the transformer backbone.
  • High-resolution local decoding remains available even when the source model stays frozen.
  • Cross-domain adaptation becomes feasible with lower compute cost than full fine-tuning.

Where Pith is reading between the lines

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

  • The same conditioning pattern could be tested on other dense tasks such as instance segmentation or depth estimation in medical volumes.
  • Freezing the ViT opens a route to deploy large vision models on modest clinical hardware while still benefiting from their priors.
  • Extending the two-way attention to multiple frozen ViT layers might further strengthen the transfer of mid-level features.

Load-bearing premise

Lightweight ViT pixel decoders lack enough local bias for precise medical masks, and learnable tokens plus two-way attention can transfer useful global priors from a frozen ViT without any end-to-end retraining.

What would settle it

A controlled experiment on the same MRI and CT test sets in which a plain UNet or an end-to-end fine-tuned ViT decoder matches or exceeds ViTC-UNet accuracy when the conditioning tokens and two-way attention are removed.

Figures

Figures reproduced from arXiv: 2605.16393 by Joel Valdivia Ortega, Marion Jasnin, Tingying Peng.

Figure 1
Figure 1. Figure 1: Information flow of ViTC-UNet: a multi-modal framework for semantic segmentation. The [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ViTC-UNet. A ViT generates image embeddings which are integrated [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of segmentation performance across (a) CT and (b) MRI modalities. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of segmentation performance between a linear baseline, the ViT-UNet hybrid [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
read the original abstract

Semantic segmentation is essential for analysing anatomical features in biomedical research, yet a performance gap remains for Vision Transformers (ViTs) in the field, particularly for sparse, fine-structured, and low signal-to-noise targets. We attribute this challenge in part to the lightweight pixel decoders commonly used in promptable ViT models, who may lack the local inductive bias needed for high-precision biomedical masks. We bridge this gap by introducing ViTC-UNet, which conditions a UNet on frozen pre-trained ViT representations through learnable tokens and a two-way attention decoder. This combines ViT global visual priors with the local inductive bias and high-resolution decoding capacity of UNets, while avoiding end-to-end ViT fine-tuning even in cross-domain settings. ViTC-UNet outperforms baseline results in semantic segmentation tasks across MRI and CT modalities, demonstrating that structure-conditioned UNet decoding can efficiently adapt large-scale visual priors to high-complexity biomedical segmentation.

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

0 major / 2 minor

Summary. The paper proposes ViTC-UNet, an architecture that conditions a UNet-style decoder on frozen pre-trained Vision Transformer (ViT) features via learnable tokens and two-way attention. This design aims to combine ViT global visual priors with UNet's local inductive bias and high-resolution decoding for semantic segmentation of sparse, fine-structured targets in biomedical MRI and CT images, while avoiding end-to-end ViT fine-tuning in cross-domain settings. The central claim is that this yields higher Dice scores than baselines on the evaluated tasks.

Significance. If the empirical results hold, the work offers a practical route to transfer large-scale visual priors from ViTs to high-complexity biomedical segmentation without the cost of ViT retraining. The approach is internally consistent, with ablations on the token and attention components supporting the design choices, and cross-domain results reported without evident contradictions.

minor comments (2)
  1. The abstract claims outperformance on MRI and CT tasks but the results section would benefit from explicit reporting of dataset sizes, number of runs, and standard deviations alongside the Dice scores to strengthen reproducibility.
  2. Figure 3 (qualitative results) could include error maps or failure cases to better illustrate where the structure-conditioned decoding improves over baselines.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our work and the recommendation for minor revision. The review correctly identifies the core contribution of conditioning a UNet decoder on frozen ViT features via learnable tokens and two-way attention to improve biomedical segmentation without full ViT fine-tuning.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper introduces ViTC-UNet as an architectural combination of frozen ViT representations with a UNet decoder via learnable tokens and two-way attention, evaluated empirically on MRI and CT segmentation tasks. No equations, derivations, or first-principles predictions are present that could reduce to fitted inputs or self-referential definitions. Claims rest on reported Dice score improvements and ablations rather than any load-bearing self-citation chain or ansatz smuggled through prior work. The method is self-contained as a practical design proposal with independent empirical support.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that frozen ViT features contain transferable global priors suitable for biomedical targets and that the proposed conditioning mechanism can supply missing local bias without retraining the backbone.

free parameters (1)
  • learnable tokens
    Parameters introduced to interface frozen ViT representations with the UNet decoder; their values are learned during training.
axioms (1)
  • domain assumption Frozen pre-trained ViT representations provide useful global visual priors that can be effectively transferred to biomedical segmentation via conditioning.
    Invoked to justify avoiding end-to-end ViT fine-tuning in cross-domain settings.
invented entities (1)
  • ViTC-UNet no independent evidence
    purpose: Hybrid architecture that conditions UNet decoding on ViT features for domain-adaptive biomedical segmentation.
    New model introduced in the paper; no independent evidence outside the proposed method.

pith-pipeline@v0.9.0 · 5691 in / 1363 out tokens · 32167 ms · 2026-05-20T22:38:29.363454+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · 8 internal anchors

  1. [1]

    Afridi, S. et al. (2026) ‘3D-VIT-unet: 3D Vision Transformer based unet-like model for volumet- ric brain tumor segmentation’, PLOS Digital Health, 5(3). doi:10.1371/journal.pdig.0001323

    work page doi:10.1371/journal.pdig.0001323 2026

  2. [2]

    Antonelli, M., Reinke, A., Bakas, S. et al. The Medical Segmentation Decathlon. Nat Commun 13, 4128 (2022). https://doi.org/10.1038/s41467-022-30695-9

    work page doi:10.1038/s41467-022-30695-9 2022

  3. [3]

    Archit, A. et al. (2025) ‘Segment anything for Microscopy’, Nature Methods, 22(3), pp. 579–591. doi:10.1038/s41592-024-02580-4

    work page doi:10.1038/s41592-024-02580-4 2025

  4. [4]

    (2017) SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation

    Badrinarayanan,V ., et al. (2017) SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(12), 2481-2495

    work page 2017

  5. [5]

    The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification

    U.Baid, et al., "The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmen- tation and Radiogenomic Classification", arXiv:2107.02314, 2021(opens in a new window)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  6. [6]

    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", Nature Scientific Data, 4:170117 (2017) DOI: 10.1038/sdata.2017.117

    work page doi:10.1038/sdata.2017.117 2017

  7. [7]

    Carion, N. et al. (2025) SAM 3: Segment Anything with Concepts. arXiv. https://arxiv.org/abs/2511.16719

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    https://arxiv.org/abs/2112.01527

    Cheng, B., et al. (2021) Masked-attention Mask Transformer for Universal Image Segmentation. arXiv. arXiv:2112.01527

    work page arXiv 2021

  9. [9]

    Chi, W. et al. (2020) ‘Deep learning-based medical image segmentation with limited labels’, Physics in Medicine & Biology, 65(23), p. 235001. doi:10.1088/1361-6560/abc363

    work page doi:10.1088/1361-6560/abc363 2020

  10. [10]

    (2026) Bi-Orthogonal Factor Decomposition for Vision Transformers

    Doshi, F.R, et al. (2026) Bi-Orthogonal Factor Decomposition for Vision Transformers. arXiv. arXiv:2601.05328

    work page arXiv 2026

  11. [11]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  12. [12]

    Fedorov, A; Schwier, M; Clunie, D; Herz, C; Pieper, S; Kikinis, R; Tempany, C; Fennessy, F. (2018). Data From QIN-PROSTATE-Repeatability. The Cancer Imaging Archive. DOI: 10.7937/K9/TCIA.2018.MR1CKGND

    work page doi:10.7937/k9/tcia.2018.mr1ckgnd 2018

  13. [13]

    TAC-UNet: transformer-assisted convolu- tional neural network for medical image segmentation

    He J, Ma Y , Yang M, Yang W, Wu C, Chen S. TAC-UNet: transformer-assisted convolu- tional neural network for medical image segmentation. Quant Imaging Med Surg. 2024 Dec 5;14(12):8824-8839. doi: 10.21037/qims-24-1229. Epub 2024 Nov 5. PMID: 39698603; PM- CID: PMC11651933

    work page doi:10.21037/qims-24-1229 2024

  14. [14]

    Hernandez Petzsche, M.R., de la Rosa, E., Hanning, U. et al. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Sci Data 9, 762 (2022). https://doi.org/10.1038/s41597-022-01875-5

    work page doi:10.1038/s41597-022-01875-5 2022

  15. [15]

    Isensee, F. et al. (2024) ‘NNU-Net Revisited: A Call for rigorous validation in 3D medical image segmentation’, Lecture Notes in Computer Science, pp. 488–498. doi:10.1007/978-3- 031-72114-4_47

    work page doi:10.1007/978-3- 2024

  16. [16]

    F., Kohl, S

    Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., & Maier-Hein, K. H. (2021). nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2), 203-211

    work page 2021

  17. [17]

    (2020) nnU-Net for Brain Tumor Segmentation

    Isensee, F., et al. (2020) nnU-Net for Brain Tumor Segmentation. arXiv. arXiv:2011.00848 10

    work page arXiv 2020

  18. [18]

    (2022) Extending nnU-Net is all you need

    Isensee, F., et al. (2022) Extending nnU-Net is all you need. arXiv. arXiv:2208.10791

    work page arXiv 2022

  19. [19]

    (2019) ‘CHAOS - Combined (CT-MR) Healthy Abdominal Organ Segmentation Challenge Data’

    Kavur, A., et al. (2019) ‘CHAOS - Combined (CT-MR) Healthy Abdominal Organ Segmentation Challenge Data’. The IEEE International Symposium on Biomedical Imaging (ISBI), Zenodo. doi:10.5281/zenodo.3431873

    work page doi:10.5281/zenodo.3431873 2019

  20. [20]

    Kavur, A.E. et al. (2021) ‘Chaos challenge - combined (CT-MR) healthy abdominal organ segmentation’, Medical Image Analysis, 69, p. 101950. doi:10.1016/j.media.2020.101950

    work page doi:10.1016/j.media.2020.101950 2021

  21. [21]

    Segment Anything

    Kirillov, A., Mintun, E., Ravi N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A., Lo, W., Dollár, P., Girshick, R.: Segment Anything. arXiv (2023) arXiv:2304.02643

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  22. [22]

    Koch, V ., et al (2024) DinoBloom: A Foundation Model for Generalizable Cell Embeddings in Hematology. arXiv. arXiv:2404.05022

    work page arXiv 2024

  23. [23]

    Lamm, L. et al. (2024) MemBrain V2: An end-to-end tool for the analysis of membranes in cryo-electron tomography [Preprint]. doi:10.1101/2024.01.05.574336

    work page doi:10.1101/2024.01.05.574336 2024

  24. [24]

    and Sharp, T.H

    Last, M.G., V oortman, L.M. and Sharp, T.H. (2025) Scaling data analyses in cellular cryoET using comprehensive segmentation [Preprint]. doi:10.1101/2025.01.16.633326

    work page doi:10.1101/2025.01.16.633326 2025

  25. [25]

    Li, F. et al. (2022) ‘Segmentation of human aorta using 3D NNU-net-oriented deep learning’, Review of Scientific Instruments, 93(11). doi:10.1063/5.0084433

    work page doi:10.1063/5.0084433 2022

  26. [26]

    Li, L. et al. (2023) ‘MyoPS: A benchmark of myocardial pathology segmentation combining three-sequence cardiac magnetic resonance images’, Medical Image Analysis, 87, p. 102808. doi:10.1016/j.media.2023.102808

    work page doi:10.1016/j.media.2023.102808 2023

  27. [27]

    (2025) ’Few-Shot Deployment of Pretrained MRI Transformers in Brain Imaging Tasks’

    Li, M., et al. (2025) ’Few-Shot Deployment of Pretrained MRI Transformers in Brain Imaging Tasks’. arXiv. arXiv:2508.05783

    work page arXiv 2025

  28. [28]

    Li, X., et al. (2025) ‘Evit-UNET: U-net like efficient vision transformer for medical image segmentation on mobile and Edge Devices’, 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pp. 1–5. doi:10.1109/isbi60581.2025.10981108

    work page doi:10.1109/isbi60581.2025.10981108 2025

  29. [29]

    (2017) Focal loss for dense object detection

    Lin, T., et al. (2017) Focal loss for dense object detection. ICCV

    work page 2017

  30. [30]

    (2025) Unified Open-World Segmentation with Multi-Modal Prompts

    Liu, Y ., et al. (2025) Unified Open-World Segmentation with Multi-Modal Prompts. ICCV . arXiv:2510.10524

    work page arXiv 2025

  31. [31]

    (2015) Fully Convolutional Networks for Semantic Segmentation

    Long, J. (2015) Fully Convolutional Networks for Semantic Segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 3431-3440

    work page 2015

  32. [32]

    ICLR (2019)

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. ICLR (2019)

    work page 2019

  33. [33]

    (2022) AbdomenCT-1K: Is Abdominal Organ Segmentation a Solved Problem?

    Ma, J., et al. (2022) AbdomenCT-1K: Is Abdominal Organ Segmentation a Solved Problem?. IEEE Transactions on Pattern Analysis and Machine Intelligence. 10.1109/TPAMI.2021.3100536

    work page doi:10.1109/tpami.2021.3100536 2022

  34. [34]

    Ma, J. et al. (2021) ‘Toward data-efficient learning: A benchmark for Covid-19 CT Lung and infection segmentation’, Medical Physics, 48(3), pp. 1197–1210. doi:10.1002/mp.14676

    work page doi:10.1002/mp.14676 2021

  35. [35]

    Ma, J. et al. (2024) ‘Segment anything in Medical Images’, Nature Communications, 15(1). doi:10.1038/s41467-024-44824-z

    work page doi:10.1038/s41467-024-44824-z 2024

  36. [36]

    Matsoukas, C., et al (2022) What Makes Transfer Learning Work For Medical Images: Feature Reuse & Other Factors. arXiv. arXiv:2203.01825

    work page arXiv 2022

  37. [37]

    (2023) Pretrained ViTs Yield Versatile Representations For Medical Images

    Matsoukas, C., et al. (2023) Pretrained ViTs Yield Versatile Representations For Medical Images. arXiv. arXiv:2303.07034

    work page arXiv 2023

  38. [38]

    (2024) LVS-Net: A Lightweight Vessels Segmentation Network for Retinal Image Analysis

    Mehmood, M. (2024) LVS-Net: A Lightweight Vessels Segmentation Network for Retinal Image Analysis. arXiv. arXiv:2412.05968v1 11

    work page arXiv 2024

  39. [39]

    Menze, B.H. et al. (2015) ‘The Multimodal Brain Tumor Image Segmentation Benchmark (brats)’, IEEE Transactions on Medical Imaging, 34(10), pp. 1993–2024. doi:10.1109/tmi.2014.2377694

    work page doi:10.1109/tmi.2014.2377694 2015

  40. [40]

    (2016) V-Net:Fully convolutional neural networks for volumetric medical image segmentation

    Milletari, F., et al. (2016) V-Net:Fully convolutional neural networks for volumetric medical image segmentation. 3DV

    work page 2016

  41. [41]

    DINOv2: Learning Robust Visual Features without Supervision

    Oquab, M., Darcet, T., Moutakanni, T., V o, H., Szafraniec, M., Khalidov, V ., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  42. [42]

    Pal, D. et al. (2025) ‘Pannet: A feature-based attention aggregation model for segmenting pancreatic ductal adenocarcinoma on contrast-enhanced CT images of the abdomen’, Medical Imaging 2025: Computer-Aided Diagnosis, p. 63. doi:10.1117/12.3048971

    work page doi:10.1117/12.3048971 2025

  43. [43]

    Payer, T. et al. (2023) ‘Medical volume segmentation by overfitting sparsely annotated data’, Journal of Medical Imaging, 10(04). doi:10.1117/1.jmi.10.4.044007

    work page doi:10.1117/1.jmi.10.4.044007 2023

  44. [44]

    Podobnik, G. et al. (2023) ‘Han-Seg: The head and neck organ-at-risk CT and mr segmentation dataset’, Medical Physics, 50(3), pp. 1917–1927. doi:10.1002/mp.16197

    work page doi:10.1002/mp.16197 2023

  45. [45]

    Radl, Lukas; Jin, Yuan; Pepe, Antonio; Li, Jianning; Gsaxner, Christina; Zhao, Fen-hua; et al. (2022). Aortic Vessel Tree (A VT) CTA Datasets and Segmentations. figshare. Dataset. https://doi.org/10.6084/m9.figshare.14806362.v1

    work page doi:10.6084/m9.figshare.14806362.v1 2022

  46. [46]

    SAM 2: Segment Anything in Images and Videos

    Ravi, N., et al. (2024) SAM 2: Segment Anything in Images and Videos. arXiv. https://arxiv.org/abs/2408.00714

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  47. [47]

    Walk in the cloud: Learning curves for point clouds shape analysis, pp

    Ranftl, R., Bochkovskiy, A. and Koltun, V . (2021) Vision Transformers for dense prediction, 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pp. 12159–12168. doi:10.1109/iccv48922.2021.01196

    work page doi:10.1109/iccv48922.2021.01196 2021

  48. [48]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    Ronneberger, O. (2015) U-Net: Convolutional Networks for Biomedical Image Segmentation. arXiv. arXiv:1505.04597

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  49. [49]

    de la Rosa, E., Reyes, M., Liew, SL. et al. DeepISLES: a clinically validated ischemic stroke segmentation model from the ISLES’22 challenge. Nat Commun 16, 7357 (2025). https://doi.org/10.1038/s41467-025-62373-x

    work page doi:10.1038/s41467-025-62373-x 2025

  50. [50]

    ICLR (2019)

    Sablayrolles A., Douze M., Schmid C., and Jégou H.: Spreading vectors for similarity search. ICLR (2019)

    work page 2019

  51. [51]

    Sang, Y . et al. (2025) Benchmark of Segmentation Techniques for Pelvic Fracture in CT and X-ray: Summary of the PENGWIN 2024 Challenge. arXiv. arXiv:2504.02382

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  52. [52]

    Soh WK and Rajapakse JC (2023) Hybrid UNet transformer architecture for ischemic stoke segmentation with MRI and CT datasets. Front. Neurosci. 17:1298514. doi: 10.3389/fnins.2023.1298514

    work page doi:10.3389/fnins.2023.1298514 2023

  53. [53]

    Støverud, K.-H. et al. (2024) ‘AeroPath: An airway segmentation benchmark dataset with challenging pathology and Baseline Method’, PLOS ONE, 19(10). doi:10.1371/journal.pone.0311416

    work page doi:10.1371/journal.pone.0311416 2024

  54. [54]

    (2025) Randomized-MLP Regularization Improves Domain Adaptation and Interpretability in DINOv2

    Valdivia Ortega, J., et al. (2025) Randomized-MLP Regularization Improves Domain Adaptation and Interpretability in DINOv2. NeurIPS. arXiv:2511.05509

    work page arXiv 2025

  55. [55]

    van der Graaf, J.W., van Hooff, M.L., Buckens, C.F.M. et al. Lumbar spine segmen- tation in MR images: a dataset and a public benchmark. Sci Data 11, 264 (2024). https://doi.org/10.1038/s41597-024-03090-w

    work page doi:10.1038/s41597-024-03090-w 2024

  56. [56]

    (2023) ‘SPIDER - Lumbar spine segmentation in MR images: a dataset and a public benchmark’

    van der Graaf, J., et al. (2023) ‘SPIDER - Lumbar spine segmentation in MR images: a dataset and a public benchmark’. Zenodo. doi:10.5281/zenodo.10159290

    work page doi:10.5281/zenodo.10159290 2023

  57. [57]

    (2023) ‘Dataset with segmentations of 117 important anatomical structures in 1228 CT images’

    Wasserthal, J. (2023) ‘Dataset with segmentations of 117 important anatomical structures in 1228 CT images’. Zenodo. doi:10.5281/zenodo.10047292. 12

    work page doi:10.5281/zenodo.10047292 2023

  58. [58]

    Wasserthal, J. et al. (2023) ‘TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images’, Radiology: Artificial Intelligence, 5(5). doi:10.1148/ryai.230024

    work page doi:10.1148/ryai.230024 2023

  59. [59]

    (2024) ’Enhancing surgical instrument segmentation: integrating vision trans- former insights with adapter’ Int J Comput Assist Radiol Surg

    Wei, M., et al. (2024) ’Enhancing surgical instrument segmentation: integrating vision trans- former insights with adapter’ Int J Comput Assist Radiol Surg. 10.1007/s11548-024-03140-z

    work page doi:10.1007/s11548-024-03140-z 2024

  60. [60]

    Xu, Q. et al. (2026) ‘Robust multi-domain digital pathology image segmentation via joint balancing representation learning’, Expert Systems with Applications, 320, p. 132093. doi:10.1016/j.eswa.2026.132093

    work page doi:10.1016/j.eswa.2026.132093 2026

  61. [61]

    (2022) AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation

    Yuanfeng, J., et al. (2022) AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation. arXiv. arXiv:2206.08023

    work page arXiv 2022

  62. [62]

    Zhao, T. et al. (2024) ‘A Foundation model for joint segmentation, detection and recogni- tion of biomedical objects across nine modalities’, Nature Methods, 22(1), pp. 166–176. doi:10.1038/s41592-024-02499-w

    work page doi:10.1038/s41592-024-02499-w 2024

  63. [63]

    and Yan, P

    Zhu, Q., Du, B. and Yan, P. (2020) ‘Boundary-weighted domain adaptive neural network for prostate mr image segmentation’, IEEE Transactions on Medical Imaging, 39(3), pp. 753–763. doi:10.1109/tmi.2019.2935018. 13 A Technical Appendices and Supplementary Material A.1 Licenses The datasets used in this paper where obtained as part of the compilation made by...

    work page doi:10.1109/tmi.2019.2935018 2020