Pith. sign in

REVIEW 5 major objections 4 minor 30 references

A frozen general vision model carries dormant spatial priors that a lightweight TokenBook can convert into guide masks, improving medical segmentation without fine-tuning the foundation model.

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

2026-08-02 19:42 UTC pith:6O5TP5IB

load-bearing objection Promising intuition, confounded evidence: the guide mask is supervised directly with ground truth, so the claim about latent VFM priors is not supported; the abstract/body mismatch and table contradictions also need fixing. the 5 major comments →

arxiv 2603.01115 v2 pith:6O5TP5IB submitted 2026-03-01 cs.CV

LUMOS: Latent Universal Medical Priors for Segmentation

Zhuonan Liang , Wei Guo , Jie Gan , Yaxuan Song , Runnan Chen , Hang Chang , Weidong Cai This is my paper
classification cs.CV
keywords medical image segmentationvision foundation modelsDINOv3frozen featuresspatial guidanceTokenBooktransfer learningguide masks
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.

General vision foundation models trained on natural images are usually thought to be of limited use for medical segmentation without fine-tuning or domain adaptation. This paper claims the opposite: low-level visual priors needed for anatomical delineation already lie dormant in such models, and a lightweight TokenBook mechanism can extract them from a frozen DINOv3 backbone as a spatial guide mask. The mask gates feature activations in a conventional segmentation network, which is trained alongside it with a guide-supervision loss that aligns the mask with ground-truth regions. Across colonoscopy, dermoscopy, and thyroid ultrasound datasets, the framework reports consistent gains in Dice, IoU, and boundary distance, with the largest improvements on weak backbones. The practical pay-off: foundation-model priors can be injected without fine-tuning the foundation model, preserving the efficiency and inductive biases of dedicated medical architectures.

Core claim

The central claim, in the paper's own terms, is that the low-level visual priors necessary for anatomical delineation already lie dormant within general vision foundation models, and that they can be amplified without fine-tuning. The framework freezes a DINOv3 backbone (a self-supervised vision transformer pretrained on natural images), distills its token features through a TokenBook of learned prototypes into a spatial guide mask, and uses that mask to gate feature activations in the segmentation backbone. A guide-supervision loss—binary cross-entropy against the ground-truth mask, optionally with a boundary hinge—trains the prototypes and weights. Experiments across three medical datasets

What carries the argument

The TokenBook mechanism is the load-bearing component. It maps frozen token features from a natural-image-pretrained vision foundation model (DINOv3 by default) to a spatial guide mask by computing similarities between each token and a set of learned prototypes, then taking a weighted sum: G(x) = Σ_i α_i · sim(T^i, P). The prototypes P and weights α_i are trained during segmentation using a BCE guide-supervision loss that aligns the mask with the ground truth. This guide mask then gates the feature activations of the segmentation backbone, injecting the foundation model's spatial regularities while leaving the backbone's own inductive biases intact.

Load-bearing premise

The load-bearing premise is that the measured improvements come from the frozen foundation model's latent priors, not from the fact that the guide mask itself is trained with ground-truth supervision through the guide-loss term.

What would settle it

Train the full framework on any of the three datasets, but replace the frozen DINOv3 token features with random Gaussian features or a frozen randomly initialized network, keeping TokenBook, guide loss, and gating otherwise identical. If segmentation performance does not drop, the method's gains are attributable to the label-trained guide pathway rather than to pre-trained vision priors.

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

If this is right

  • If frozen VFM features carry sufficient spatial priors, medical segmentation can be improved without fine-tuning or domain adaptation of the foundation model, reducing training cost and data needs.
  • The gating is architecture-agnostic and can wrap convolutional or transformer-based segmentation backbones, preserving their inductive biases while adding external spatial guidance.
  • Parameter-efficient adaptation (LoRA) of the frozen generator can further sharpen the guidance, suggesting the dormant priors can be tuned rather than relearned.
  • The largest gains appear on backbones that perform poorly alone, indicating the guide acts as a regularizer that rescues under-performing architectures.
  • The reported HD95 reductions imply the guide improves not only localization but also boundary delineation, which is a key clinical requirement.

Where Pith is reading between the lines

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

  • The current experiments do not isolate the frozen VFM priors from the label-trained guide pathway: the prototypes and weights are fit with a BCE loss against the ground truth, so some or all of the gain could come from that auxiliary supervised pathway rather than from dormant VFM knowledge.
  • A direct test is to replace the frozen DINOv3 token features with random or differently initialized features while keeping TokenBook, guide loss, and gating identical; if performance holds, the pre-trained priors are not the driver.
  • The SigLIP sensitivity offers a handle on mechanism: if token granularity is the issue, then patch-size-matching SigLIP to DINOv3 should recover much of the gap, which would implicate token-space structure rather than the representation objective.
  • The same gating recipe could transfer to other dense prediction tasks (detection, registration, anomaly localization) where a coarse spatial prior from a frozen VFM may substitute for expensive supervision.

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

5 major / 4 minor

Summary. The manuscript, submitted under the title LUMOS, contains an abstract describing LUMOS with Pathfinder and Inspiror components, while the body is a different paper titled 'GuiDINO: Rethinking Vision Foundation Model in Medical Image Segmentation' with a TokenBook mechanism. The body proposes using a frozen DINOv3 token encoder to compute token-prototype similarities, which are aggregated into a spatial guide mask that gates feature activations in conventional segmentation backbones (UNet, nnWNet). The guide mask is trained with binary cross-entropy against the ground-truth mask (Eq. 3), optionally with a boundary hinge loss and LoRA adaptation. The paper claims consistent segmentation improvements across Kvasir-SEG, ISIC 2017, and TN3K datasets, and interprets the gains as evidence that frozen general VFMs contain dormant spatial priors useful for medical segmentation.

Significance. The idea that a frozen general VFM can serve as a spatial prior generator without fine-tuning is potentially valuable for efficient medical segmentation, and the proposed gating mechanism is simple and plausibly cheap to integrate. The release of code is a positive aspect, and the qualitative guide masks in Fig. 2 do appear to localize targets roughly. However, the current evidence does not establish the central claim: the guide mask is task-supervised through BCE against ground truth, so gains could come from the auxiliary supervised pathway rather than from latent VFM priors. Moreover, the paper's own tables contradict several textual claims of consistent improvement. The significance is therefore conditional on substantial additional experimentation and careful reinterpretation.

major comments (5)
  1. [Abstract, §1, §3] The abstract describes LUMOS with Pathfinder/Inspiror, DINO matched-backbone gains, and SigLIP sensitivity. The body is titled GuiDINO and presents only a TokenBook mechanism and DINOv3 experiments; no SigLIP results, no Pathfinder, no Inspiror. The two frameworks are never reconciled, making it impossible to verify the paper's declared contributions. The authors must decide which method is being submitted and align the title, abstract, and body.
  2. [§3.1–3.2, Eq. (1), Eq. (3)] The guide mask G(x) is computed from learnable prototypes P and learnable weights α_i, and is trained directly with BCE against the ground-truth mask (Eq. 3). Thus G(x) is not an unsupervised or frozen quantity. Any downstream improvement over a backbone-only baseline could arise from the additional parameters/capacity of the TokenBook branch, the extra dense BCE supervision, or the gating operation itself, rather than from the specific DINOv3 pretrained features. To support the 'latent universal medical prior' interpretation, the authors need control experiments: random or untrained VFM features, replacement of DINOv3 with a task-trained encoder of similar capacity, and a no-guide baseline with equivalent parameter count. No such ablation is reported.
  3. [§4.3, Table 1] The text states that GuiDINO-W 'outperforms the baseline models across all three datasets,' but Table 1 shows that on ISIC 2017, GuiDINO-W achieves IoU 82.32 and HD95 13.14, both worse than nnWNet's 82.62 and 12.73. On TN3K, SegDINO has a better HD95 (18.62) than GuiDINO-W (24.80). The claim of consistent superiority is not supported by the paper's own numbers.
  4. [§4.4, Table 2] The ablation text asserts that integrating GuiDINO 'consistently improves segmentation performance across both backbones and all datasets.' Table 2 contradicts this: GuiDINO-W on ISIC loses 0.30 IoU and worsens HD95 by 0.41 relative to nnWNet; GuiDINO-U on TN3K has a +0.38 HD95 degradation. The word 'consistently' is false and also conflicts with the abstract's 'stable matched-backbone gains.' The claims need to be revised to reflect the mixed results.
  5. [§4.5, Table 3] The LoRA comparison further weakens the guidance hypothesis. Seg-GuiDINO without LoRA is worse than Seg-DINO on Kvasir (−1.54 IoU) and TN3K (−12.76 IoU). When LoRA is added, TN3K HD95 worsens from 16.13 to 21.41 relative to Seg-GuiDINO without LoRA. The discussion acknowledges inconsistency but does not explain how these results support the central claim that DINO guidance benefits segmentation. At minimum, the paper should report which differences are statistically significant given the three seeds.
minor comments (4)
  1. [§4.1] The dataset name is written 'TK3N' in the text but appears as 'TN3K' in all tables; please make this consistent and verify it matches reference [5].
  2. [Eq. (1)] The notation for token features is inconsistent ('T i' vs 'T_i'); clarify that the sum is over spatial positions and define the dimensionality of the similarity output.
  3. [Figure 1, §4.5] The caption says the DINOv3 backbone is frozen, but §4.5 describes LoRA adaptation of the DINOv3 guide backbone. Specify in the figure and methods whether the backbone is always frozen or optionally adapted.
  4. [§4.2, Table 1] All results are reported as means over three seeds without error bars or significance tests. Given the small observed differences (e.g., ±0.3 IoU), provide standard deviations or confidence intervals.

Circularity Check

1 steps flagged

Guide mask is trained with ground-truth BCE (Eq. 3), so its claimed ability to capture the target is forced by construction; attribution of gains to frozen VFM priors is therefore partially circular.

specific steps
  1. self definitional [Sec. 3.2, Eq. (3); Sec. 4.5 / Fig. 2 caption]
    "Lguide = − 1/N Σ [yi log(gi) + (1−yi) log(1−gi)] ... It can be observed that the guidance from DINO can capture the rough location of the target, which can potentially help the segmentation backbone to focus on relevant regions"

    The guide mask G(x) is not a frozen or unsupervised quantity: Eq. (3) trains it directly with binary cross-entropy against the ground-truth mask y. Therefore the Fig. 2 / Sec. 4.5 observation that the guidance 'capture[s] the rough location of the target' is guaranteed by the training objective and provides no independent evidence about dormant DINOv3 priors. The paper's central claim that general VFMs act as spatial prior generators is supported only by a GT-supervised auxiliary branch whose output is aligned to GT by construction; the reported gains could come from extra parameters, the dense BCE supervision, or the gating operation rather than from the specific frozen VFM features. No ablation replaces the frozen features with random or task-trained features to isolate the VFM contribut

full rationale

The paper is empirical rather than a formal derivation, so circularity appears in the interpretation of evidence rather than in a mathematical chain. The numerical comparisons against nnUNet, nnWNet, SegDINO, etc. are external benchmarks and are not circular. However, the headline claim that frozen VFMs supply 'latent universal medical priors' is not isolated experimentally: the guide mask is trained with BCE against the ground-truth mask (Eq. 3), so its ability to highlight target regions is definitional. No random-feature, feature-replacement, or equal-capacity task-trained control is provided, so improvements cannot be attributed specifically to the frozen DINOv3 token space as opposed to the GT-supervised TokenBook pathway. This is a central interpretive circularity, although not a total one: the frozen backbone weights are genuinely not trained and the final segmentation predictions are not identical to the guide mask by construction. Table 2 also shows GuiDINO-W losing 0.30 IoU on ISIC and worsening HD95 by 0.41, contradicting the text's 'consistent improvements' claim, but that inconsistency is a correctness concern rather than circularity. Score 6 reflects the partial reduction of the key evidence to a label-fitted input.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The framework rests on a small set of learned components, but the crucial attribution to frozen VFM priors is not independently testable because the guide mask is GT-supervised. The learned prototypes, weights, and hyperparameters are the effective fitted parameters of the method.

free parameters (5)
  • TokenBook prototypes P = learned, exact values not reported
    Learned in Eq. (1) and supervised by ground truth via Eq. (3); they are the main vehicle for translating VFM tokens into a task-specific guide.
  • Token contribution weights α_i = learned, exact values not reported
    Learned per-token weights in Eq. (1); fit on each dataset.
  • Guide supervision weight λ = not reported
    Hyperparameter in Eq. (2) controlling L_guide; no ablation or value given.
  • Number of prototypes N and similarity function choice = not reported
    Architectural choices in Eq. (1) set by hand; alternatives are not compared.
  • LoRA rank / adaptation hyperparameters = not reported
    LoRA variants in Table 3 require rank and alpha choices that are not reported.
axioms (4)
  • domain assumption Frozen token-based VFMs (DINOv3) encode patch-level spatial regularities transferable to medical images.
    Core premise; asserted in the abstract and §3.1 but never isolated from GT-supervised guide training.
  • ad hoc to paper Token–prototype similarity (dot or cosine) yields a useful spatial guide mask.
    Design choice in Eq. (1) with no comparison to alternatives or theoretical support.
  • domain assumption Gating backbone features with the guide mask is an effective and bias-preserving injection mechanism.
    Architectural assumption in §3.1; no ablation against concatenation, cross-attention, or feature modulation.
  • domain assumption nnUNet-style augmentation and inference transfer correctly to these datasets.
    Used in §4.2; not independently validated.

pith-pipeline@v1.3.0-alltime-deepseek · 8928 in / 12286 out tokens · 115892 ms · 2026-08-02T19:42:29.840090+00:00 · methodology

0 comments
read the original abstract

General vision foundation models (VFMs) have been primarily developed on natural images, and their utility for medical image segmentation is therefore often considered to depend on costly adaptation or domain-specific fine-tuning. In this paper, we revisit this assumption from a different perspective: rather than requiring VFM segmentors to relearn visual regularities, we investigate whether the low-level visual priors necessary for anatomical delineation already lie dormant within general VFMs. We observe that frozen VFMs, despite lacking medical supervision, encode transferable visual regularities. These properties are not exclusive to natural images but are also fundamental to medical image understanding. Motivated by this observation, we propose Latent Universal Medical PriOrs for Segmentation (LUMOS), a novel framework that amplifies general VFM priors to conventional medical segmentors. LUMOS consists of two key components: (1) Pathfinder that distills visual cues from a frozen vision foundation model, and (2) Inspiror that sparks the conventional medical networks with spatial guidance from distilled visual regularities. In this way, the segmentor is relieved from learning complex visual regularities entirely from limited medical annotations and can instead focus on task-specific anatomical delineation. Across diverse medical datasets and token-based VFMs, LUMOS shows that general VFMs can serve as spatial prior generators when their frozen token spaces preserve patch-level pattern relevance. DINO provides stable matched-backbone gains, while SigLIP exposes VFM-specific sensitivity caused by its different token granularity and representation objective.

Figures

Figures reproduced from arXiv: 2603.01115 by Hang Chang, Jie Gan, Runnan Chen, Weidong Cai, Wei Guo, Yaxuan Song, Zhuonan Liang.

Figure 1
Figure 1. Figure 1: Overview of the proposed GuiDINO framework. The frozen DINOv3 backbone serves as a guide generator, extracting dense token features that are converted into spatial guide masks via the TokenBook mechanism. These guide masks are then used to gate feature activations in the segmentation backbone, which produces the final segmentation output. extracts dense token features and converts them into spatial guide m… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative results for segmentation. First row: original input image. Sec￾ond row: actual segmentation masks. Third row: guidance mask from GuiDINO. The guidance mask highlights the rough location of the target, which can assist the segmentation backbone in focusing on relevant regions [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] 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

30 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    IEEE Transactions on Medi- cal ImagingPP(2025)

    Archit, A., Freckmann, L., Pape, C.: Medicosam: Robust improve- ment of sam for medical imaging. IEEE Transactions on Medi- cal ImagingPP(2025). https://doi.org/10.1109/TMI.2025.3644811, https://www.ncbi.nlm.nih.gov/pubmed/41406266

    arXiv 2025

  2. [2]

    In: European conference on computer vision 2022

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision 2022. pp. 205–218. Computer Vision – ECCV 2022 Workshops, Springer Nature Switzerland (2023). https://doi.org/10.1007/978-3- 031-25066-8_9

    doi:10.1007/978-3- 2022

  3. [3]

    EngMedicine3(2) (2026)

    Chen, Q., Liu, A., Zhang, J., Yang, C., Zhang, Y.: Foundation models in medical imaging: A review. EngMedicine3(2) (2026). https://doi.org/10.1016/j.engmed.2026.100123 10 Z. Liang et al

    arXiv 2026

  4. [4]

    In: 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)

    Codella, N.C.F., Gutman, D., Celebi, M.E., Helba, B., Marchetti, M.A., Dusza, S.W., Kalloo, A., Liopyris, K., Mishra, N., Kittler, H., Halpern, A.: Skin lesion analysis toward melanoma detection: A challenge at the 2017 international sympo- sium on biomedical imaging (isbi), hosted by the international skin imaging collab- oration (isic). In: 2018 IEEE 15...

    arXiv 2017

  5. [5]

    In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)

    Gong, H., Chen, G., Wang, R., Xie, X., Mao, M., Yu, Y., Chen, F., Li, G.: Multi- task learning for thyroid nodule segmentation with thyroid region prior. In: 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI). pp. 257–261 (2021). https://doi.org/10.1109/isbi48211.2021.9434087

    arXiv 2021

  6. [6]

    IEEE Transactions on Medi- cal Imaging42(9), 2763–2775 (2023)

    He, A., Wang, K., Li, T., Du, C., Xia, S., Fu, H.: H2former: An efficient hierarchical hybrid transformer for medical image segmentation. IEEE Transactions on Medi- cal Imaging42(9), 2763–2775 (2023). https://doi.org/10.1109/TMI.2023.3264513, https://www.ncbi.nlm.nih.gov/pubmed/37018111

    arXiv 2023

  7. [7]

    In: European Conference of Computer Vision 2024

    Huang, J., Zhou, Y., Luo, Y., Liu, G., Guo, H., Yang, G.: Representing topologi- cal self-similarity using fractal feature maps for accurate segmentation of tubular structures. In: European Conference of Computer Vision 2024. pp. 143–160. Com- puter Vision – ECCV 2024, Springer Nature Switzerland (2025)

    2024

  8. [8]

    Nature Methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmenta- tion. Nature Methods18(2), 203–211 (2021). https://doi.org/10.1038/s41592-020- 01008-z, https://www.ncbi.nlm.nih.gov/pubmed/33288961

    arXiv 2021

  9. [9]

    In: MultiMedia Modeling 2020.pp.451–462.MultiMediaModeling,SpringerInternationalPublishing(2020)

    Jha, D., Smedsrud, P.H., Riegler, M.A., Halvorsen, P., de Lange, T., Johansen, D., Johansen, H.D.: Kvasir-seg: A segmented polyp dataset. In: MultiMedia Modeling 2020.pp.451–462.MultiMediaModeling,SpringerInternationalPublishing(2020). https://doi.org/10.1007/978-3-030-37734-2_37

    doi:10.1007/978-3-030-37734-2_37 2020

  10. [10]

    2023 IEEE/CVF International Conference on Computer Vision, ICCV pp

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. 2023 IEEE/CVF International Conference on Computer Vision, ICCV pp. 3992– 4003 (2023). https://doi.org/10.1109/Iccv51070.2023.00371

    arXiv 2023

  11. [11]

    Lee, S., Kang, S., Shim, H.: Self-supervised vision trans- formers are efficient segmentation learners for imperfect la- bels (January 01 2024). https://doi.org/10.48550/arXiv.2401.12535, https://ui.adsabs.harvard.edu/abs/2024arXiv240112535L, the AAAI Confer- ence on Artificial Intelligence 2024 Edge Intelligence Workshop (EIW)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2401.12535 2024

  12. [12]

    Proceedings of the AAAI Conference on Artificial In- telligence39(5), 4652–4660 (2025)

    Li, C., Liu, X., Li, W., Wang, C., Liu, H., Liu, Y., Chen, Z., Yuan, Y.: U-kan makes strong backbone for medical image segmenta- tion and generation. Proceedings of the AAAI Conference on Artificial In- telligence39(5), 4652–4660 (2025). https://doi.org/10.1609/aaai.v39i5.32491, https://ojs.aaai.org/index.php/AAAI/article/view/32491

    doi:10.1609/aaai.v39i5.32491 2025

  13. [13]

    In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Li, L., Zhou, Y., Yang, G.: Robust source-free domain adaptation for fundus image segmentation. In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 7825–7834 (2024). https://doi.org/10.1109/wacv57701.2024.00766

    arXiv 2024

  14. [14]

    https://doi.org/10.48550/arXiv.2509.02379, https://ui.adsabs.harvard.edu/abs/2025arXiv250902379L

    Li, Y., Wu, Y., Lai, Y., Hu, M., Yang, X.: Meddinov3: How to adapt vision foundation models for medical image segmenta- tion? (September 01 2025). https://doi.org/10.48550/arXiv.2509.02379, https://ui.adsabs.harvard.edu/abs/2025arXiv250902379L

    doi:10.48550/arxiv.2509.02379 2025

  15. [15]

    https://doi.org/10.48550/arXiv.2509.06467, https://ui.adsabs.harvard.edu/abs/2025arXiv250906467L, technical Report

    Liu, C., Chen, Y., Shi, H., Lu, J., Jian, B., Pan, J., Cai, L., Wang, J., Yu, J., Gao, Z., Zhang, X., Bai, L., Zhang, Y., Li, J., Bercea, C.I., Title Suppressed Due to Excessive Length 11 Ouyang, C., Chen, C., Xiong, Z., Wiestler, B., Wachinger, C., Duncan, J.S., Rueckert, D., Bai, W., Arcucci, R.: Does dinov3 set a new medical vision standard? benchmarki...

    doi:10.48550/arxiv.2509.06467 2025

  16. [16]

    Nature Communications15(1), 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical im- ages. Nature Communications15(1), 654 (2024). https://doi.org/10.1038/s41467- 024-44824-z, https://www.ncbi.nlm.nih.gov/pubmed/38253604

    arXiv 2024

  17. [17]

    Quantitative Imaging in Medicine and Surgery15(6), 5825–5858 (2025), https://qims.amegroups.org/article/view/138057

    Noh, S., Lee, B.D.: A narrative review of foundation models for medi- cal image segmentation: zero-shot performance evaluation on diverse modali- ties. Quantitative Imaging in Medicine and Surgery15(6), 5825–5858 (2025), https://qims.amegroups.org/article/view/138057

    2025

  18. [18]

    https://doi.org/10.48550/arXiv.2304.07193, https://ui.adsabs.harvard.edu/abs/2023arXiv230407193O

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Bal- las, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rab- bat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, J., La- batut, P., Joulin, A., Bojanowski, P.: Dinov2: Learning robust visual ...

    Pith/arXiv arXiv doi:10.48550/arxiv.2304.07193 2023

  19. [19]

    In: Medical Image Computing and Computer-Assisted Intervention (MICCAI 2015)

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention (MICCAI 2015). pp. 234–241. Lecture Notes in Computer Sci- ence, Springer International Publishing (2015). https://doi.org/10.1007/978-3-319- 24574-4_28, http://arxiv.org/pdf/1505.04597

    Pith/arXiv arXiv 2015

  20. [20]

    In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2025

    Schulthess, N., Konukoglu, E.: Anomaly detection by clustering dino embeddings using a dirichlet process mixture. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2025. pp. 46–56. Springer Nature Switzerland (2026)

    2025

  21. [21]

    https://doi.org/10.48550/arXiv.2508.10104, https://ui.adsabs.harvard.edu/abs/2025arXiv250810104S

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: Dinov3 (August 01 2025)....

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.10104 2025

  22. [22]

    BMC Medical Imaging24(1), 241 (2024)

    Sun, J., Chen, K., He, Z., Ren, S., He, X., Liu, X., Peng, C.: Medical image analysis using improved sam-med2d: segmentation and classification perspectives. BMC Medical Imaging24(1), 241 (2024). https://doi.org/10.1186/s12880-024-01401-6, https://www.ncbi.nlm.nih.gov/pubmed/39285324

    arXiv 2024

  23. [23]

    Computers in Biology and Medicine187, 109725 (2025)

    Wahd, A.S., Felfeliyan, B., Zhou, Y., Ghosh, S., McArthur, A., Zhang, J., Jaremko, J.L., Hareendranathan, A.: Sam2rad: A segmentation model for medical images with learnable prompts. Computers in Biology and Medicine187, 109725 (2025). https://doi.org/10.1016/j.compbiomed.2025.109725, https://www.ncbi.nlm.nih.gov/pubmed/39914197

    arXiv 2025

  24. [24]

    https://doi.org/10.48550/arXiv.2508.14809, https://ui.adsabs.harvard.edu/abs/2025arXiv250814809W

    Wang, S., Safari, M., Hu, M., Li, Q., Chang, C.W., Qiu, R.L., Yang, X.: Dinov3 with test-time training for medical image regis- tration (August 01 2025). https://doi.org/10.48550/arXiv.2508.14809, https://ui.adsabs.harvard.edu/abs/2025arXiv250814809W

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.14809 2025

  25. [25]

    Liang et al

    Xu, G., Udupa, J.K., Lu, W., Zhang, Y.: Exploiting dinov3-based self-supervised features for robust few-shot medical image segmen- 12 Z. Liang et al. tation (January 01 2026). https://doi.org/10.48550/arXiv.2601.08078, https://ui.adsabs.harvard.edu/abs/2026arXiv260108078X, 36 pages, 11 figures

    work page doi:10.48550/arxiv.2601.08078 2026

  26. [26]

    https://doi.org/10.48550/arXiv.2509.00833, https://ui.adsabs.harvard.edu/abs/2025arXiv250900833Y

    Yang, S., Wang, H., Xing, Z., Chen, S., Zhu, L.: Segdino: An ef- ficient design for medical and natural image segmentation with dino-v3 (August 01 2025). https://doi.org/10.48550/arXiv.2509.00833, https://ui.adsabs.harvard.edu/abs/2025arXiv250900833Y

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.00833 2025

  27. [27]

    Comput Biol Medical137, 104815 (2021)

    Yeung, M., Sala, E., Schonlieb, C.B., Rundo, L.: Focus u-net: A novel dual attention-gated cnn for polyp segmentation during colonoscopy. Comput Biol Medical137, 104815 (2021). https://doi.org/10.1016/j.compbiomed.2021.104815, https://www.ncbi.nlm.nih.gov/pubmed/34507156

    arXiv 2021

  28. [28]

    In: Medical Image Computing and Computer Assisted In- tervention – MICCAI 2025

    Zhang, X., Chen, E.Z., Zhao, L., Chen, X., Liu, Y., Maihe, B., Duncan, J.S., Chen, T., Sun, S.: Adapting vision foundation models for real-time ultrasound image segmentation. In: Medical Image Computing and Computer Assisted In- tervention – MICCAI 2025. pp. 24–34. Springer Nature Switzerland (2026). https://doi.org/10.1007/978-3-032-04971-1_3

    work page doi:10.1007/978-3-032-04971-1_3 2025

  29. [29]

    In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024

    Zheng, X., Zhang, Y., Zhang, H., Liang, H., Bao, X., Jiang, Z., Lao, Q.: Curriculum prompting foundation models for medical image segmentation. In: Medical Image Computing and Computer Assisted Intervention – MICCAI 2024. pp. 487–497. Springer Nature Switzerland (2024)

    2024

  30. [30]

    In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zhou, Y., Li, L., Lu, L., Xu, M.: nnwnet: Rethinking the use of transformers in biomedical image segmentation and calling for a unified evaluation benchmark. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 20852–20862 (2025). https://doi.org/10.1109/cvpr52734.2025.01942

    arXiv 2025