Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Recognizing Surgical Phases Anywhere: Few-Shot Test-time Adaptation and Task-graph Guided Refinement

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Surgical Phase Anywhere claims a frozen vision-language foundation model can be adapted to a new hospital's workflow with a few labeled images, text phase definitions, and a task graph, beating full-shot models at 32-shot.

desk verdict A solid few-shot adaptation framework for surgical phase recognition whose real gains are within the same backbone; the headline claim of beating full-shot models is undercut by a backbone mismatch, and the task graph's fidelity is never directly tested. read the letter →

arxiv 2506.20254 v2 pith:2AVGWXCD submitted 2025-06-25 cs.CV

Kun Yuan , Tingxuan Chen , Shi Li , Joel L. Lavanchy , Christian Heiliger , Ege Özsoy , Yiming Huang , Long Bai
show 4 more authors
Nassir Navab Vinkle Srivastav Hongliang Ren Nicolas Padoy
This is my paper
classification cs.CV
keywords surgicalphaserecognitionfew-shotlearningvision-languagemodelstest-timeadaptationdiffusionmodeltaskgraphcross-institutionalgeneralizationdatascience
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

The paper's claim is that surgical phase recognition can be made portable: a hospital should be able to take a frozen vision-language foundation model and adapt it to its own operating room with a handful of labeled images, phase descriptions written in plain text, and a task graph listing the allowed phase transitions and their duration bounds. The authors introduce Surgical Phase Anywhere (SPA), which combines few-shot spatial alignment, a diffusion model trained on synthetic sequences generated from that task graph, and self-supervised test-time adaptation that forces three prediction streams to agree; they report that at 32 labeled images per phase it beats every few-shot baseline they compare against on four benchmarks and exceeds a full-shot model on gastric bypass (BernBypass F1 43.54 against 42.4). Why it matters: if the claim holds, deploying accurate phase recognition at a new site stops being a data-collection project and becomes a matter of writing down the local protocol. The cost of that convenience is a single assumption, which the paper's own ablation exposes: the task graph must match the institution's real procedure, or performance falls hard.

What carries the argument

The load-bearing machinery is the pair formed by the task graph and the two adapters built around it. The task graph G=(V,E) is a directed graph whose nodes are the institution's phases and whose edges are permissible transitions, annotated with per-phase duration bounds [L_i^min, L_i^max]; it feeds a Markovian generator that draws the next phase uniformly from valid successors when the current phase has not exceeded its upper bound, producing the synthetic sequences used to train the temporal module. That module is a diffusion model over hidden states that one-hot encode phase labels: a Gaussian forward process destroys the sequence structure, and the learned reverse process denoises coarse predictions so they respect the protocol. The other adapter is the test-time loss $L = L_{\text{mutual}}(S_{\text{ref}}, S_{\text{vl}}) + L_{\text{mutual}}(S_{\text{fs}}, S_{\text{ref}})$, where $L_{\text{mutual}}(A,B) = -(1/L)\sum_l\sum_k A_{l,k}\log B_{l,k}$, which pushes the three prediction streams into frame-wise agreement and updates two lightweight linear layers during inference.

What would settle it

On a held-out institution, run SPA twice with identical few-shot data: once with the correct task graph and once with a randomized graph of the same edge count and duration budget. If the randomized-graph variant performs as well as the correct-graph variant, the graph's protocol content is not what drives the temporal gain and the central explanation fails.

Watch

Extended reading notes

Core claim

The central claim, stated the way a sympathetic reader would, is that spatial, temporal, and test-time shifts in surgical video are separable problems, each solvable with minimal annotation. Spatial adaptation trains a text-driven linear classifier on the frozen vision and text encoders, using the few labeled images plus textual phase names, so the model learns the institution's visual vocabulary. Temporal adaptation turns the institution's protocol into a directed task graph G=(V,E) with per-phase duration bounds, samples synthetic phase sequences from it with a Markovian rule, and trains a diffusion model in a hidden state space over phase encodings; at inference the diffusion model denoises coarse frame predictions into protocol-conforming sequences. Test-time adaptation adds two linear layers on the encoders and optimizes a mutual-agreement loss between reference-image, vision-language, and few-shot prediction streams, self-supervised on the test video itself, before the diffusion refinement runs. The paper reports the assembled pipeline outperforms prior few-shot methods on Cholec80, StrasBypass, BernBypass, and Autolaparo, and that the ablation isolates large contributions from both the task-graph module and test-time adaptation.

Load-bearing premise

The temporal module assumes the hand-written task graph — the allowed phase transitions and their min/max durations — faithfully encodes the target institution's real protocol, because every synthetic training sequence for the diffusion model is generated from that graph; the paper's own ablation shows a foreign graph can cut F1 from 55.80 to 42.86 on StrasBypass, so a wrong graph is not neutral, it is actively harmful.

Editorial extensions

If this is right

  • A hospital could install institution-specific phase recognition from 32 labeled images per phase plus a written protocol: at that budget the paper reports BernBypass F1 of 43.54, above the 42.4 of a full-shot model.
  • Even without a task graph, test-time adaptation alone raises F1 by roughly 6 to 14 points on every benchmark, at a measured cost of 22 seconds per 30-minute video.
  • Task-graph-guided diffusion cleans up the output qualitatively: predictions become contiguous phase blocks with clear boundaries instead of fragmented label noise, which is what makes them usable in the operating room.
  • Task graphs do not transfer symmetrically: the ablation shows Strasbourg's graph helps on BernBypass while Bern's graph hurts StrasBypass, so the graph must be authored at the target site.
  • Because only lightweight linear layers are trained on frozen encoders, customization runs in minutes on a single GPU, and the authors report the framework needs 200 to 300 times less labeled data than the full-shot alternative.

Reading between the lines

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

  • The mutual-agreement loss could double as an out-of-distribution alarm: videos where the reference, vision-language, and few-shot streams disagree are precisely the ones a surgeon should review, turning the adaptation step into patient-level monitoring.
  • The task graph is the last manual input; a natural next step is to infer it from unlabeled video at the target site, for instance by detecting phase boundaries from embedding shifts, which would make the pipeline nearly annotation-free.
  • Because the diffusion model trains only on graph-generated sequences, its temporal beliefs are capped by the graph's accuracy: a typo in a duration bound would silently bias the model in a way that supervised training on real video would not.
  • At 22 seconds per video, test-time adaptation is cheap enough to run per case or per surgeon, so the model's accuracy could compound over an operating day rather than remaining fixed between training and deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes SPA (Surgical Phase Anywhere), a framework for few-shot surgical phase recognition built on the frozen PeskaVLP vision-language foundation model. SPA combines three adaptation stages: a few-shot text-driven linear classifier over vision prototypes and text embeddings; a task-graph-guided diffusion model that generates synthetic phase sequences from a hand-specified institutional task graph and refines coarse predictions for temporal consistency; and a test-time adaptation step that aligns three prediction streams (reference, vision-language, and few-shot) through a self-supervised mutual-agreement loss. Experiments are reported on Cholec80, StrasbourgBypass, BernBypass, and Autolaparo at 1-, 16-, and 32-shot settings, with ablations separating the test-time adaptation and task-graph diffusion components. The central claim is state-of-the-art few-shot performance, including an assertion that SPA outperforms a full-shot RN50 model with only 32 labeled samples per class.

Significance. If the reported results are valid, SPA is practically valuable: it adapts a generalist surgical foundation model to a new institution using only textual phase definitions, a small labeled image set, and a hand-specified task graph, which matches a realistic clinical deployment scenario. The paper ships code, evaluates across four datasets spanning three procedure types, and provides an ablation study that isolates the contributions of test-time adaptation and task-graph diffusion. However, the headline comparison against full-shot models is currently confounded by the choice of backbone, and the task graph's construction and sensitivity are under-specified. Because the temporal prior is central to the few-shot gains, these issues must be resolved before the paper's strongest claims can be accepted.

major comments (3)
  1. [§4, Table 1] The claim that SPA 'outperforms full-shot models' with 32-shot data, e.g., BernBypass F1 43.54 vs 42.4, is not an apples-to-apples comparison. The full-shot row is an RN50 model taken from prior papers, whereas SPA and the few-shot baselines are built on the PeskaVLP foundation model, which is a substantially stronger backbone. This confounds the method with the underlying representation and does not demonstrate that 32 labeled samples suffice to surpass a full-shot model of comparable capacity. Please add full-shot results obtained by training a comparable head on the same PeskaVLP features (or full fine-tuning of PeskaVLP), or explicitly qualify the claim so that it does not imply a controlled method comparison.
  2. [§2.2, Eq. (2), Table 2] The task graph is the load-bearing temporal prior for the reported few-shot gains, yet the paper never states how G = (V, E) and the bounds [L_i^min, L_i^max] were constructed for each dataset, whether edges or duration bounds were tuned using validation or test knowledge, or how sensitive the results are to plausible graph errors. Table 2 (TG-Cross) shows that using BernBypass's task graph on StrasBypass drops 32-shot F1 from 55.80 to 42.86, essentially erasing most of the TG-In benefit, which confirms that graph fidelity strongly controls the outcome. Please report the graph construction protocol for each dataset, include robustness experiments with perturbed edges and duration bounds, and clarify that no test-set information was used to define the graphs. Without this, the 32-shot superiority could reflect favorable manual graph engineering rather than the adaptation framework itself.
  3. [§2.2, Eq. (2)] Equation (2) defines the synthetic-sequence generation rule using only the upper bound L_i^max: the next phase is sampled whenever the current duration has not exceeded L_i^max, and otherwise forced to transition. The lower bound L_i^min, introduced earlier in the same section, never appears in the generation rule. Consequently, generated sequences can contain phases of duration one frame, making the prior weaker than the stated duration bounds imply. Please either modify the generation rule to enforce L_i^min or explicitly justify why the lower bound is omitted.
minor comments (4)
  1. [§3, Table 1] The caption says results are averaged over three subsets but does not specify how the subsets are sampled, whether they are disjoint, or how seeds are chosen. Please report the subset construction procedure and, ideally, per-subset results to allow the reader to gauge variance.
  2. [§2.1] There is a typo in 'using the the frozen pre-trained visual encoder' and the notation 'ft i' in Eq. (1) should be typeset as a transpose (e.g., f_i^T).
  3. [Table 2] The column headers contain two successive '+TTA' entries; one of them appears to be mislabeled. Please clarify which ablation each column corresponds to in the table and the text.
  4. [References] Reference [26] lists the arXiv identifier as 2410.002632, which appears to be a typo; please verify the correct identifier.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the task-graph prior, few-shot classifier, and TTA are fitted to separate signals and no prediction reduces to a parameter fitted to that same prediction.

full rationale

SPA's three modules are fitted to distinct inputs. The few-shot classifier (Eq. 1) trains only the prototype weights w and multipliers alpha on labeled images, with text embeddings t_k fixed. The diffusion model (Eqs. 2-4) is trained on synthetic phase sequences generated from the hand-specified task graph G=(V,E) and duration upper bounds; the graph is an input prior, not a learned function of the test labels. TTA (Sec. 2.3) optimizes linear layers f_v and f_t with a self-supervised mutual-agreement loss on the unlabeled test video, then the diffusion model refines the fused prediction; the loss does not use ground-truth phases. Consequently, the reported F1 scores are measured on held-out labels that never entered any fitted parameter. The main caveat is task-graph fidelity: Table 2's TG-Cross row shows that an institution-mismatched graph can reduce performance (e.g., StrasBypass 32-shot 55.80 to 42.86), and Eq. (2) never uses the stated lower bound L_i^min. These are correctness/robustness concerns about how the external prior is constructed, not circularity. Self-citations to PeskaVLP [26] and SVL-Pretrain [27] are ordinary dependencies on an open-access frozen backbone; they do not assert a uniqueness theorem or smuggle in the target result.

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

The paper introduces no hidden analytic constants fitted to the target result. The trained weights of the classifier, diffusion model, and TTA layers are standard learnable parameters, and the task graph is a hand-specified input. The load-bearing assumptions are the fidelity of the task graph and the transferability of the frozen foundation model, both listed as axioms.

assumptions (5)
  • domain assumption The frozen PeskaVLP encoders provide transferable surgical visual and textual representations.
    SPA freezes these encoders and only trains linear layers; if the embeddings do not generalize to new institutions or procedures, the few-shot spatial adaptation cannot work. See Section 2.1.
  • domain assumption The institution-specific task graph G=(V,E) and phase duration bounds [L_i^min, L_i^max] correctly represent the target surgical protocol.
    All synthetic sequences for training the diffusion model are generated from this graph via Eq. 2; any error or incompleteness in the graph is propagated into the temporal refinement. Table 2 (TG-Cross) shows that an incompatible graph can reduce F1 by 12.94 on StrasBypass at 32-shot.
  • domain assumption Mutual agreement among the three prediction streams (reference, vision-language, few-shot) is a reliable self-supervision signal.
    The TTA loss in Section 2.3 assumes that co-agreement implies correctness; this heuristic is not validated independently on surgical videos and could reinforce shared errors.
  • domain assumption The natural language phase descriptions d_k are semantically aligned with the visual content of the phase in the target institution.
    Text embeddings are blended into the classifier weights via Eq. 1; if a description is ambiguous or mismatched to the local video appearance, the classifier will be biased. See Section 2.1.
  • standard math The diffusion forward and reverse processes in hidden state space can model discrete phase transitions after mixing with Gaussian noise.
    The method treats one-hot phase labels as continuous hidden states and applies Gaussian diffusion; this is a standard assumption for graph-constrained sequence generation, but the recovery of exact discrete transitions from denoised states is not proven in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recognizing Surgical Phases Anywhere: Few-Shot Test-time Adaptation and Task-graph Guided Refinement." pith.science (2026). https://pith.science/paper/2AVGWXCD

@misc{pith2026250620254,
  author       = {Pith},
  title        = {Pith review of: Recognizing Surgical Phases Anywhere: Few-Shot Test-time Adaptation and Task-graph Guided Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AVGWXCD}},
  note         = {Machine review of arXiv:2506.20254}
}
read the original abstract

The complexity and diversity of surgical workflows, driven by heterogeneous operating room settings, institutional protocols, and anatomical variability, present a significant challenge in developing generalizable models for cross-institutional and cross-procedural surgical understanding. While recent surgical foundation models pretrained on large-scale vision-language data offer promising transferability, their zero-shot performance remains constrained by domain shifts, limiting their utility in unseen surgical environments. To address this, we introduce Surgical Phase Anywhere (SPA), a lightweight framework for versatile surgical workflow understanding that adapts foundation models to institutional settings with minimal annotation. SPA leverages few-shot spatial adaptation to align multi-modal embeddings with institution-specific surgical scenes and phases. It also ensures temporal consistency through diffusion modeling, which encodes task-graph priors derived from institutional procedure protocols. Finally, SPA employs dynamic test-time adaptation, exploiting the mutual agreement between multi-modal phase prediction streams to adapt the model to a given test video in a self-supervised manner, enhancing the reliability under test-time distribution shifts. SPA is a lightweight adaptation framework, allowing hospitals to rapidly customize phase recognition models by defining phases in natural language text, annotating a few images with the phase labels, and providing a task graph defining phase transitions. The experimental results show that the SPA framework achieves state-of-the-art performance in few-shot surgical phase recognition across multiple institutions and procedures, even outperforming full-shot models with 32-shot labeled data. Code is available at https://github.com/CAMMA-public/SPA

Figures

Figures reproduced from arXiv: 2506.20254 by the authors.

Figure 1
Figure 1. Conventional adaptation struggles with performance drops across institutions due to spatial and procedural variations. Our institution-specific adaptation minimizes manual effort by requiring only textual phase definitions, a few labeled images, and a task graph from the target hospital, to adapt spatial understanding and enforce institution-specific temporal constraints. Deep learning models have made significant p… view at source ↗
Figure 2
Figure 2. Pipeline of SPA. During training, we first train (a) few-shot classifier for spa￾tial adaptation, and (b) task-graph guided diffusion model for temporal adaptation. During inference, TTA updates model parameters using self-supervision, followed by the diffusion model to ensure temporally coherent phase transitions. 2.1 Spatial adaptation via few-shot learning As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. 32-shot phase recognition shows SPA reduces noise and fragmentation [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Geo-RepNet: Geometry-Aware Representation Learning for Surgical Phase Recognition in Endoscopic Submucosal Dissection

    cs.CV 2025-07 conditional novelty 5.0 of 10

    Adding depth-derived geometric priors to a RepVGG backbone improves surgical phase classification on a new nine-phase ESD dataset.

Reference graph

Works this paper leans on

29 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Alapatt, D., Murali, A., Srivastav, V., Consortium, A., Mascagni, P., Padoy, N.: Jumpstarting surgical computer vision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 328–338. Springer (2024)

    work page 2024

  2. [2]

    In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part III 23

    Czempiel, T., Paschali, M., Keicher, M., Simson, W., Feussner, H., Kim, S.T., Navab, N.: Tecno: Surgical phase recognition with multi-stage temporal con- volutional networks. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part III 23. pp. 343–352. Sprin...

    work page 2020

  3. [3]

    Annals of surgery 273(4), 684–693 (2021)

    Garrow, C.R., Kowalewski, K.F., Li, L., Wagner, M., Schmidt, M.W., Engelhardt, S., Hashimoto, D.A., Kenngott, H.G., Bodenstedt, S., Speidel, S., et al.: Machine learning for surgical phase recognition: a systematic review. Annals of surgery 273(4), 684–693 (2021)

    work page 2021

  4. [4]

    In: Advancements In Med- ical Foundation Models: Explainability, Robustness, Security, and Beyond (2024)

    Honarmand, M., Jamal, M.A., Mohareri, O.: Vidlpro: A video-language pre- training framework for robotic and laparoscopic surgery. In: Advancements In Med- ical Foundation Models: Explainability, Robustness, Security, and Beyond (2024)

    work page 2024

  5. [5]

    arXiv preprint arXiv:2411.15421 (2024)

    Hu, M., Yuan, K., Shen, Y., Tang, F., Xu, X., Zhou, L., Li, W., Chen, Y., Xu, Z., Peng, Z., et al.: Ophclip: Hierarchical retrieval-augmented learning for ophthalmic surgical video-language pretraining. arXiv preprint arXiv:2411.15421 (2024)

    arXiv 2024

  6. [6]

    Nature medicine29(9), 2307–2316 (2023)

    Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T.J., Zou, J.: A visual–language foundation model for pathology image analysis using medical twitter. Nature medicine29(9), 2307–2316 (2023)

    2023

  7. [7]

    Advances in neural information processing systems36(2024)

    Ikezogwo, W., Seyfioglu, S., Ghezloo, F., Geva, D., Sheikh Mohammed, F., Anand, P.K., Krishna, R., Shapiro, L.: Quilt-1m: One million image-text pairs for histopathology. Advances in neural information processing systems36(2024)

    work page 2024

  8. [8]

    arXiv preprint arXiv:2501.09436 (2025)

    Jaspers, T.J., de Jong, R.L., Li, Y., Kusters, C.H., Bakker, F.H., van Jaarsveld, R.C., Kuiper, G.M., van Hillegersberg, R., Ruurda, J.P., Brinkman, W.M., et al.: Scaling up self-supervised learning for improved surgical foundation models. arXiv preprint arXiv:2501.09436 (2025)

    arXiv 2025

Show all 29 references
  1. [9]

    International journal of computer assisted radiology and surgery pp

    Lavanchy, J.L., Ramesh, S., Dall’Alba, D., Gonzalez, C., Fiorini, P., Müller-Stich, B.P., Nett, P.C., Marescaux, J., Mutter, D., Padoy, N.: Challenges in multi-centric generalization: phase and step recognition in roux-en-y gastric bypass surgery. International journal of comp...

    2024

  2. [10]

    Advances in Neural Information Processing Systems37, 6126–6153 (2025)

    Ma, C., Jiang, H., Chen, W., Li, Y., Wu, Z., Yu, X., Liu, Z., Guo, L., Zhu, D., Zhang, T., et al.: Eye-gaze guided multi-modal alignment for medical representa- tion learning. Advances in Neural Information Processing Systems37, 6126–6153 (2025)

    2025

  3. [11]

    Nature Biomedical Engineering1(9), 691–696 (2017)

    Maier-Hein, L., Vedula, S.S., Speidel, S., Navab, N., Kikinis, R., Park, A., Eisen- mann, M., Feussner, H., Forestier, G., Giannarou, S., et al.: Surgical data science for next-generation interventions. Nature Biomedical Engineering1(9), 691–696 (2017)

    2017

  4. [12]

    npj Digital Medicine5(1), 163 (2022)

    Mascagni, P., Alapatt, D., Sestini, L., Altieri, M.S., Madani, A., Watanabe, Y., Alseidi, A., Redan, J.A., Alfieri, S., Costamagna, G., et al.: Computer vision in surgery: from potential to clinical value. npj Digital Medicine5(1), 163 (2022)

    2022

  5. [13]

    arXiv preprint arXiv:2212.04155 (2022)

    Murali, A., Alapatt, D., Mascagni, P., Vardazaryan, A., Garcia, A., Okamoto, N., Mutter, D., Padoy, N.: Latent graph representations for critical view of safety assessment. arXiv preprint arXiv:2212.04155 (2022)

    2022 arXiv

  6. [14]

    In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention

    Özsoy, E., Pellegrini, C., Keicher, M., Navab, N.: Oracle: Large vision-language models for knowledge-guided holistic or domain modeling. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 455–465. Springer (2024)

    2024

  7. [15]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PMLR (2021)

    2021

  8. [16]

    Medical Image Analysis88, 102844 (2023) SPA: Surgical Phase Anywhere 11

    Ramesh, S., Srivastav, V., Alapatt, D., Yu, T., Murali, A., Sestini, L., Nwoye, C.I., Hamoud, I., Sharma, S., Fleurentin, A., et al.: Dissecting self-supervised learning methods for surgical computer vision. Medical Image Analysis88, 102844 (2023) SPA: Surgical Phase Anywhere 11

    2023

  9. [17]

    In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention

    Shakeri,F.,Huang,Y.,Silva-Rodríguez,J.,Bahig,H.,Tang,A.,Dolz,J.,BenAyed, I.: Few-shot adaptation of medical vision-language models. In: International Con- ference on Medical Image Computing and Computer-Assisted Intervention. pp. 553–563. Springer (2024)

    2024

  10. [18]

    Medical Image Analysis99, 103357 (2025)

    Silva-Rodriguez, J., Chakor, H., Kobbi, R., Dolz, J., Ayed, I.B.: A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis99, 103357 (2025)

    2025

  11. [19]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Silva-Rodriguez, J., Hajimiri, S., Ben Ayed, I., Dolz, J.: A closer look at the few- shot adaptation of large vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 23681–23690 (2024)

    2024

  12. [20]

    IEEE transactions on medical imaging36(1), 86–97 (2016)

    Twinanda, A.P., Shehata, S., Mutter, D., Marescaux, J., De Mathelin, M., Padoy, N.: Endonet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging36(1), 86–97 (2016)

    2016

  13. [21]

    IEEE Transactions on Medical Imaging (2024)

    Wang, H., Yang, G., Zhang, S., Qin, J., Guo, Y., Xu, B., Jin, Y., Zhu, L.: Video- instrument synergistic network for referring video instrument segmentation in robotic surgery. IEEE Transactions on Medical Imaging (2024)

    2024

  14. [22]

    arXiv preprint arXiv:2210.10163 (2022)

    Wang, Z., Wu, Z., Agarwal, D., Sun, J.: Medclip: Contrastive learning from un- paired medical images and text. arXiv preprint arXiv:2210.10163 (2022)

    2022 arXiv

  15. [23]

    In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention

    Wang, Z., Lu, B., Long, Y., Zhong, F., Cheung, T.H., Dou, Q., Liu, Y.: Autolaparo: A new dataset of integrated multi-tasks for image-guided surgical automation in laparoscopic hysterectomy. In: International Conference on Medical Image Com- puting and Computer-Assisted Interve...

    2022

  16. [24]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Wu, C., Zhang, X., Zhang, Y., Wang, Y., Xie, W.: Medklip: Medical knowledge enhanced language-image pre-training for x-ray diagnosis. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21372–21383 (2023)

    2023

  17. [25]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Yuan, K., Srivastav, V., Navab, N., Padoy, N.: Hecvl: Hierarchical video-language pretraining for zero-shot surgical phase recognition. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 306–316. Springer (2024)

    2024

  18. [26]

    arXiv preprint arXiv:2410.002632(2024)

    Yuan, K., Srivastav, V., Navab, N., Padoy, N., et al.: Procedure-aware surgi- cal video-language pretraining with hierarchical knowledge augmentation. arXiv preprint arXiv:2410.002632(2024)

    2024 arXiv

  19. [27]

    Medical Image Analysis p

    Yuan, K., Srivastav, V., Yu, T., Lavanchy, J.L., Marescaux, J., Mascagni, P., Navab, N., Padoy, N.: Learning multi-modal representations by watching hundreds of surgical video lectures. Medical Image Analysis p. 103644 (2025)

    2025

  20. [28]

    arXiv preprint arXiv:2111.03930 (2021)

    Zhang, R., Fang, R., Zhang, W., Gao, P., Li, K., Dai, J., Qiao, Y., Li, H.: Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930 (2021)

    2021 arXiv

  21. [29]

    arXiv preprint arXiv:2506.19694 (2025)

    Zhou, Y., Bi, Y., Tong, W., Wang, W., Navab, N., Jiang, Z.: Ultraad: Fine-grained ultrasound anomaly classification via few-shot clip adaptation. arXiv preprint arXiv:2506.19694 (2025)

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.