Adaptation of Multi-modal Representation Models for Multi-task Surgical Computer Vision

Britty Baby; Nicolas Padoy; Soham Walimbe; Vinkle Srivastav

REVIEW 5 major objections 6 minor 17 references

Adaptation of Multi-modal Representation Models for Multi-task Surgical Computer Vision

T0 review · 5 major / 6 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 MML-SurgAdapt shows that a single CLIP-based model trained with one positive label per image can match task-specific models on phase recognition, CVS assessment, and action triplet recognition.

desk verdict Worth reading for label-efficient multi-task surgical video, but the missing split-disjointness check on Endoscapes2023 undercuts the parity claims until resolved. read the letter →

arxiv 2507.05020 v2 pith:NAPUMY3D submitted 2025-07-07 cs.CV cs.AI

Soham Walimbe , Britty Baby , Vinkle Srivastav , Nicolas Padoy This is my paper

classification cs.CVcs.AI

keywords Vision-LanguageModelsSinglePositiveMulti-LabelCLIPadaptationMulti-tasksurgicalAIPhaserecognitionCriticalViewofSafetyActiontripletsAnnotation-efficientlearning

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

Surgical video analysis typically needs separate models for phase recognition, Critical View of Safety assessment, and action triplet recognition. This paper tries to establish that one CLIP-based model, trained with a single positive label per image across all three tasks, can match those task-specific models. The authors combine Cholec80, Endoscapes2023, and CholecT50 into a 110-label multi-task setup and use Single Positive Multi-Label learning plus Hill loss to handle the missing labels. They report phase F1 of 74.2, CVS mAP of 55.9, and triplet AP-IVT of 30.3, on par with or better than task-specific baselines while using 23% fewer labels. If correct, this makes multi-task surgical AI annotation cheaper and moves toward one adaptable model per procedure.

What carries the argument

The engine is a CLIP image/text encoder pair fine-tuned end-to-end, with the 110 task labels turned into fixed natural-language prompts such as 'a photo of the phase Calot Triangle Dissection'. A graph convolution network built from text-embedding similarities refines the label embeddings, and cosine similarity between image and label embeddings is sigmoid-transformed into probabilities. The Hill loss does the heavy lifting for partial annotations: it reweights negative labels so that labels already predicted positive are down-weighted as likely false negatives, and applies a modified MSE loss to negatives. This combination is what lets one positive label per image stand in for a full multi-label annotation.

What would settle it

Take a held-out set of images from CholecT50 with complete triplet annotations, train MML-SurgAdapt under the single-positive protocol, and measure AP-IVT against a version trained with all labels present. If the single-positive version loses more than 5 points on this subset while matching on the paper's test split, then the parity claim depends on the test labels being sparse too.

Watch

Extended reading notes

Core claim

The paper's central claim is that MML-SurgAdapt, a CLIP ViT-L/14 model adapted with a graph convolution layer and Hill loss, achieves task-general multi-label prediction of surgical phase, Critical View of Safety, and action triplets from images of laparoscopic cholecystectomy even when the training set carries only one positive label per image. On the combined 110-label task set, the model reaches phase F1 74.2, CVS mAP 55.9, and triplet AP-IVT 30.3, compared with 65.9, 57.4, and 29.9 for the strongest task-specific or vision-only baselines listed. The authors also show it outperforms existing SPML frameworks (DualCoOp, VLPL, HSPNet) and that replacing the full partial-positive label set with one randomly chosen positive per image costs only about 0.5% performance while cutting the annotation burden by 23%. This is presented as the first use of single-positive multi-label learning to integrate multiple surgical tasks into a single model.

Load-bearing premise

The load-bearing premise is that setting every unobserved label to negative during training does not bias the model too much, so one positive label per image plus Hill loss reweighting can stand in for full annotations.

Editorial extensions

If this is right

A single model trained on one positive label per image can serve phase, CVS, and triplet recognition simultaneously, eliminating the need for three separate task-specific models.
Annotation cost drops by 23% with only about 0.5% performance loss, so clinical labelling effort can be redirected to more complex labels such as CVS.
Hill loss outperforms WAN and SPLC on this surgical multi-task setup, giving a concrete loss choice for future SPML surgical work.
Because the model beats existing SPML frameworks, the combination of CLIP adaptation, GCN label refinement, and Hill loss appears to transfer better to surgical data than prior SPML designs.
The framework is not tied to these three tasks; any surgical task whose labels can be phrased as prompts can be added to the 110-label vocabulary.

Reading between the lines

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

If the 0.5% gap between SPML and partial-positive training is stable across random single-positive draws, then the information lost by dropping extra positives is small; a natural test is to vary the number of positives per image and measure the performance curve.
The small GCN gain (~0.5%) suggests the label refinement module may matter less than the CLIP prior and the Hill reweighting; ablating the GCN entirely might enable cheaper fine-tuning.
Because the evaluation covers a single procedure, the claim that this generalizes to other surgeries is untested; a strong test would be adding a fourth task, such as tool presence or bleeding detection, and checking whether the 110-label model retains parity.
Treating unobserved labels as negatives could fail on tasks where false negatives are not uniformly distributed; the paper's Hill reweighting smooths this, but a task with many more labels per image than three would stress the assumption.

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.

Desk Editor's Note

read the letter

The one thing to know: this is a legitimate label-efficient multi-task adaptation of CLIP for surgical video, with a sensible use of single-positive multi-label learning. The catch is that the paper never verifies that the Endoscapes2023 test split is disjoint from the Cholec80 or CholecT50 training videos, so the headline 'comparable to task-specific models' is not yet proven.

The genuinely new part is applying SPML to merge three surgical datasets with partial annotations into one 110-label multi-task model. That is a useful step for the community. The paper also does a few things well: it includes a loss-function comparison with standard deviations over five runs, it is honest about the GCN ablation being small (0.5% with and without), and the code is public. The Hill loss choice is well motivated and the reweighting story is plausible.

The soft spots are real but mostly addressable. The biggest is the split overlap. Section 3 says overlapping videos were removed from Cholec80, but it says nothing about Endoscapes2023 or CholecT50 beyond 'split per [5]' and 'Rendezvous split'. If Endoscapes2023 test frames come from surgeries that also appear in Cholec80 or CholecT50 training, then the CVS and triplet numbers in Table 2 could be inflated for reasons orthogonal to the method. That needs a documented video-level verification. Also, Table 2 has no error bars, even though Table 3 reports standard deviations; presenting single runs for the central comparison is a step backward. Model selection is ambiguous: three validation criteria are listed, but it is not clear how they are combined. SPML baselines settings are under-specified. These are not load-bearing if the split overlap is resolved.

Who is this for? Researchers working on surgical phase, CVS, or triplet recognition who want a single model and fewer labels. The method is a reasonable tool, not a breakthrough. It deserves a serious referee; the conditional should hinge on the authors either proving split exclusivity or rerunning with a clean split and reporting error bars in the main table. I would engage with it, but with a raised eyebrow.

Referee Report

5 major / 6 minor

Summary. The paper introduces MML-SurgAdapt, a CLIP-based multi-task framework for laparoscopic cholecystectomy that jointly addresses phase recognition, Critical View of Safety (CVS) assessment, and action triplet recognition. The method converts task-specific labels into text prompts, applies a Graph Convolutional Network to refine label embeddings, and trains with Single Positive Multi-Label (SPML) learning using Hill loss to handle missing labels when merging Cholec80, Endoscapes2023, and CholecT50. The authors report that MML-SurgAdapt performs comparably to task-specific baselines and outperforms existing SPML frameworks, while reducing the number of labels by 23%.

Significance. If the central claims hold, the paper makes a practical contribution: it is, to my knowledge, the first application of SPML to integrate multiple surgical tasks in a single VLM-based model, with public datasets, released code, and a loss-function comparison that includes standard deviations over five runs. These concrete strengths make the contribution reproducible in principle. However, the significance is currently tempered by incomplete evidence on split disjointness, an underspecified model-selection rule, and the absence of uncertainty estimates in the main comparison table, all of which are needed before the parity claim can be accepted.

major comments (5)

[Section 3, Dataset splits] The manuscript does not provide evidence that the Endoscapes2023 test split is disjoint from the Cholec80 and CholecT50 training and validation splits. It states that overlapping videos were removed from Cholec80, but it does not report overlap counts, video identifiers, or a verification step for the other two datasets. Because the central parity claims in Table 2 rest on evaluating genuinely unseen test videos, please provide a video-level overlap table or an automated check showing that no test video of any of the three datasets appears in any training or validation split of any dataset, and clarify which dataset's test videos were affected by the removal.
[Section 3, Implementation details] The model-selection procedure is underspecified. The sentence 'The best model is saved based on three validation criteria: highest mAP, lowest loss on partial positive validation ground truths and lowest loss on single positive validation ground truths' does not explain how these criteria are combined when they conflict. This creates a risk that the reported test numbers in Table 2 are selected post hoc; please define a single, pre-registered selection rule (for example, validation mAP on the union of tasks) and report the corresponding test results.
[Table 2 and Section 4.1] The main comparison table reports point estimates without standard deviations or confidence intervals, although Section 4.2 shows that five-run variability is non-negligible (for example, Hill CVS mAP is 55.9±2.8 in Table 3, and WAN triplet AP-IVT is 30.9±0.5). Please report mean±standard deviation, or an equivalent uncertainty measure, for all entries in Table 2, or state explicitly that the table reports a single seed and justify why the comparisons remain meaningful.
[Section 4.1, SPML Models] The SPML baselines are under-specified. The statement 'We use the default loss functions as specified in the original papers for each model' is insufficient for reproducibility and fairness. Please report the training epochs, learning rate, batch size, backbone, prompt templates, and dataset-specific settings used for DualCoOp, VLPL, and HSPNet, so that readers can assess whether the comparison is fair.
[Section 2.1 and Section 4.2] The 23% label-reduction claim is not derived. Please specify the exact formula used to compare the number of positive labels in the SPML setup with the partial-positive setup, and state whether the reduction is measured per image, per label, or per annotation event.

minor comments (6)

[Section 4.1] The heading 'Comparision between Architectures' contains a typo; it should read 'Comparison between Architectures'.
[Figure 1] The figure caption contains corrupted glyph sequences (for example, '7 7 7' and '*UDSK') that appear to be a font-embedding artifact; please regenerate the figure and caption.
[Section 2.2 and Section 3] The hyperparameters τ, the number of GCN layers L, and the top-K value used to sparsify the adjacency matrix are used in the model but their numerical values are never reported; please list them in Implementation details.
[Section 3, Dataset splits] The text says the Endoscapes2023 split is 'performed according to [5]' but does not state how the official split interacts with the overlap-removal procedure; please clarify this interaction explicitly.
[Section 4.3] The GCN ablation reports only a 0.5% improvement; please state which metric or metrics this improvement is measured on (phase F1, CVS mAP, triplet AP, or an average across tasks).
[Table 3] The table title 'Combined Performance comparison of objective functions' is ambiguous; please clarify whether 'Combined' means averaged across tasks, evaluated on the combined multi-task model, or something else.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical comparisons on held-out public test labels, with no construction making a prediction equal to its input.

full rationale

The paper's derivation chain is an empirical training-and-evaluation loop: MML-SurgAdapt is trained on partially annotated images, and the reported phase F1, CVS AP/mAP, and triplet AP values are measured against held-out test labels from Cholec80, Endoscapes2023, and CholecT50. No equation or fitted parameter is reused as the predicted quantity. The SPML setup intentionally creates false negatives by treating unobserved labels as negatives, and the Hill loss is chosen to mitigate them; this is a stated modeling assumption, not a circular reduction. Self-citations appear for dataset splits and baselines (e.g., Rendezvous, Endoscapes2023), but these support external benchmarks and are not invoked as a uniqueness theorem or as a substitute for evaluation. The '23% label reduction' is computed from the number of labels used in the SPML versus Partial Positive setups, not derived from task performance, so it is not a fitted-input prediction. The concern raised by the skeptic about potential overlap between Endoscapes2023 test videos and Cholec80/CholecT50 training videos is a data-leakage/correctness risk, but it is not an instance of definitional circularity in the paper's reasoning. Overall, the central claims stand on independent held-out evaluation and do not reduce to their inputs by construction.

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

The central claim depends on standard pretrained CLIP features and on the transfer of SPML assumptions from natural images to surgery; no new entities are introduced, and the hyperparameters of Hill loss and the GCN are not fully specified.

free parameters (5)

Hill loss reweighting lambda = 1.5
Set to 1.5 in Section 3; no sensitivity analysis is reported.
Hill loss logit margin m = not reported
Appears in Eq. 5; the value is not stated in the paper, likely taken from the original Hill loss paper.
Temperature tau = not reported
Used in sigmoid scaling in Eq. 4; likely inherited from CLIP, but not specified.
Number of GCN layers L = not reported
GCN depth is not reported, so the complexity of the label graph module is unspecified.
Top-K neighbors in adjacency matrix = not reported
Sparsification of the adjacency matrix retains only the top-K connections, but K is not stated.

assumptions (4)

domain assumption Pretrained CLIP weights transfer to surgical video frames
The text and image encoders are initialized with CLIP ViT-L/14 pretrained on web image-text data; the framework assumes these features transfer to surgery. Section 2.2.
domain assumption Hill loss reweights false negatives sufficiently for surgical SPML
The paper assumes Hill loss mitigates the false negatives created by treating unobserved labels as negative, transferring a natural-image SPML assumption to surgery. Section 2.2, Eq. 6-7.
domain assumption The 110-label union is a coherent multi-label space
The paper assumes phase, CVS, and triplet labels from the three datasets can be combined without conflict. Section 2.1.
domain assumption Removing overlapping videos fully prevents train/test leakage
The paper assumes that removing overlapping videos from Cholec80 with CholecT50 eliminates data leakage across train/test splits. Section 3, Dataset splits.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptation of Multi-modal Representation Models for Multi-task Surgical Computer Vision." pith.science (2026). https://pith.science/paper/NAPUMY3D

@misc{pith2026250705020,
  author       = {Pith},
  title        = {Pith review of: Adaptation of Multi-modal Representation Models for Multi-task Surgical Computer Vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NAPUMY3D}},
  note         = {Machine review of arXiv:2507.05020}
}

read the original abstract

Surgical AI often involves multiple tasks within a single procedure, like phase recognition or assessing the Critical View of Safety in laparoscopic cholecystectomy. Traditional models, built for one task at a time, lack flexibility, requiring a separate model for each. To address this, we introduce MML-SurgAdapt, a unified multi-task framework with Vision-Language Models (VLMs), specifically CLIP, to handle diverse surgical tasks through natural language supervision. A key challenge in multi-task learning is the presence of partial annotations when integrating different tasks. To overcome this, we employ Single Positive Multi-Label (SPML) learning, which traditionally reduces annotation burden by training models with only one positive label per instance. Our framework extends this approach to integrate data from multiple surgical tasks within a single procedure, enabling effective learning despite incomplete or noisy annotations. We demonstrate the effectiveness of our model on a combined dataset consisting of Cholec80, Endoscapes2023, and CholecT50, utilizing custom prompts. Extensive evaluation shows that MML-SurgAdapt performs comparably to task-specific benchmarks, with the added advantage of handling noisy annotations. It also outperforms the existing SPML frameworks for the task. By reducing the required labels by 23%, our approach proposes a more scalable and efficient labeling process, significantly easing the annotation burden on clinicians. To our knowledge, this is the first application of SPML to integrate data from multiple surgical tasks, presenting a novel and generalizable solution for multi-task learning in surgical computer vision. Implementation is available at: https://github.com/CAMMA-public/MML-SurgAdapt

Figures

Figures reproduced from arXiv: 2507.05020 by the authors.

**Figure 1.** Proposed model architecture of MML-SurgAdapt (3) an evaluation of loss functions and benchmarks against task-specific models; and (4) a scalable framework for surgical and broader multimodal applications. 2 Methodology To address multi-task learning in the surgical domain, we use the popular VisionLanguage Model (VLM), CLIP [8]. For our multi-label classification setup, we convert task-specific labels into textual … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages

[5]

arXiv preprint ar Xiv:2312.12429 (2023)

Murali, A., Alapatt, D., Mascagni, P., Vardazaryan, A., G arcia, A., Okamoto, N., Costamagna, G., Mutter, D., Marescaux, J., Dallemagne, B., et al.: The endoscapes dataset for surgical scene segmentation, object detection , and critical view of safety assessment: oﬃcial splits and benchmark. arXiv preprint ar Xiv:2312.12429 (2023)

arXiv 2023
[1]

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

Alapatt, D., Murali, A., Srivastav, V., Consortium, A., M ascagni, P., Padoy, N.: Jumpstarting surgical computer vision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 3 28–338. Springer (2024)

work page 2024
[2]

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

Cole, E., Mac Aodha, O., Lorieul, T., Perona, P., Morris, D ., Jojic, N.: Multi-label learning from single positive labels. In: Proceedings of th e IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 933–942 (20 21)

work page
[3]

In: Proceedings of the IEEE conference on computer vision and pa ttern recognition

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pa ttern recognition. pp. 770–778 (2016)

2016
[4]

arXiv preprint arXiv:1609.02907 (2016)

Kipf, T.N., Welling, M.: Semi-supervised classiﬁcation with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

arXiv 2016
[6]

In: Internationa l Conference on Medi- cal Image Computing and Computer-Assisted Intervention

Murali, A., Alapatt, D., Mascagni, P., Vardazaryan, A., G arcia, A., Okamoto, N., Mutter, D., Padoy, N.: Encoding surgical videos as latent sp atiotemporal graphs for object and anatomy-driven reasoning. In: Internationa l Conference on Medi- cal Image Computing and Computer-Assisted Intervention. p p. 647–657. Springer (2023)

work page 2023
[7]

Medical I mage Analysis 78, 102433 (2022) 10 S

Nwoye, C.I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni , P., Mutter, D., Marescaux, J., Padoy, N.: Rendezvous: Attention mechanism s for the recognition of surgical action triplets in endoscopic videos. Medical I mage Analysis 78, 102433 (2022) 10 S. Walimbe et al

work page 2022
[8]

In: International conferen ce on machine learning

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

work page 2021

Show all 17 references

[9]

International journal of computer assisted radiology and s urgery 16, 1111–1119 (2021)

Ramesh, S., Dall’Alba, D., Gonzalez, C., Yu, T., Mascagni , P., Mutter, D., Marescaux, J., Fiorini, P., Padoy, N.: Multi-task temporal convolutional networks for joint recognition of surgical phases and steps in gastri c bypass procedures. International journal of computer ass...

2021
[10]

Advances in Neural Information P rocessing Systems 35, 30569–30582 (2022)

Sun, X., Hu, P., Saenko, K.: Dualcoop: Fast adaptation to multi-label recognition with limited annotations. Advances in Neural Information P rocessing Systems 35, 30569–30582 (2022)

2022
[11]

IEEE transactions on medical imaging 36(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 la paroscopic videos. IEEE transactions on medical imaging 36(1), 86–97 (2016)

2016
[12]

In: Proceedings of the 31st ACM International Confe rence on Multimedia

Wang, A., Chen, H., Lin, Z., Ding, Z., Liu, P., Bao, Y., Yan , W., Ding, G.: Hierar- chical prompt learning using clip for multi-label classiﬁc ation with single positive labels. In: Proceedings of the 31st ACM International Confe rence on Multimedia. pp. 5594–5604 (2023)

2023
[13]

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

Xing, X., Xiong, Z., Stylianou, A., Sastry, S., Gong, L., Jacobs, N.: Vision- language pseudo-labels for single-positive multi-label l earning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Reco gnition. pp. 7799–7808 (2024)

2024
[14]

In: Proceedings of the IEEE/CVF Conferenc e on Computer Vi- sion and Pattern Recognition

Yang, L., Zhang, R.Y., Wang, Y., Xie, X.: Mma: Multi-moda l adapter for vision- language models. In: Proceedings of the IEEE/CVF Conferenc e on Computer Vi- sion and Pattern Recognition. pp. 23826–23837 (2024)

2024
[15]

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

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

2024
[16]

arXiv preprint arXiv:2112.07368 (2021)

Zhang, Y., Cheng, Y., Huang, X., Wen, F., Feng, R., Li, Y., Guo, Y.: Simple and robust loss design for multi-label learning with missin g labels. arXiv preprint arXiv:2112.07368 (2021)

2021 arXiv
[17]

International Journal of Computer Vision 130(9), 2337–2348 (2022)

Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International Journal of Computer Vision 130(9), 2337–2348 (2022)

2022

Pith tools

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

[1] [5]

arXiv preprint ar Xiv:2312.12429 (2023)

Murali, A., Alapatt, D., Mascagni, P., Vardazaryan, A., G arcia, A., Okamoto, N., Costamagna, G., Mutter, D., Marescaux, J., Dallemagne, B., et al.: The endoscapes dataset for surgical scene segmentation, object detection , and critical view of safety assessment: oﬃcial splits and benchmark. arXiv preprint ar Xiv:2312.12429 (2023)

arXiv 2023

[2] [1]

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

Alapatt, D., Murali, A., Srivastav, V., Consortium, A., M ascagni, P., Padoy, N.: Jumpstarting surgical computer vision. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 3 28–338. Springer (2024)

work page 2024

[3] [2]

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

Cole, E., Mac Aodha, O., Lorieul, T., Perona, P., Morris, D ., Jojic, N.: Multi-label learning from single positive labels. In: Proceedings of th e IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 933–942 (20 21)

work page

[4] [3]

In: Proceedings of the IEEE conference on computer vision and pa ttern recognition

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pa ttern recognition. pp. 770–778 (2016)

2016

[5] [4]

arXiv preprint arXiv:1609.02907 (2016)

Kipf, T.N., Welling, M.: Semi-supervised classiﬁcation with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016)

arXiv 2016

[6] [6]

In: Internationa l Conference on Medi- cal Image Computing and Computer-Assisted Intervention

Murali, A., Alapatt, D., Mascagni, P., Vardazaryan, A., G arcia, A., Okamoto, N., Mutter, D., Padoy, N.: Encoding surgical videos as latent sp atiotemporal graphs for object and anatomy-driven reasoning. In: Internationa l Conference on Medi- cal Image Computing and Computer-Assisted Intervention. p p. 647–657. Springer (2023)

work page 2023

[7] [7]

Medical I mage Analysis 78, 102433 (2022) 10 S

Nwoye, C.I., Yu, T., Gonzalez, C., Seeliger, B., Mascagni , P., Mutter, D., Marescaux, J., Padoy, N.: Rendezvous: Attention mechanism s for the recognition of surgical action triplets in endoscopic videos. Medical I mage Analysis 78, 102433 (2022) 10 S. Walimbe et al

work page 2022

[8] [8]

In: International conferen ce on machine learning

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

work page 2021