Cycle Context Verification for In-Context Medical Image Segmentation

Huazhu Fu; Liangli Zhen; Shishuai Hu; Yong Xia; Zehui Liao

REVIEW 4 major objections 5 minor 32 references

Cycle Context Verification for In-Context Medical Image Segmentation

T0 review · 4 major / 5 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 The paper claims that a cycle self-verification score, computed by swapping query and context roles, can serve as an implicit measure of initial query segmentation quality, and that optimizing a query-specific prompt against this score…

desk verdict A genuinely novel test-time self-verification idea for ICL segmentation, with consistent but modest gains and an unvalidated proxy assumption at the load-bearing point. read the letter →

arxiv 2507.08357 v1 pith:ADNY6YJ7 submitted 2025-07-11 cs.CV

Shishuai Hu , Zehui Liao , Liangli Zhen , Huazhu Fu , Yong Xia This is my paper · ORCID

classification cs.CV

keywords in-contextlearningmedicalimagesegmentationcyclecontextverificationquery-specificprompttest-timeoptimizationself-verificationfoundationmodelsDicecoefficient

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

This paper claims that an in-context medical segmentation model can grade its own prediction without any ground truth on the query. After the model segments the query image, Cycle Context Verification (CCV) swaps roles: the query image plus its predicted mask become the context, and each original in-context image becomes the new query. The Dice accuracy of those secondary re-predictions, called A, is taken as an implicit score of the initial query mask. A query-specific additive prompt is then optimized to raise A, and the resulting enhanced query image yields the final segmentation. If the proxy is valid, CCV is a training-free, model-agnostic way to improve in-context segmentation when annotations are scarce.

What carries the argument

The central object is the cycle ICL pipeline and its proxy score A. Given an initial prediction $y_t^p = M(x_t, S)$, the model is re-fed with swapped roles: the new context is $\{(x_t, y_t^p)\}$ and the new query is an original in-context image $x_n$, producing $y_n^p = M(x_n, \{(x_t, y_t^p)\})$. The accuracy is $A = \frac{1}{N}\sum_{n=1}^{N} \mathrm{DSC}(y_n^p, y_n)$, averaged over all in-context pairs, and it serves both as the verification signal and as the loss: $L = 1 - A$. A learnable, query-specific prompt $P$ enters as $\hat{x}_t = x_t + P$, and optimization halts when $A > 0.9$ or after 20 iterations, converting the proxy into improved alignment without updating the frozen model.

What would settle it

Compute, on a set of test queries with fixed in-context pairs, the correlation between A and the true query Dice, $\mathrm{DSC}(y_t^p, y_t)$. If the correlation is absent or negative, A is not an implicit measure of query quality. A sharper version is to record A and the true query Dice at every prompt-optimization iteration: the central claim predicts they rise together, so any run where A increases while the true query Dice decreases would disprove the proxy.

Watch

Extended reading notes

Core claim

The core discovery, on the paper's own terms, is that verification accuracy A computed by feeding the predicted query mask back as context and measuring how well the model re-segments the known in-context images serves as an implicit measure of the initial query prediction's reliability. The paper asserts that if the initial query prediction deviates substantially from the ground truth, the swapped-context re-predictions degrade, so optimizing A with a learnable, spatially aligned prompt P added to the query image should improve both A and the final query mask. Experiments on seven medical image segmentation datasets using UniverSeg and SegGPT show that equipping Random Selection, VPR, DualSC, or InMeMo with CCV raises average Dice by roughly 1.8 to 3.0 points, and ablations show that both the cycle pipeline and the prompt optimization are needed.

Load-bearing premise

The whole procedure rests on the assumption that the accuracy of the swapped, secondary predictions on the known in-context masks genuinely tracks the quality of the initial query mask, an assumption the paper states but does not derive or measure.

Editorial extensions

If this is right

CCV improves segmentation for both a medical-specific ICL backbone (UniverSeg) and a generalist backbone (SegGPT) across all eight tasks tested, so the benefit does not depend on a single architecture.
Because CCV operates at test time with the model frozen, it can be stacked on existing context-selection methods (Random Selection, VPR, DualSC) and an existing prompt-based method (InMeMo), each gaining roughly 1.8 to 3.0 average Dice points.
A query-specific prompt optimized per query can outperform a task-specific prompt such as InMeMo when in-context data are scarce, since it does not require labeled data to train a shared prompt.
Ablations indicate that neither component alone is enough: the cycle pipeline by itself with whole-model updating hurts performance, and prompt optimization without the cycle objective gives only a small gain; the two together are what works.

Reading between the lines

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

A testable extension is to measure the correlation between A and the true query Dice across a large set of queries; the paper's central premise predicts a clear positive correlation, and that number would tell practitioners when to trust the proxy.
Because P is optimized to make in-context re-predictions accurate, a risk is that P overfits the in-context masks rather than improving the query mask; checking whether the final query mask improves on queries whose in-context pairs are very dissimilar would expose such overfitting.
The framework generalizes beyond medical imaging: any in-context segmentation model with at least one labeled reference mask could use the same cycle score as a test-time objective, though the medical setting benefits most precisely because annotations are scarce.

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

Short version: this is a genuinely new test-time trick for ICL segmentation, and the empirical evidence is decent—gains on seven datasets with two backbones—but the core assumption that the cycle accuracy A tracks query quality is never validated. I think it's worth a serious referee, but the authors need to show A correlates with query Dice, or at least add error bars and a cleaner ablation.

The cyclic swap idea is the real contribution: take the query prediction as the new context, re-predict an in-context mask, and use its Dice against the known in-context label as a proxy for whether the query prediction was any good. Then optimize a per-query additive prompt against that proxy. That's not something I've seen in VPR, DualSC, or InMeMo—it's a self-verification loop, not another retrieval or prompt-learning scheme. The paper is also honest about the setting: frozen model, no extra training, plug-and-play on top of existing context selection. The visuals in Fig. 2 do show the prompt changing the prediction for the better.

The soft spots are real but not disqualifying on their own. The proxy correlation is asserted in Section 2.2 with no data behind it. The prompt is an unconstrained additive perturbation optimized directly against the small in-context label set, so it could in principle overfit the verification objective without helping the query—and with N=1 for SegGPT the risk is concrete. That said, the consistent improvements across eight tasks and two backbones are suggestive, and I don't think there's a hidden circularity: the labels used for A are legitimate in-context ground truths, not query labels. The bigger gaps are statistical: single-run numbers, no significance tests. Table 2's ablation is muddy: 'removing prompt optimization' actually means updating the whole model, which collapses, so it doesn't isolate the contribution of the prompt. I'd also want the optimizer details and inference-time cost.

Who should read it: anyone working on ICL for medical segmentation or test-time adaptation. The practical payoff is small but the mechanism is transferable.

If I were the editor: yes, send to referees. The idea is novel enough that the field should get a clean version of it, but the proxy needs validation or at least a correlation plot, and the results need error bars.

Referee Report

4 major / 5 minor

Summary. The paper proposes Cycle Context Verification (CCV), a test-time framework for in-context medical image segmentation. Given a query image and in-context image-mask pairs, CCV first obtains an initial query prediction, then swaps roles: the query image and its predicted mask become the new context, and an in-context image becomes the new query. The accuracy A of the resulting secondary predictions against the available in-context masks (Dice, Eq. 3) is used as an implicit measure of the initial query prediction quality. A query-specific additive prompt P is optimized to maximize A (equivalently minimize L = 1 - A, Eq. 5), and the final segmentation is produced with the optimized input. Experiments with UniverSeg and SegGPT on seven datasets, combined with three context-selection baselines and the InMeMo enhancement method, report consistent DSC improvements.

Significance. If the central proxy claim is valid, CCV is an appealing train-free, model-agnostic, plug-and-play method that requires no additional annotated data beyond the in-context pairs already used. The paper has concrete strengths: it evaluates on seven datasets with two distinct ICL backbones, integrates with several context-selection baselines, and reports consistent improvements in Table 1. The method is simple and potentially useful for limited-context clinical scenarios. However, the significance of the contribution rests entirely on the unvalidated assertion that maximizing A improves query segmentation; the current evidence is not yet sufficient to support that assertion, and the statistical basis of the reported gains is thin.

major comments (4)

[Sec. 2.2, Eq. (3)] The load-bearing claim that A is an implicit measure of the reliability of y_p_t is asserted without derivation, correlation analysis, or failure-case study. A direct validation would be a per-sample analysis correlating A with DSC(y_p_t, y_t_gt), or a controlled experiment where optimizing P to maximize A is compared with optimizing P against a constant or random objective. Without this, optimizing L = 1 - A can overfit the unconstrained additive prompt P (Eq. 4) to the in-context ground-truth masks (only 8 or 1 pairs), improving A without improving and possibly worsening the query segmentation.
[Table 1] All reported numbers are single-run, with no error bars, multiple seeds, or significance tests. The gains of 1.79-3.02 DSC points could be within run-to-run variation, especially because context selection and prompt initialization involve randomness. The claim of superiority over existing methods needs paired significance testing or at least multiple runs with different context samples.
[Sec. 3.3, Table 2] The ablation labeled 'removing prompt optimization' does not actually remove prompt optimization while keeping the cycle pipeline; it updates the entire ICL model with a small learning rate for 20 iterations. This change conflates 'no prompt optimization' with 'destructive fine-tuning of the backbone.' A cleaner control would keep the model frozen and either remove the prompt update entirely or optimize the prompt with a non-informative objective. As reported, the ablation does not establish that the prompt-based optimization component is essential.
[Sec. 2.3, Eq. (4)] The query-specific prompt P is an unconstrained additive perturbation to the query image, with no norm bound or regularization. The optimization runs up to 20 iterations with a threshold T=0.9 that may not be reachable. The paper does not report magnitudes of learned P or a sensitivity analysis for T and E. This makes the overfitting concern from the first major comment concrete, because P can grow large to force high A on the in-context labels without preserving meaningful image content.

minor comments (5)

[Sec. 3.3, Table 3] The row labels 'Update First' and 'Update Second' are not defined in the table or caption; the text mentions updating during generation of ŷp_t and ŷp_n, but the mapping to the rows should be stated explicitly.
[Sec. 3.1, Implementation Details] The preprocessing description is ambiguous: for UniverSeg images are resized to 128x128 and converted to gray-scale, while for SegGPT they are resized to 448x448 and converted to three-channel color. The paper should clarify whether the same normalization is applied for both backbones and why gray-scale conversion is used for UniverSeg.
[Fig. 1] The figure contains LaTeX artifacts such as 'f$' and 'M' in the diagram, and the legend for the red arrows (cycle data flow) should be explained more clearly so that the role swap between query and in-context pair is immediately visible.
[Sec. 2.2] The sentence 'Since the ground truth mask y_n is available' should specify that this is the ground-truth mask of the in-context pair, not the query; the current wording could be misinterpreted as using the query's ground truth.
[Throughout] Minor typos and formatting issues: 'NVIDIA3090' should be 'NVIDIA 3090', and the repeated use of 'implicit measure' could be replaced with a more precise term such as 'surrogate' or 'proxy' where appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CCV optimizes a query prompt against ground-truth in-context labels; the unvalidated 'implicit measure' claim is an assumption, not a reduction.

full rationale

The derivation chain is self-contained and does not reduce to its inputs. The verification accuracy A in Eq. (3) is computed from secondary predictions y^p_n = M(x_n, {(x_t, y^p_t)}) evaluated with ground-truth masks y_n of the in-context pairs, which are external labels independent of the query's own ground truth. Updating P to maximize A therefore constitutes test-time adaptation on available labeled context pairs, not fitting the query's label. The paper's statement that 'A serves as an implicit measure of the reliability of y^p_t' is an asserted heuristic, not a definitional equivalence: A is a function of y^p_t, the frozen model M, and the in-context labels, and the claimed correlation is empirically plausible but unverified. A missing correlation or failure-case study is a correctness and robustness concern, not circularity. The only first-party citations (Refs [14,29]) are background/survey/dataset references in the introduction and do not support the core mechanism. No equation is defined in terms of the quantity it predicts, and no fitted parameter is renamed as a prediction. Hence no specific reduction of the kind required for a circularity finding is present.

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

The method relies on a proxy accuracy computed with in-context ground-truth masks, which is a legitimate test-time signal, but the paper does not validate the proxy's correlation with query accuracy. Several hyperparameters are empirically chosen or unspecified, and the whole-model fine-tuning ablation contradicts the stated infeasibility of fine-tuning.

free parameters (5)

query-specific prompt P = same size as query image; zero-initialized, optimized per query
The main entity being optimized; its values are fit to the proxy A on in-context pairs (Section 2.3, Eq. (4)).
accuracy threshold T = 0.9
Empirically set to halt prompt optimization when proxy DSC exceeds this value (Section 3.1).
max iterations E = 20
Empirically set to cap the number of optimization steps (Section 3.1).
prompt learning rate = unspecified
Required to reproduce the optimization; not reported in Section 3.1.
prompt optimizer = unspecified
Not reported in Section 3.1; needed for exact reproduction.

assumptions (4)

domain assumption The secondary prediction accuracy A on in-context pairs is a valid proxy for the quality of the initial query prediction y_p_t.
Stated in Section 2.2: 'Thus, A serves as an implicit measure of the reliability of y_p_t.' No proof or empirical correlation analysis is provided.
domain assumption The ICL model can effectively consume a predicted mask (x_t, y_p_t) as an in-context pair without catastrophic degradation of the secondary prediction.
Required by Eq. (2) in Section 2.2; not validated in the paper.
domain assumption Fine-tuning the ICL model during inference is computationally infeasible and risks catastrophic forgetting.
Stated in Section 1 and the Abstract; however, the ablation in Section 3.3 actually fine-tunes the entire model for 20 iterations, contradicting the infeasibility claim.
domain assumption Optimizing the prompt P on the limited in-context pairs generalizes to the query image without overfitting.
Implicit in Section 2.3; no regularization or validation on the query is used during prompt optimization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cycle Context Verification for In-Context Medical Image Segmentation." pith.science (2026). https://pith.science/paper/ADNY6YJ7

@misc{pith2026250708357,
  author       = {Pith},
  title        = {Pith review of: Cycle Context Verification for In-Context Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADNY6YJ7}},
  note         = {Machine review of arXiv:2507.08357}
}

read the original abstract

In-context learning (ICL) is emerging as a promising technique for achieving universal medical image segmentation, where a variety of objects of interest across imaging modalities can be segmented using a single model. Nevertheless, its performance is highly sensitive to the alignment between the query image and in-context image-mask pairs. In a clinical scenario, the scarcity of annotated medical images makes it challenging to select optimal in-context pairs, and fine-tuning foundation ICL models on contextual data is infeasible due to computational costs and the risk of catastrophic forgetting. To address this challenge, we propose Cycle Context Verification (CCV), a novel framework that enhances ICL-based medical image segmentation by enabling self-verification of predictions and accordingly enhancing contextual alignment. Specifically, CCV employs a cyclic pipeline in which the model initially generates a segmentation mask for the query image. Subsequently, the roles of the query and an in-context pair are swapped, allowing the model to validate its prediction by predicting the mask of the original in-context image. The accuracy of this secondary prediction serves as an implicit measure of the initial query segmentation. A query-specific prompt is introduced to alter the query image and updated to improve the measure, thereby enhancing the alignment between the query and in-context pairs. We evaluated CCV on seven medical image segmentation datasets using two ICL foundation models, demonstrating its superiority over existing methods. Our results highlight CCV's ability to enhance ICL-based segmentation, making it a robust solution for universal medical image segmentation. The code will be available at https://github.com/ShishuaiHu/CCV.

Figures

Figures reproduced from arXiv: 2507.08357 by the authors.

**Figure 2.** Visualization of four learned query-specific prompt [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 22 canonical work pages

[1]

Ahmed, Z., Panhwar, S.Q., Baqai, A., Umrani, F.A., Ahmed, M., Khan, A.: Deep learningbasedautomateddetectionofintraretinalcystoidfluid.InternationalJour- nal of Imaging Systems and Technology32(3), 902–917 (2022)

work page 2022
[2]

Data in brief28, 104863 (2020)

Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

2020
[3]

Advances in Neural Information Processing Systems35, 25005–25017 (2022)

Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompt- ing via image inpainting. Advances in Neural Information Processing Systems35, 25005–25017 (2022)

work page 2022
[4]

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

Butoi, V.I., Ortiz, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Uni- verseg: Universal medical image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21438–21451 (2023)

work page 2023
[5]

Medical Image Analysis98, 103310 (2024)

Chen, C., Miao, J., Wu, D., Zhong, A., Yan, Z., Kim, S., Hu, J., Liu, Z., Sun, L., Li, X., et al.: Ma-sam: Modality-agnostic sam adaptation for 3d medical image segmentation. Medical Image Analysis98, 103310 (2024)

work page 2024
[6]

IEEE Transactions on Medical Imaging (2024)

Cheng, Z., Wang, S., Xin, T., Zhou, T., Zhang, H., Shao, L.: Few-shot medi- cal image segmentation via generating multiple representative descriptors. IEEE Transactions on Medical Imaging (2024)

2024
[7]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Chang, B., et al.: A survey on in-context learning. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 1107–1128 (2024)

work page 2024
[8]

IEEE Transactions on Medical Imaging (2024) 10 S

Gao, J., Lao, Q., Kang, Q., Liu, P., Du, C., Li, K., Zhang, L.: Boosting your context by dual similarity checkup for in-context learning medical image segmentation. IEEE Transactions on Medical Imaging (2024) 10 S. Hu et al

work page 2024

Show all 32 references

[9]

Huang, Y., Yang, X., Liu, L., Zhou, H., Chang, A., Zhou, X., Chen, R., Yu, J., Chen, J., Chen, C., et al.: Segment anything model for medical images? Medical Image Analysis 92, 103061 (2024)

2024
[10]

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

Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 4015–4026 (2023)

2023
[11]

arXiv preprint arXiv:2309.06256 (2023)

Lin, Y., Tan, L., Lin, H., Zheng, Z., Pi, R., Zhang, J., Diao, S., Wang, H., Zhao, H., Yao, Y., et al.: Speciality vs generality: An empirical study on catastrophic for- getting in fine-tuning foundation models. arXiv preprint arXiv:2309.06256 (2023)

2023 arXiv
[12]

Nature Communications15(1), 654 (2024)

Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications15(1), 654 (2024)

2024
[13]

Medical Image Analysis 89, 102918 (2023)

Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: an experimental study. Medical Image Analysis 89, 102918 (2023)

2023
[14]

Visual Intelligence 3(1), 1 (2025)

Mei, J., Zhou, T., Huang, K., Zhang, Y., Zhou, Y., Wu, Y., Fu, H.: A survey on deep learning for polyp segmentation: Techniques, challenges and future trends. Visual Intelligence 3(1), 1 (2025)

2025
[15]

arXiv preprint arXiv:2501.19393 (2025)

Muennighoff, N., Yang, Z., Shi, W., Li, X.L., Fei-Fei, L., Hajishirzi, H., Zettle- moyer, L., Liang, P., Candès, E., Hashimoto, T.: s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393 (2025)

2025 arXiv
[16]

In: Advances in Visual Computing: 16th International Symposium, ISVC 2021, virtual event, October 4-6, 2021, proceedings, part II

Ngoc Lan, P., An, N.S., Hang, D.V., Long, D.V., Trung, T.Q., Thuy, N.T., Sang, D.V.: Neounet: Towards accurate colon polyp segmentation and neoplasm detec- tion. In: Advances in Visual Computing: 16th International Symposium, ISVC 2021, virtual event, October 4-6, 2021, procee...

2021
[17]

Medical Image Analysis59, 101570 (2020)

Orlando, J.I., Fu, H., Breda, J.B., Van Keer, K., Bathula, D.R., Diaz-Pinto, A., Fang, R., Heng, P.A., Kim, J., Lee, J., et al.: Refuge challenge: A unified frame- work for evaluating automated methods for glaucoma assessment from fundus pho- tographs. Medical Image Analysis59...

2020
[18]

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
[19]

IEEE Transactions on Medical Imaging 34(11), 2366–2378 (2015)

Sirinukunwattana, K., Snead, D.R., Rajpoot, N.M.: A stochastic polygons model for glandular structures in colon histology images. IEEE Transactions on Medical Imaging 34(11), 2366–2378 (2015)

2015
[20]

arXiv preprint arXiv:2408.03314 (2024)

Snell, C., Lee, J., Xu, K., Kumar, A.: Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314 (2024)

2024 arXiv
[21]

arXiv preprint arXiv:2304.04748 (2023)

Sun, Y., Chen, Q., Wang, J., Wang, J., Li, Z.: Exploring effective factors for im- proving visual in-context learning. arXiv preprint arXiv:2304.04748 (2023)

2023
[22]

In: European Conference on Computer Vision

Suo, W., Lai, L., Sun, M., Zhang, H., Wang, P., Zhang, Y.: Rethinking and im- proving visual prompt selection for in-context learning segmentation. In: European Conference on Computer Vision. pp. 18–35. Springer (2024)

2024
[23]

Kaggle (2021), https://doi.org/10.34740/kaggle/dsv/3122958

Tahir, A.M., Chowdhury, M.E.H., Qiblawey, Y., Khandakar, A., Rahman, T., Ki- ranyaz, S., Khurshid, U., Ibtehaz, N., Mahmud, S., Ezeddin, M.: Covid-qu-ex dataset. Kaggle (2021), https://doi.org/10.34740/kaggle/dsv/3122958

2021
[24]

Computers in Biology and Medicine148, 105852 (2022) Cycle Context Verification for In-Context Medical Image Segmentation 11

Venugopal, V., Joseph, J., Das, M.V., Nath, M.K.: Dtp-net: A convolutional neural network model to predict threshold for localizing the lesions on dermatological macro-images. Computers in Biology and Medicine148, 105852 (2022) Cycle Context Verification for In-Context Medical...

2022
[25]

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

Wang, X., Wang, W., Cao, Y., Shen, C., Huang, T.: Images speak in images: A generalist painter for in-context visual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6830–6839 (2023)

2023
[26]

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

Wang, X., Zhang, X., Cao, Y., Wang, W., Shen, C., Huang, T.: Seggpt: Towards segmenting everything in context. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1130–1140 (2023)

2023
[27]

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

Wu, C., Restrepo, D., Shuai, Z., Liu, Z., Shen, L.: Efficient in-context medical seg- mentation with meta-driven visual prompt selection. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 255–265. Springer (2024)

2024
[28]

arXiv preprint arXiv:2405.15279 (2024)

Xu,C.,Liu,C.,Wang,Y.,Yao,Y.,Fu,Y.:Towardsglobaloptimalvisualin-context learning prompt selection. arXiv preprint arXiv:2405.15279 (2024)

2024 arXiv
[29]

Visual Intelligence2(1), 19 (2024)

Zeng, Q., Xie, Y., Lu, Z., Xia, Y.: A human-in-the-loop method for pulmonary nodule detection in ct scans. Visual Intelligence2(1), 19 (2024)

2024
[30]

Zhang, J., Wang, B., Li, L., Nakashima, Y., Nagahara, H.: Instruct me more! randompromptingforvisualin-contextlearning.In:ProceedingsoftheIEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2597–2606 (2024)

2024
[31]

Zhang, Y., Zhou, K., Liu, Z.: What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems 36, 17773–17794 (2023)

2023
[32]

Nature Methods pp

Zhao, T., Gu, Y., Yang, J., Usuyama, N., Lee, H.H., Kiblawi, S., Naumann, T., Gao, J., Crabtree, A., Abel, J., et al.: A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities. Nature Methods pp. 1–11 (2024)

2024

Pith tools

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

[1] [1]

Ahmed, Z., Panhwar, S.Q., Baqai, A., Umrani, F.A., Ahmed, M., Khan, A.: Deep learningbasedautomateddetectionofintraretinalcystoidfluid.InternationalJour- nal of Imaging Systems and Technology32(3), 902–917 (2022)

work page 2022

[2] [2]

Data in brief28, 104863 (2020)

Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

2020

[3] [3]

Advances in Neural Information Processing Systems35, 25005–25017 (2022)

Bar, A., Gandelsman, Y., Darrell, T., Globerson, A., Efros, A.: Visual prompt- ing via image inpainting. Advances in Neural Information Processing Systems35, 25005–25017 (2022)

work page 2022

[4] [4]

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

Butoi, V.I., Ortiz, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Uni- verseg: Universal medical image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21438–21451 (2023)

work page 2023

[5] [5]

Medical Image Analysis98, 103310 (2024)

Chen, C., Miao, J., Wu, D., Zhong, A., Yan, Z., Kim, S., Hu, J., Liu, Z., Sun, L., Li, X., et al.: Ma-sam: Modality-agnostic sam adaptation for 3d medical image segmentation. Medical Image Analysis98, 103310 (2024)

work page 2024

[6] [6]

IEEE Transactions on Medical Imaging (2024)

Cheng, Z., Wang, S., Xin, T., Zhou, T., Zhang, H., Shao, L.: Few-shot medi- cal image segmentation via generating multiple representative descriptors. IEEE Transactions on Medical Imaging (2024)

2024

[7] [7]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Chang, B., et al.: A survey on in-context learning. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 1107–1128 (2024)

work page 2024

[8] [8]

IEEE Transactions on Medical Imaging (2024) 10 S

Gao, J., Lao, Q., Kang, Q., Liu, P., Du, C., Li, K., Zhang, L.: Boosting your context by dual similarity checkup for in-context learning medical image segmentation. IEEE Transactions on Medical Imaging (2024) 10 S. Hu et al

work page 2024