Mask to Concept: Auto-Promptable SAM3 via Efficient Test-Time Concept Embedding Search for Few-Shot Annotation

Jia Chen; Qiang Hu; Qiang Li; Quan Zhou; Shaoqing Zhai; Zhiwei Wang

arxiv: 2606.26711 · v2 · pith:3QTSX4KMnew · submitted 2026-06-25 · 💻 cs.CV

Mask to Concept: Auto-Promptable SAM3 via Efficient Test-Time Concept Embedding Search for Few-Shot Annotation

Quan Zhou , Shaoqing Zhai , Qiang Hu , Jia Chen , Qiang Li , Zhiwei Wang This is my paper

Pith reviewed 2026-07-01 06:33 UTC · model grok-4.3

classification 💻 cs.CV

keywords few-shot segmentationmedical image annotationconcept embeddingSAM3test-time adaptationactive learninguncertainty estimation

0 comments

The pith

A learnable concept embedding optimized inside frozen SAM3 enables auto-prompted medical image segmentation from few labeled examples.

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

The paper shows that SAM3 can be made into an auto-promptable annotator for medical images by searching for a suitable concept embedding using only a few labeled masks. The embedding is initialized and refined by gradient updates that reduce the error between the model's segmentation and the given masks. This process happens without changing the model weights or adding extra networks. An uncertainty module identifies predictions that need human review, and the corrected masks improve the embedding in subsequent steps. If successful, this reduces the cost of creating large annotated medical datasets by limiting expert involvement to uncertain cases only.

Core claim

Mask to Concept (M2C) adapts SAM3 for few-shot medical annotation by initializing a learnable concept embedding that prompts the frozen model, then updates the embedding through backpropagation of the segmentation loss computed against the provided masks. A Hybrid Uncertainty Estimation module computes prediction entropy and inconsistency with box prompts to select samples for human correction, feeding the refined masks back to further optimize the embedding.

What carries the argument

The learnable concept embedding updated at test time by minimizing segmentation error gradients to generate prompts for the frozen SAM3 model.

If this is right

The method reaches state-of-the-art accuracy on medical few-shot segmentation tasks.
Annotation requires fewer human interventions due to the active learning loop with uncertainty estimates.
No external feature matchers or auxiliary networks are needed, keeping the approach lightweight.
The framework supports continuous improvement as more corrected masks become available.

Where Pith is reading between the lines

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

Similar test-time embedding search could adapt other frozen foundation models to new domains without retraining.
The approach might scale to video or 3D medical data if the embedding optimization generalizes across dimensions.
Reducing the number of initial labeled examples below current few-shot levels could be tested by strengthening the uncertainty feedback loop.

Load-bearing premise

The gradients from segmentation errors on a small set of medical images are sufficient to discover visual concepts that transfer across different images in the same domain.

What would settle it

If optimizing the concept embedding on five labeled CT scans fails to produce accurate segmentations on a held-out set of similar scans from the same hospital, the claim would be falsified.

Figures

Figures reproduced from arXiv: 2606.26711 by Jia Chen, Qiang Hu, Qiang Li, Quan Zhou, Shaoqing Zhai, Zhiwei Wang.

**Figure 1.** Figure 1: Overview of the proposed human-in-the-loop annotation framework. It consists of two core iterative modules: Mask to Concept (M2C) and Hybrid Uncertainty Estimation (HUE). During each iteration, given a few human-labeled samples from the previous iteration, M2C employs a continuous searching on the dataset-specific concept embedding to automatically segment the remaining unlabeled samples. Following this, … view at source ↗

**Figure 2.** Figure 2: Visualization of 1-shot results on Kvasir-SEG (top) and ISIC-2017 (bottom). 3.2 Implementation Details We implement our method based on the official pre-trained weight of SAM3 [18]. To ensure a fair comparison, all SAM-based baselines (i.e., ProtoSAM, Matcher, SPFS-SAM, and MAUP) utilize SAM-H [3]. During the mask to concept continual searching, we freeze the entire SAM3 backbone and exclusively optimize … view at source ↗

**Figure 3.** Figure 3: Annotation efficiency curves on Kvasir-SEG and ISIC-2017 datasets. out further fine-tuning. In our FSS evaluation protocol, the dataset is split into support and query sets using a 1 : 9 ratio. For K-shot segmentation, we test performance on the query set by forming episodes where one query sample is paired with K randomly selected support samples. To ensure statistical reliability, each query image under… view at source ↗

**Figure 4.** Figure 4: Ablation of HUE on Kvasir-SEG [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Transforming foundation segmentation models from human-prompted tools into auto-promptable annotators is critical for scalable medical data annotation. Current methods commonly depend on external feature matchers or auxiliary networks to automate geometric prompting, but introducing architectural overhead and limiting performance scalability. Although SAM3 natively supports concept segmentation via reusable text prompts, its direct use in medical imaging is hindered by a lack of fine-grained clinical knowledge and the ambiguity of human-written descriptions. In this work, we propose Mask to Concept (M2C), an efficient framework that adapts SAM3 for medical few-shot annotation without external modules, parameter retraining, or manual text engineering. Using only a few labeled images, M2C enables SAM3 to automatically search for transferable visual concepts entirely within its frozen architecture: it initializes a learnable concept embedding, uses it to prompt segmentation, and updates the embedding by gradients of minimizing the concept segmentation error. We further introduce a Hybrid Uncertainty Estimation (HUE) module that calculates the prediction entropy and maps concept predictions back to the box prompts, measuring concept-geometry prompting inconsistency. Highly uncertain samples are flagged actively for human correction, and the corrected masks are then fed back to M2C to continuously search for more precise concept embeddings, forming a self-enhancing annotation loop with minimal expert effort. Experiments on medical segmentation benchmarks show that our method achieves SOTA few-shot segmentation performance and outstanding annotation efficiency, offering a practical and efficient pathway toward scalable medical image labeling. Codes are at https://github.com/Huster-Hq/M2C.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

M2C tries to adapt frozen SAM3 for medical few-shot annotation by gradient-optimizing one concept embedding at test time plus an uncertainty-driven active loop, but the abstract supplies no numbers to support the SOTA claim.

read the letter

The paper's core move is to initialize a learnable embedding, feed it as a prompt into frozen SAM3, and update it by backpropagating segmentation loss on a handful of labeled medical images. A hybrid uncertainty module then flags high-entropy or geometrically inconsistent predictions for human correction, feeding the fixed masks back into the same embedding search. This keeps everything inside the original model without external matchers or retraining.

That approach is genuinely new in the way it closes the loop entirely within SAM3's frozen weights and uses the inconsistency signal to drive the active correction. It directly targets the practical pain point of medical annotation where you cannot afford to train new heads or collect large labeled sets.

The soft spot is the missing evidence. The abstract states SOTA few-shot results and strong annotation efficiency but gives no metrics, no baselines, no dataset splits, and no ablations. Without those, the central claim cannot be checked. The stress-test concern also lands: a single vector updated on a few images from a natural-image model can easily fit low-level textures or support-set artifacts rather than clinically meaningful structures, and the uncertainty check does not obviously prevent that.

The work is aimed at researchers building annotation pipelines around foundation models in medical imaging. If the full experiments include proper comparisons and controls, it would be worth a referee's time to verify whether the embedding actually generalizes. Based on the abstract alone the empirical part is too thin, but the method itself is coherent enough to send out for review rather than desk reject.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes Mask to Concept (M2C), a framework that adapts the frozen SAM3 model for few-shot medical image annotation by initializing a learnable concept embedding, prompting segmentation with it, and updating the embedding solely via gradients of segmentation error on a small number of labeled images. No external modules, retraining, or manual text prompts are used. A Hybrid Uncertainty Estimation (HUE) module computes prediction entropy and maps concept outputs back to box prompts to detect inconsistency; high-uncertainty cases are flagged for human correction whose masks are fed back into the embedding search, forming an iterative loop. The authors claim this yields SOTA few-shot segmentation performance and high annotation efficiency on medical benchmarks.

Significance. If the empirical results hold, the approach could enable practical, low-effort adaptation of foundation segmentation models to medical domains without architectural changes or large-scale retraining, supporting scalable annotation pipelines. The public code release at https://github.com/Huster-Hq/M2C is a clear strength that aids reproducibility.

major comments (2)

[Abstract] Abstract: the central claim of 'SOTA few-shot segmentation performance' is asserted without any metrics, baselines, dataset names, ablation tables, or quantitative results, so the soundness of the empirical contribution cannot be assessed from the provided text.
[Abstract] Method description (abstract): the single learnable concept embedding is optimized exclusively by back-propagating segmentation loss on the support-set masks; nothing in the described procedure (gradient search or HUE inconsistency check) supplies an independent mechanism to ensure the embedding encodes clinically meaningful, transferable structures rather than support-set-specific textures or artifacts, which is load-bearing for generalization claims given the natural-to-medical domain gap.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We address each point below and will revise the manuscript where appropriate to strengthen clarity.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 'SOTA few-shot segmentation performance' is asserted without any metrics, baselines, dataset names, ablation tables, or quantitative results, so the soundness of the empirical contribution cannot be assessed from the provided text.

Authors: We agree that the abstract would benefit from quantitative anchors to support the SOTA claim. The full manuscript reports detailed results including Dice/IoU metrics, multiple baselines, dataset names, and ablations. We will revise the abstract to incorporate key numerical highlights and dataset references while remaining within length limits. revision: yes
Referee: [Abstract] Method description (abstract): the single learnable concept embedding is optimized exclusively by back-propagating segmentation loss on the support-set masks; nothing in the described procedure (gradient search or HUE inconsistency check) supplies an independent mechanism to ensure the embedding encodes clinically meaningful, transferable structures rather than support-set-specific textures or artifacts, which is load-bearing for generalization claims given the natural-to-medical domain gap.

Authors: The optimization occurs directly on target-domain support masks, and the HUE module supplies an auxiliary consistency signal by mapping concept outputs back to box prompts and quantifying entropy-based disagreement; this loop is iterated with human-corrected masks. Transferability is validated empirically on held-out test images from the same medical benchmarks rather than through an external supervisory signal. We will expand the abstract's method summary to better emphasize the role of HUE-driven refinement in promoting robustness, while the full paper contains supporting ablations. revision: partial

Circularity Check

0 steps flagged

No circularity; method is explicit test-time optimization on external labels

full rationale

The paper presents a standard gradient-based optimization loop that fits a single concept embedding to minimize segmentation loss on a small set of provided labeled masks, then applies the resulting embedding to new images. This is a direct fitting procedure with no equations that reduce the claimed output to the input by construction. The HUE module derives uncertainty directly from model entropy and prompt inconsistency without reference to fitted targets. No self-citations serve as load-bearing uniqueness theorems, no ansatzes are smuggled, and no predictions are statistically forced from the same data used for fitting. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the learnable embedding is presented as part of the method rather than a new postulated physical entity.

pith-pipeline@v0.9.1-grok · 5825 in / 1118 out tokens · 36659 ms · 2026-07-01T06:33:08.055369+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

23 extracted references · 7 canonical work pages · 4 internal anchors

[1]

Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation.Medical image analysis, 63:101693, 2020

Nima Tajbakhsh, Laura Jeyaseelan, Qian Li, Jeffrey N Chiang, Zhihao Wu, and Xiaowei Ding. Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation.Medical image analysis, 63:101693, 2020

2020
[2]

Sali:Short-termalignmentandlong-terminteractionnetworkforcolonoscopyvideo polyp segmentation

Qiang Hu, Zhenyu Yi, Ying Zhou, Fang Peng, Mei Liu, Qiang Li, and Zhiwei Wang. Sali:Short-termalignmentandlong-terminteractionnetworkforcolonoscopyvideo polyp segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 531–541. Springer, 2024

2024
[3]

Segment anything

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

2023
[4]

SAM 2: Segment Anything in Images and Videos

Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[5]

Segment anything in medical images.Nature Communications, 15(1):654, 2024

Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature Communications, 15(1):654, 2024

2024
[6]

Universeg: Universal medical image segmentation

Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Universeg: Universal medical image segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 21438–21451, 2023

2023
[7]

Tyche: Stochastic in-context learning for medical image segmentation

Marianne Rakic, Hallee E Wong, Jose Javier Gonzalez Ortiz, Beth A Cimini, John V Guttag, and Adrian V Dalca. Tyche: Stochastic in-context learning for medical image segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11159–11173, 2024. 10 Q. Zhou et al

2024
[8]

Multiverseg: scalable interactive segmentation of biomedical imaging datasets with in-context guidance

Hallee E Wong, Jose Javier Gonzalez Ortiz, John Guttag, and Adrian V Dalca. Multiverseg: scalable interactive segmentation of biomedical imaging datasets with in-context guidance. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20966–20980, 2025

2025
[9]

Few-shot medical image segmentation via generating multiple representative descriptors.IEEE Transactions on Medical Imaging, 43(6):2202–2214, 2024

Ziming Cheng, Shidong Wang, Tong Xin, Tao Zhou, Haofeng Zhang, and Ling Shao. Few-shot medical image segmentation via generating multiple representative descriptors.IEEE Transactions on Medical Imaging, 43(6):2202–2214, 2024

2024
[10]

Samix: Reinforcing sam2 with semantic adapter and reference selecting policy for mix-supervised segmentation

QiangHu,JiajieWei,ZhenyuYi,ZhifenYan,YingjieGuo,HongkuanShi,Ge-Peng Ji, Qiang Li, and Zhiwei Wang. Samix: Reinforcing sam2 with semantic adapter and reference selecting policy for mix-supervised segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17948–17958, 2026

2026
[11]

arXiv preprint arXiv:2305.13310 (2023)

Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023

work page arXiv 2023
[12]

Protosam: One-shot medical image segmentation with foundational models.arXiv preprint arXiv:2407.07042, 2024

Lev Ayzenberg, Raja Giryes, and Hayit Greenspan. Protosam: One-shot medical image segmentation with foundational models.arXiv preprint arXiv:2407.07042, 2024

work page arXiv 2024
[13]

Maup: Training-free multi-center adaptive uncertainty-aware prompting for cross-domain few-shot medical image segmen- tation

Yazhou Zhu and Haofeng Zhang. Maup: Training-free multi-center adaptive uncertainty-aware prompting for cross-domain few-shot medical image segmen- tation. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, pages 326–336. Springer, 2025

2025
[14]

arXiv preprint arXiv:2305.03048 (2023)

Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot.arXiv preprint arXiv:2305.03048, 2023

work page arXiv 2023
[15]

Self-prompting large vision models for few-shot medical image segmentation

Qi Wu, Yuyao Zhang, and Marawan Elbatel. Self-prompting large vision models for few-shot medical image segmentation. InMICCAI workshop on domain adaptation and representation transfer, pages 156–167. Springer, 2023

2023
[16]

Synpo: Boosting training-free few-shot medical segmentation via high-quality negative prompts

Yufei Liu, Haoke Xiao, Jiaxing Chai, Yongcun Zhang, Rong Wang, Zijie Meng, and Zhiming Luo. Synpo: Boosting training-free few-shot medical segmentation via high-quality negative prompts. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 594–603. Springer, 2025

2025
[17]

First-framesupervisedvideopolyp segmentation via propagative and semantic dual-teacher network

QiangHu,MeiLiu,QiangLi,andZhiweiWang. First-framesupervisedvideopolyp segmentation via propagative and semantic dual-teacher network. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 1–5. IEEE, 2025

2025
[18]

SAM 3: Segment Anything with Concepts

Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025
[19]

Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

2022
[20]

An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to- image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022
[21]

Kvasir-seg: A segmented polyp dataset

Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. Auto-promptable SAM3 via Concept Embeddi...

2020
[22]

Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic)...

2017
[23]

Decoupled Weight Decay Regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017

[1] [1]

Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation.Medical image analysis, 63:101693, 2020

Nima Tajbakhsh, Laura Jeyaseelan, Qian Li, Jeffrey N Chiang, Zhihao Wu, and Xiaowei Ding. Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation.Medical image analysis, 63:101693, 2020

2020

[2] [2]

Sali:Short-termalignmentandlong-terminteractionnetworkforcolonoscopyvideo polyp segmentation

Qiang Hu, Zhenyu Yi, Ying Zhou, Fang Peng, Mei Liu, Qiang Li, and Zhiwei Wang. Sali:Short-termalignmentandlong-terminteractionnetworkforcolonoscopyvideo polyp segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 531–541. Springer, 2024

2024

[3] [3]

Segment anything

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

2023

[4] [4]

SAM 2: Segment Anything in Images and Videos

Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024

[5] [5]

Segment anything in medical images.Nature Communications, 15(1):654, 2024

Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature Communications, 15(1):654, 2024

2024

[6] [6]

Universeg: Universal medical image segmentation

Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R Sabuncu, John Guttag, and Adrian V Dalca. Universeg: Universal medical image segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 21438–21451, 2023

2023

[7] [7]

Tyche: Stochastic in-context learning for medical image segmentation

Marianne Rakic, Hallee E Wong, Jose Javier Gonzalez Ortiz, Beth A Cimini, John V Guttag, and Adrian V Dalca. Tyche: Stochastic in-context learning for medical image segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11159–11173, 2024. 10 Q. Zhou et al

2024

[8] [8]

Multiverseg: scalable interactive segmentation of biomedical imaging datasets with in-context guidance

Hallee E Wong, Jose Javier Gonzalez Ortiz, John Guttag, and Adrian V Dalca. Multiverseg: scalable interactive segmentation of biomedical imaging datasets with in-context guidance. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 20966–20980, 2025

2025

[9] [9]

Few-shot medical image segmentation via generating multiple representative descriptors.IEEE Transactions on Medical Imaging, 43(6):2202–2214, 2024

Ziming Cheng, Shidong Wang, Tong Xin, Tao Zhou, Haofeng Zhang, and Ling Shao. Few-shot medical image segmentation via generating multiple representative descriptors.IEEE Transactions on Medical Imaging, 43(6):2202–2214, 2024

2024

[10] [10]

Samix: Reinforcing sam2 with semantic adapter and reference selecting policy for mix-supervised segmentation

QiangHu,JiajieWei,ZhenyuYi,ZhifenYan,YingjieGuo,HongkuanShi,Ge-Peng Ji, Qiang Li, and Zhiwei Wang. Samix: Reinforcing sam2 with semantic adapter and reference selecting policy for mix-supervised segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17948–17958, 2026

2026

[11] [11]

arXiv preprint arXiv:2305.13310 (2023)

Yang Liu, Muzhi Zhu, Hengtao Li, Hao Chen, Xinlong Wang, and Chunhua Shen. Matcher: Segment anything with one shot using all-purpose feature matching. arXiv preprint arXiv:2305.13310, 2023

work page arXiv 2023

[12] [12]

Protosam: One-shot medical image segmentation with foundational models.arXiv preprint arXiv:2407.07042, 2024

Lev Ayzenberg, Raja Giryes, and Hayit Greenspan. Protosam: One-shot medical image segmentation with foundational models.arXiv preprint arXiv:2407.07042, 2024

work page arXiv 2024

[13] [13]

Maup: Training-free multi-center adaptive uncertainty-aware prompting for cross-domain few-shot medical image segmen- tation

Yazhou Zhu and Haofeng Zhang. Maup: Training-free multi-center adaptive uncertainty-aware prompting for cross-domain few-shot medical image segmen- tation. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, pages 326–336. Springer, 2025

2025

[14] [14]

arXiv preprint arXiv:2305.03048 (2023)

Renrui Zhang, Zhengkai Jiang, Ziyu Guo, Shilin Yan, Junting Pan, Xianzheng Ma, Hao Dong, Peng Gao, and Hongsheng Li. Personalize segment anything model with one shot.arXiv preprint arXiv:2305.03048, 2023

work page arXiv 2023

[15] [15]

Self-prompting large vision models for few-shot medical image segmentation

Qi Wu, Yuyao Zhang, and Marawan Elbatel. Self-prompting large vision models for few-shot medical image segmentation. InMICCAI workshop on domain adaptation and representation transfer, pages 156–167. Springer, 2023

2023

[16] [16]

Synpo: Boosting training-free few-shot medical segmentation via high-quality negative prompts

Yufei Liu, Haoke Xiao, Jiaxing Chai, Yongcun Zhang, Rong Wang, Zijie Meng, and Zhiming Luo. Synpo: Boosting training-free few-shot medical segmentation via high-quality negative prompts. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 594–603. Springer, 2025

2025

[17] [17]

First-framesupervisedvideopolyp segmentation via propagative and semantic dual-teacher network

QiangHu,MeiLiu,QiangLi,andZhiweiWang. First-framesupervisedvideopolyp segmentation via propagative and semantic dual-teacher network. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Pro- cessing (ICASSP), pages 1–5. IEEE, 2025

2025

[18] [18]

SAM 3: Segment Anything with Concepts

Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, An- drew Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

work page internal anchor Pith review Pith/arXiv arXiv 2025

[19] [19]

Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

2022

[20] [20]

An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion

Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to- image generation using textual inversion.arXiv preprint arXiv:2208.01618, 2022

work page internal anchor Pith review Pith/arXiv arXiv 2022

[21] [21]

Kvasir-seg: A segmented polyp dataset

Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In MultiMedia Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II 26, pages 451–462. Springer, 2020. Auto-promptable SAM3 via Concept Embeddi...

2020

[22] [22]

Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic)...

2017

[23] [23]

Decoupled Weight Decay Regularization

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017