Pith. sign in

REVIEW 4 major objections 8 minor 45 references

MedSeg-R: Medical Image Segmentation with Clinical Reasoning

T0 review · 4 major / 8 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 Clinical priors lift SAM medical segmentation to 86.7% Dice

desk verdict Interesting architecture with real Dice gains, but the test-time text-prior pipeline is under-specified enough that the gains may be coming from privileged text rather than the proposed modules. read the letter →

arxiv 2506.18669 v1 pith:2DSO2NP7 submitted 2025-06-23 cs.CV

Hao Shao , Qibin Hou This is my paper
classification cs.CV
keywords medicalimagesegmentationSegmentAnythingModelclinicalreasoningstructuredsemanticpriorssmalllesioncross-attributeattentiondynamicconvolutiondeformable
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 aims to show that medical image segmentation improves when a frozen Segment Anything Model backbone is guided by structured clinical priors — location, texture, and shape — extracted from diagnostic text. The proposed framework, MedSeg-R, first parses a medical report into three attribute embeddings, fuses them with a transformer, then uses them to modulate the backbone via spatial attention, dynamic convolution, and deformable sampling. This is designed to resolve inter-class ambiguity and compensate for severe class imbalance that hurts small lesions. If the account is right, lightweight guidance modules can be plugged into SAM-based systems to yield consistent Dice gains on overlapping and small structures.

What carries the argument

The load-bearing mechanism is the attribute-guided modulation inserted into a frozen SAM encoder. Three text attributes from CLIP are transformed into (1) a spatial attention map (location), (2) a 7x7 convolution kernel (texture), and (3) deformable-convolution kernels and offsets (shape). A final gating fusion combines the modulated feature map, the original SAM feature, and the pooled semantic vector Tattr, and feeds the result to SAM's mask decoder. The cognitive stage's cross-attribute transformer block lets the three priors interact before pooling, which the ablation shows is worth roughly one Dice point.

What would settle it

Run MedSeg-R on Synapse with the three attribute strings replaced by deliberately wrong descriptions (for example, 'the gallbladder is in the left lower quadrant, spiky, and hyperechoic'). If Dice drops by less than a point relative to correct attributes, the structured priors are not carrying the reported gain; if the drop is large, the method's dependence on an external, unspecified report pipeline is confirmed.

Watch

Extended reading notes

Core claim

MedSeg-R's central claim is that fine-grained, attribute-level semantic priors — instead of a single global sentence embedding — are what make clinical reasoning helpful for segmentation. Injecting these priors early into the SAM feature extractor, rather than into a separate decoder, produces a model that is both more accurate and cheaper than SEG-SAM. On the SA-Med2D benchmark the method reaches 74.27% average Dice under point prompts and 81.90% under box prompts, exceeding SEG-SAM by +1.99 and +1.17 points respectively; on the Synapse multi-organ CT set it reaches 86.71% mDice with HD95 of 6.13, beating SEG-SAM by +2.62 mDice, with the largest gains on the gallbladder (+5.03) and pancreas (+2.22).

Load-bearing premise

The whole gain depends on having an accurate structured description — position, texture, shape — of the target structure at test time, produced by an LLM and expert refinement; the paper does not specify which LLM, which prompts, or how the experts edited the reports for the SA-Med2D and Synapse test sets.

Editorial extensions

If this is right

  • On SA-Med2D, MedSeg-R beats SEG-SAM by +1.99 Dice under point prompts and +1.17 under box prompts, and leads in all eight tested modalities.
  • On Synapse, MedSeg-R reaches 86.71 mDice and 6.13 HD95, surpassing SEG-SAM by +2.62 mDice, with the largest gains on small, overlapping organs such as the gallbladder (+5.03) and pancreas (+2.22).
  • The method works prompt-free on Synapse, indicating reduced reliance on user-provided spatial cues.
  • The SAM backbone is frozen and only the guidance modules and mask decoder are tuned (68.49M parameters, 48 FPS), so the gains come without extra decoding branches.

Reading between the lines

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

  • An immediate reproducibility check stands out: because the LLM and expert-refinement pipeline is unspecified, a third party could re-run the Synapse evaluation with reports generated by a public LLM using the paper's own templates; whether the +2.62 gain survives that substitution is an open question the paper does not answer.
  • The ablations show location contributes most of the gain; a reasonable extension is a deployment variant that uses only location priors when texture and shape annotations are too costly, accepting a small Dice loss for much simpler data preparation.
  • Since the backbone stays frozen and only 68.49M parameters are tuned, the same attribute-modulation blocks could likely be attached to SAM 2 or other promptable backbones; the paper does not test that, but the efficiency numbers (48 FPS versus 13 FPS for MedSAM) suggest the overhead is small enough to try.
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

4 major / 8 minor

Summary. The paper proposes MedSeg-R, a two-stage framework for medical image segmentation built on a frozen SAM image encoder. In a 'cognitive' stage, structured semantic priors (location, texture, shape) are extracted from clinical reports via a CLIP text encoder and refined with cross-attribute attention. In a 'perceptual' stage, these priors modulate SAM's image features through spatial attention, a 7x7 texture convolution, and deformable convolution for shape, followed by a gating fusion of the original SAM features, the modulated features, and the semantic prior. The method is evaluated on SA-Med2D (eight modalities, point and box prompts) and Synapse (prompt-free multi-organ CT), with reported improvements over SEG-SAM and other baselines, particularly for small and overlapping structures. Ablations on KiTS19 and Kvasir study the contribution of each module and prior type.

Significance. If the results hold, the paper makes a useful contribution by showing that structured clinical priors can be injected into a frozen SAM backbone with relatively few trainable parameters (68.49M, 48 FPS), while reproducing several baselines under the same training protocol (Table 1 and Table 2). The per-modality breakdown and the overlap-robustness analysis in Fig. 3 are valuable. However, the central claim depends on the provenance of the text priors, which is not specified in the main text; if per-image expert-refined LLM reports are used at test time, the comparison with prompt-driven baselines is not apples-to-apples. The reported small-object point-prompt Dice exceeding box-prompt Dice also needs clarification. The methodological idea is sound and potentially of interest to the segmentation community, but the experimental protocol must be clarified and possibly revised before the specific quantitative claims can be accepted.

major comments (4)
  1. [Appendix A.3] The provenance of the structured priors is the load-bearing element of the evaluation and is unspecified. Fig. 5 and Appendix A.3 describe a pipeline where an LLM generates a medical report from each input image, the report is 'reviewed and refined by medical experts', and the revised text is parsed into location, texture, and shape attributes. The main text (Section 3.2, Eqs. 1-3) never states whether these attribute texts are per-image or fixed per-class templates, which LLM and prompts are used, whether the same expert-refined reports are available at training and test time, and how this was applied to each benchmark (SA-Med2D, Synapse, KiTS19, Kvasir). If per-image expert-refined reports were used at inference, then MedSeg-R is not an image-to-mask model; it consumes privileged textual information that the baselines do not, and the reported gains in Tables 1-2 and Figs. 3-4 could reflect the oracle-like text priors rather than the proposed attention, dynamic convolution, and deformable sampling modules. Please provide a complete, reproducible protocol for generating the priors, including the LLM, prompts, and whether expert refinement is part of the method or only a data-preparation step.
  2. [Section 4.2] The small-object results are internally inconsistent: for MedSeg-R, the average small-object Dice under point prompts is 59.10%, while under box prompts it is 51.33%, and a similar inversion holds for SEG-SAM (57.41% vs 49.57%). This contradicts the overall averages in Table 1, where box prompts outperform point prompts for every method, and it is also contrary to the expectation that more precise spatial prompts should help small objects. Please clarify how the 'small object' subset is defined in each prompt condition, whether the same images and objects are used in both evaluations, and whether the inversion is caused by different sampling of the test set, a different threshold, or a typographical error.
  3. [Section 4.4] All ablation studies are conducted on the KiTS19 and Kvasir datasets, which are not described in Section 4.1 and are never used in the main comparisons. The reader cannot assess how these results relate to the SA-Med2D and Synapse benchmarks, since the data split, prompt setting, and training protocol are not reported. Please add a description of these datasets and the exact experimental setup, or move the ablations to the main benchmark so that the contribution of each module is evaluated under the same conditions as the central claims.
  4. [Section 3.2] The method description does not formally state the input space of the cognitive stage. Eq. (1) assumes the attribute texts T_txt_alpha are given, but the paper does not specify whether these texts are derived per image at test time (and if so, with which model), or whether they are fixed per class. This ambiguity affects the mathematical soundness of the method as described: if the text varies per image, the method is conditional on an additional input modality that is not part of the problem formulation; if the text is fixed per class, then the claimed 'per-image' clinical reasoning in Fig. 5 is not actually used. Please clarify this point, as it determines whether the method can be used without a human-in-the-loop or an LLM at inference.
minor comments (8)
  1. [Abstract] The abstract contains a duplicated phrase: 'inspired by inspired by clinical reasoning'.
  2. [Section 4.1] There is a typo in 'All modelities are trained'; it should be 'All modalities are trained'.
  3. [Table 1 caption] The caption says 'Med2D-16M dataset', while the text and Appendix describe 'SA-Med2D-20M' and 'SA-Med2D-20M' with 4.6M images and ~20M masks; please align the naming and describe the relationship between '16M' and '20M' if they refer to the same data.
  4. [Section 4.2, Small object segmentation] The definition of small objects as occupying 'less than 5% of the image area' is not tied to the preprocessing step in Appendix A.2 that discards masks smaller than 0.153% of the image area. Please clarify whether the 5% threshold is applied to the object area relative to the image, and report the number of objects and modalities that fall into this category.
  5. [Fig. 2 and Fig. 5] The captions and the embedded text in the figures are dense and hard to read, especially the attribute texts in Fig. 2 and the LLM/refinement blocks in Fig. 5. Please enlarge fonts and use a clearer layout.
  6. [Section 2 / References] The in-text name 'Iris Gao et al. [2025]' does not match the reference list entry 'Yunhe Gao et al. [2025]' for 'Show and Segment'; please correct the inconsistency.
  7. [Section 5, Limitations] The limitations paragraph states that 'these priors are still learned implicitly from data rather than being explicitly encoded', which conflicts with the paper's description of the cognitive stage as explicitly parsing structured attributes from text. Please rephrase to clarify what is learned implicitly.
  8. [Appendix A.2] The threshold of 0.153% of the image area for discarding masks is introduced without justification; please cite the source of this value or provide a sensitivity analysis.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported Dice gains are empirical measurements from trained models, and no equation reduces to a fitted parameter or to a self-citation chain.

full rationale

MedSeg-R is an architecture-and-training paper: the claimed improvements on SA-Med2D and Synapse are measured model outputs, not quantities derived by construction from fitted constants. The structured priors (location, texture, shape) enter as text embeddings that modulate SAM features via attention, dynamic convolution, and deformable sampling (Sec. 3.2-3.3); their contribution is assessed by ablations (Tables 3-6), which is an empirical, falsifiable protocol rather than a definitional identity. The only potentially fragile point is Appendix A.3: the LLM-generated, expert-refined reports are produced from the input image itself, so the 'priors' are not independent external knowledge. But that is a self-conditioning design choice and an experimental-protocol concern about privileged test-time information, not circularity in the derivation sense: the paper never claims to derive the segmentation from first principles, and no equation reduces to its own input. Self-citations to prior works (Polyper, Camoformer) appear only as background and are not load-bearing. The stated limitation that priors are 'learned implicitly from data rather than being explicitly encoded' further confirms that the method is trained and evaluated empirically. Accordingly, no specific circular step can be quoted and exhibited, and the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard training assumptions and on the availability of accurate structured text priors, which is the main added dependency beyond the image itself.

free parameters (2)
  • Small-object threshold = 5% of image area
    Used to define small objects in evaluation; results for small-object Dice depend on this hand-chosen cutoff (Sec. 4.2).
  • Texture kernel size = 7x7
    Chosen for the dynamic convolution filter in the perceptual stage (Sec. 3.3); no sensitivity analysis is provided.
assumptions (3)
  • domain assumption LLM-generated reports correctly extract location, texture, and shape priors for the target structures.
    The cognitive stage relies on an LLM and expert refinement to generate structured reports (Appendix A.3 and Fig. 5); the paper does not specify the LLM, prompts, or acceptance criteria.
  • domain assumption The training distribution (25% subset of SA-Med2D-20M and full Synapse) is representative of the test distribution.
    Models are trained on a 25% subset; performance is evaluated on the full test set (Sec. 4.1).
  • standard math Standard neural network components (transformer, convolution, deformable convolution) behave as expected.
    The method uses standard building blocks; no proofs are needed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MedSeg-R: Medical Image Segmentation with Clinical Reasoning." pith.science (2026). https://pith.science/paper/2DSO2NP7

@misc{pith2026250618669,
  author       = {Pith},
  title        = {Pith review of: MedSeg-R: Medical Image Segmentation with Clinical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2DSO2NP7}},
  note         = {Machine review of arXiv:2506.18669}
}
read the original abstract

Medical image segmentation is challenging due to overlapping anatomies with ambiguous boundaries and a severe imbalance between the foreground and background classes, which particularly affects the delineation of small lesions. Existing methods, including encoder-decoder networks and prompt-driven variants of the Segment Anything Model (SAM), rely heavily on local cues or user prompts and lack integrated semantic priors, thus failing to generalize well to low-contrast or overlapping targets. To address these issues, we propose MedSeg-R, a lightweight, dual-stage framework inspired by inspired by clinical reasoning. Its cognitive stage interprets medical report into structured semantic priors (location, texture, shape), which are fused via transformer block. In the perceptual stage, these priors modulate the SAM backbone: spatial attention highlights likely lesion regions, dynamic convolution adapts feature filters to expected textures, and deformable sampling refines spatial support. By embedding this fine-grained guidance early, MedSeg-R disentangles inter-class confusion and amplifies minority-class cues, greatly improving sensitivity to small lesions. In challenging benchmarks, MedSeg-R produces large Dice improvements in overlapping and ambiguous structures, demonstrating plug-and-play compatibility with SAM-based systems.

Figures

Figures reproduced from arXiv: 2506.18669 by the authors.

Figure 1
Figure 1. Comparison of MedSeg-R with existing approaches. (a) CNN/Transformer-based models, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the MedSeg-R framework. The input image features from SAM’s encoder [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Combined analysis of anatomical overlap robustness (left) and small-object segmentation [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Left: Box plot of dice scores for three small volume organs, gallbladder, pancreas, and aorta, on the Synapse dataset. Each distribution compares MedSeg-R, SEG-SAM, and SAMed under identical settings, with MedSeg-R showing consistently higher medians and tighter spread…
Figure 5
Figure 5. Figure 5: An illustration of the process for extracting structured semantic attributes from medical [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Impact of data volume and prompt strategy on performance. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Segmentation results of different methods on the Synapse dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Segmentation results of different methods on the Endoscopy modality. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Segmentation results of different methods on the Derscopy modality. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

  2. [2]

    Transdeeplab: Convolution-free transformer-based deeplab v3+ for medical image segmentation

    Reza Azad, Moein Heidari, Moein Shariatnia, Ehsan Khodapanah Aghdam, Sanaz Karimijafarbigloo, Ehsan Adeli, and Dorit Merhof. Transdeeplab: Convolution-free transformer-based deeplab v3+ for medical image segmentation. In International Workshop on PRedictive Intelligence In MEdicine, pages 91--102. Springer, 2022

    work page 2022

  3. [3]

    Dae-former: Dual attention-guided efficient transformer for medical image segmentation

    Reza Azad, Ren \'e Arimond, Ehsan Khodapanah Aghdam, Amirhossein Kazerouni, and Dorit Merhof. Dae-former: Dual attention-guided efficient transformer for medical image segmentation. In International workshop on predictive intelligence in medicine, pages 83--95. Springer, 2023

    work page 2023

  4. [4]

    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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21438--21451, 2023

    work page 2023

  5. [5]

    Swin-unet: Unet-like pure transformer for medical image segmentation

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205--218. Springer, 2022

    work page 2022

  6. [6]

    Transunet: Transformers make strong encoders for medical image segmentation

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021

    arXiv 2021

  7. [7]

    gscorecam: What objects is clip looking at? In Proceedings of the Asian Conference on Computer Vision, pages 1959--1975, 2022

    Peijie Chen, Qi Li, Saad Biaz, Trung Bui, and Anh Nguyen. gscorecam: What objects is clip looking at? In Proceedings of the Asian Conference on Computer Vision, pages 1959--1975, 2022

    work page 1959

  8. [8]

    Sam fails to segment anything?--sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more

    Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?--sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2304.09148, 2023

    arXiv 2023

Show all 45 references
  1. [9]

    Sam-med2d

    Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184, 2023

    2023 arXiv

  2. [10]

    Unleashing the potential of sam for medical adaptation via hierarchical decoding

    Zhiheng Cheng, Qingyue Wei, Hongru Zhu, Yan Wang, Liangqiong Qu, Wei Shao, and Yuyin Zhou. Unleashing the potential of sam for medical adaptation via hierarchical decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3511--3522, 2024

    2024

  3. [11]

    3d u-net: learning dense volumetric segmentation from sparse annotation

    \"O zg \"u n C i c ek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger. 3d u-net: learning dense volumetric segmentation from sparse annotation. In Medical Image Computing and Computer-Assisted Intervention--MICCAI 2016: 19th International Conference, At...

    2016

  4. [12]

    Deformable convolutional networks

    Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764--773, 2017

    2017

  5. [13]

    Clustering propagation for universal medical image segmentation

    Yuhang Ding, Liulei Li, Wenguan Wang, and Yi Yang. Clustering propagation for universal medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3357--3369, 2024

    2024

  6. [14]

    Improving anatomical plausibility in medical image segmentation via hybrid graph neural networks: applications to chest x-ray analysis

    Nicol \'a s Gaggion, Lucas Mansilla, Candelaria Mosquera, Diego H Milone, and Enzo Ferrante. Improving anatomical plausibility in medical image segmentation via hybrid graph neural networks: applications to chest x-ray analysis. IEEE Transactions on Medical Imaging, 42 0 (2): ...

    2022

  7. [15]

    Show and segment: Universal medical image segmentation via in-context learning

    Yunhe Gao, Di Liu, Zhuowei Li, Yunsheng Li, Dongdong Chen, Mu Zhou, and Dimitris N Metaxas. Show and segment: Universal medical image segmentation via in-context learning. arXiv preprint arXiv:2503.19359, 2025

    2025 arXiv

  8. [16]

    Deformable gabor feature networks for biomedical image classification

    Xuan Gong, Xin Xia, Wentao Zhu, Baochang Zhang, David Doermann, and Li'an Zhuo. Deformable gabor feature networks for biomedical image classification. In Proceedings of the IEEE/CVF Winter Conference on applications of computer vision, pages 4004--4012, 2021

    2021

  9. [17]

    How to efficiently adapt large segmentation model (sam) to medical images

    Xinrong Hu, Xiaowei Xu, and Yiyu Shi. How to efficiently adapt large segmentation model (sam) to medical images. arXiv preprint arXiv:2306.13731, 2023

    2023 arXiv

  10. [18]

    3d u 2-net: A 3d universal u-net for multi-domain medical image segmentation

    Chao Huang, Hu Han, Qingsong Yao, Shankuan Zhu, and S Kevin Zhou. 3d u 2-net: A 3d universal u-net for multi-domain medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 291--299. Springer, 2019

    2019

  11. [19]

    Seg-sam: Semantic-guided sam for unified medical image segmentation

    Shuangping Huang, Hao Liang, Qingfeng Wang, Chulong Zhong, Zijian Zhou, and Miaojing Shi. Seg-sam: Semantic-guided sam for unified medical image segmentation. arXiv preprint arXiv:2412.12660, 2024

    2024 arXiv

  12. [20]

    Ldmres-net: A lightweight neural network for efficient medical image segmentation on iot and edge devices

    Shahzaib Iqbal, Tariq M Khan, Syed S Naqvi, Asim Naveed, Muhammad Usman, Haroon Ahmed Khan, and Imran Razzak. Ldmres-net: A lightweight neural network for efficient medical image segmentation on iot and edge devices. IEEE journal of biomedical and health informatics, 28 0 (7):...

    2023

  13. [21]

    nnu-net: Self-adapting framework for u-net-based medical image segmentation

    Fabian Isensee, Jens Petersen, Andre Klein, David Zimmerer, Paul F Jaeger, Simon Kohl, Jakob Wasserthal, Gregor Koehler, Tobias Norajitra, Sebastian Wirkert, et al. nnu-net: Self-adapting framework for u-net-based medical image segmentation. arXiv preprint arXiv:1809.10486, 2018

    2018 arXiv

  14. [22]

    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. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015--4026, 2023

    2023

  15. [23]

    Medclip-sam: Bridging text and image towards universal medical image segmentation

    Taha Koleilat, Hojat Asgariandehkordi, Hassan Rivaz, and Yiming Xiao. Medclip-sam: Bridging text and image towards universal medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 643--653. Springer, 2024

    2024

  16. [24]

    Tp-drseg: improving diabetic retinopathy lesion segmentation with explicit text-prompts assisted sam

    Wenxue Li, Xinyu Xiong, Peng Xia, Lie Ju, and Zongyuan Ge. Tp-drseg: improving diabetic retinopathy lesion segmentation with explicit text-prompts assisted sam. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 743--753. Springer, 2024

    2024

  17. [25]

    Towards universal text-driven ct image segmentation

    Yuheng Li, Yuxiang Lai, Maria Thor, Deborah Marshall, Zachary Buchwald, David S Yu, and Xiaofeng Yang. Towards universal text-driven ct image segmentation. arXiv preprint arXiv:2503.06030, 2025

    2025 arXiv

  18. [26]

    Fast and low-gpu-memory abdomen ct organ segmentation: the flare challenge

    Jun Ma, Yao Zhang, Song Gu, Xingle An, Zhihe Wang, Cheng Ge, Congcong Wang, Fan Zhang, Yu Wang, Yinan Xu, et al. Fast and low-gpu-memory abdomen ct organ segmentation: the flare challenge. Medical Image Analysis, 82: 0 102616, 2022

    2022

  19. [27]

    Segment anything in medical images

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15 0 (1): 0 654, 2024

    2024

  20. [28]

    Attention u-net: Learning where to look for the pancreas

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018

    2018 arXiv

  21. [29]

    Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation

    Md Mostafijur Rahman and Radu Marculescu. Multi-scale hierarchical vision transformer with cascaded attention decoding for medical image segmentation. In Medical Imaging with Deep Learning, pages 1526--1544. PMLR, 2024

    2024

  22. [30]

    Sam 2: Segment anything in images and videos

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

    2024 arXiv

  23. [31]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part...

    2015

  24. [32]

    An evaluation of atlas selection methods for atlas-based automatic segmentation in radiotherapy treatment planning

    Bas Schipaanboord, Djamal Boukerroui, Devis Peressutti, Johan van Soest, Tim Lustberg, Andre Dekker, Wouter van Elmpt, and Mark J Gooding. An evaluation of atlas selection methods for atlas-based automatic segmentation in radiotherapy treatment planning. IEEE transactions on m...

    2019

  25. [33]

    Polyper: Boundary sensitive polyp segmentation

    Hao Shao, Yang Zhang, and Qibin Hou. Polyper: Boundary sensitive polyp segmentation. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 4731--4739, 2024

    2024

  26. [34]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  27. [35]

    Sam-med3d: towards general-purpose segmentation models for volumetric medical images

    Haoyu Wang, Sizheng Guo, Jin Ye, Zhongying Deng, Junlong Cheng, Tianbin Li, Jianpin Chen, Yanzhou Su, Ziyan Huang, Yiqing Shen, et al. Sam-med3d: towards general-purpose segmentation models for volumetric medical images. arXiv preprint arXiv:2310.15161, 2023

    arXiv

  28. [36]

    Transbts: Multimodal brain tumor segmentation using transformer

    Wang Wenxuan, Chen Chen, Ding Meng, Yu Hong, Zha Sen, and Li Jiangyun. Transbts: Multimodal brain tumor segmentation using transformer. In International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, pages 109--119, 2021

    2021

  29. [37]

    Medical sam adapter: Adapting segment anything model for medical image segmentation

    Junde Wu, Ziyue Wang, Mingxuan Hong, Wei Ji, Huazhu Fu, Yanwu Xu, Min Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. Medical image analysis, 102: 0 103547, 2025

    2025

  30. [38]

    Cotr: Efficiently bridging cnn and transformer for 3d medical image segmentation

    Yutong Xie, Jianpeng Zhang, Chunhua Shen, and Yong Xia. Cotr: Efficiently bridging cnn and transformer for 3d medical image segmentation. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2021: 24th International Conference, Strasbourg, France, September 27...

    2021

  31. [39]

    Uniseg: A prompt-driven universal segmentation model as well as a strong representation learner

    Yiwen Ye, Yutong Xie, Jianpeng Zhang, Ziyang Chen, and Yong Xia. Uniseg: A prompt-driven universal segmentation model as well as a strong representation learner. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 508--518. Springer, 2023

    2023

  32. [40]

    Camoformer: Masked separable attention for camouflaged object detection

    Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  33. [41]

    Customized segment anything model for medical image segmentation

    Kaidong Zhang and Dong Liu. Customized segment anything model for medical image segmentation. arXiv preprint arXiv:2304.13785, 2023

    2023 arXiv

  34. [42]

    Organ-aware multi-scale medical image segmentation using text prompt engineering

    Wenjie Zhang, Ziyang Zhang, Mengnan He, and Jiancheng Ye. Organ-aware multi-scale medical image segmentation using text prompt engineering. arXiv preprint arXiv:2503.13806, 2025

    2025

  35. [43]

    Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers

    Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on ...

    2021

  36. [44]

    Pg-sam: Prior-guided sam with medical for multi-organ segmentation

    Yiheng Zhong, Zihong Luo, Chengzhi Liu, Feilong Tang, Zelin Peng, Ming Hu, Yingzhen Hu, Jionglong Su, Zongyuan Geand, and Imran Razzak. Pg-sam: Prior-guided sam with medical for multi-organ segmentation. arXiv preprint arXiv:2503.18227, 2025

    2025 arXiv

  37. [45]

    Unet++: A nested u-net architecture for medical image segmentation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep learning in medical image analysis and multimodal learning for clinical decision support: 4th international workshop, DLM...

    2018

Pith tools

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