Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read ProLearn claims that a prototype library distilled once from paired medical reports can replace textual input in language-guided segmentation for both training and inference, with accuracy that holds when only 1% of reports are available.

desk verdict Clever prototype-based idea for text-free medical segmentation, but the prototype initialization cannot work at the claimed low text rates, so the headline results are not yet credible. read the letter →

arxiv 2507.11055 v3 pith:VUIC7MNS submitted 2025-07-15 cs.CV

classification cs.CV
keywords medicalimagesegmentationlanguage-guidedprototypelearningtext-freeinferencesemanticapproximationtextualreliancevision-languagepretraining
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

Medical language-guided segmentation has so far required paired image-report data at both training and inference: unpaired images are wasted, and inference only works retrospectively when reports already exist. This paper proposes ProLearn, which builds a discrete prototype space once from the available paired reports and then lets any image query that space to receive an approximate textual guidance vector. Training can therefore include image-only data and inference needs no text at all. On three medical segmentation datasets the authors report that accuracy holds almost steady as report availability falls from 50% to 1%, that ProLearn beats language-guided baselines in those scarce-text settings, and that its image-only performance exceeds unimodal segmenters. The mechanism is also light: constant-time prototype lookup with roughly a million parameters, in contrast to LLM-based report generation.

What carries the argument

The central object is the Prototype-driven Semantic Approximation (PSA) module, a discrete query-and-response memory. Construction: segmentation-relevant tokens are selected from each report by a cross-attention threshold $\tau$; HDBSCAN clusters the resulting semantics into $N$ surrogate labels; K-means splits each label into $M$ sub-clusters; the sample nearest each centroid yields a linked pair $(q_{ij}, r_{ij})$ of image query prototype and report response prototype. At query time, cosine similarity ranks the query prototypes, the top-$k$ are chosen, and the response candidates are combined by softmax-weighted aggregation to form the semantic guidance $r^*$ for the U-Net decoder. The mechanism converts text from a per-case input into a fixed lookup table that can be queried by image features alone.

What would settle it

Take a held-out set whose reports contain a lesion descriptor absent from the initialization reports, and compare ProLearn's Dice on those images against an image-only U-Net; if the text-free guidance does not improve on those cases, the claim that the prototype space covers unseen semantics is refuted.

Watch

Extended reading notes

Core claim

ProLearn's central claim is that 'textual reliance' is not intrinsic to language-guided segmentation: the useful content of a clinical report can be captured in advance by a discrete prototype space, so that later images can retrieve the needed semantic guidance on their own. The Prototype-driven Semantic Approximation (PSA) module is initialized once from $K$ paired image-report samples: both modalities are encoded with BioMedCLIP, tokens whose cross-attention scores exceed $\tau$ are kept as segmentation-relevant semantics, HDBSCAN groups these into $N=6$ surrogate labels, and K-means splits each label into $M=64$ sub-clusters. The image embedding closest to each sub-cluster centroid becomes the query prototype $q_{ij}$ and its paired report embedding becomes the response prototype $r_{ij}$, forming the space $S = (S_Q, S_R)$. At run time, an image feature is compared by cosine similarity to every $q_{ij}$; the top-$k$ linked response prototypes are blended with softmax weights to produce the approximated textual feature $r^*$ that guides the decoder. In the reported experiments this removes text from inference entirely while keeping Dice essentially flat from 50% down to 1% report availability, and it lets the model outperform report-dependent methods plus image-only baselines in the same settings.

Load-bearing premise

The load-bearing premise is that every segmentation-relevant meaning a future report could contain is already represented in the small library built once from the paired reports available at initialization; a case whose needed meaning is missing from that library will receive a wrong or generic guidance.

Editorial extensions

If this is right

  • Inference no longer requires a clinical report: images query the prototype space directly, so segmentation can precede reporting and can run on devices where an LLM would not fit.
  • Image-only data become usable for training: unpaired images retrieve the same approximated semantics, so report scarcity stops wasting most of a dataset.
  • Performance degrades much more slowly as text disappears: across all three benchmarks the Dice drop from 50% to 1% text is below 0.04, while report-dependent baselines collapse far more.
  • The text-free mechanism is cheap: constant-time lookup with roughly a million parameters, reported as about 1000x fewer parameters and 100x faster inference than the LLM-based SGSeg approach.
  • With as little as 1% of reports available, image-only segmentation exceeds dedicated unimodal baselines such as U-Net, Attention U-Net, and VLP-adapted models on the same datasets.

Reading between the lines

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

  • The paper leaves implicit that the same prototype library could serve as a reusable prior: a new site or scanner would only need a handful of paired reports to recalibrate the query-response space, turning text dependence into a one-time setup cost.
  • A stress test beyond the paper's three datasets would measure whether the fixed library ($N=6$, $M=64$) covers report vocabulary not seen at initialization; if retrieval quality drops on unseen descriptors, the closed-vocabulary premise sets the method's ceiling, not its headline margin.
  • Because responses are softmax-weighted blends of retrieved prototypes, visually similar but clinically distinct lesions could receive indistinguishable guidance; attaching a retrieval-confidence score would make text-free predictions more auditable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes ProLearn, a prototype-driven framework for medical language-guided segmentation that aims to remove the need for paired image-text input at inference and reduce textual dependence during training. The core PSA module clusters paired image-text features into a discrete prototype space (via HDBSCAN and K-means) and answers image queries with an aggregated text feature. Experiments on QaTa-COV19, MosMedData+, and Kvasir-SEG compare ProLearn against language-guided baselines under reduced text availability and against unimodal methods in supposedly image-only settings, reporting gains in Dice and mIoU, as well as large reductions in parameter count and inference time relative to LLM-based SGSeg.

Significance. If the central claims are established, the work addresses a real limitation of language-guided segmentation: clinical segmentation typically precedes reporting, and paired text is scarce. The prototype idea is conceptually simple, parameter-efficient, and the authors provide code and an appendix with pseudocode. However, the current evidence does not yet support the headline claims. The specified prototype space cannot be populated by the described clustering procedure for the smallest text-availability settings, and the experimental protocol does not actually test the claimed use of image-only data. The reported gains are also not backed by error bars or significance tests. These are load-bearing issues rather than presentation concerns.

major comments (3)
  1. [Section 3.1 and Appendix A.2] The prototype space is constructed as N×M prototypes with N=6 and M=64 (Appendix A.2), requiring 384 distinct paired samples to populate via K-means. In the limited-text protocol of Section 4.3, 1% text yields approximately 57 paired samples for QaTa-COV19, 22 for MosMedData+, and 8 for Kvasir-SEG, all far below 384. With K-means, requesting M=64 sub-clusters from a surrogate cluster of size less than 64 is not well-defined, so the mechanism described cannot produce the prototype space used for the 1% (and some 5%/10%) results in Table 1 and Table 2. The authors must either specify how prototypes are formed when K < N×M (for example, by reducing M adaptively per surrogate label), or clarify that initialization uses the full training set before text subsampling; the latter would invalidate the comparison against baselines that receive only the limited paired data.
  2. [Section 4.3 and Table 2] The limited-text protocol discards unpaired images ('discarding the remaining unpaired images to simulate their inaccessibility'), so no image-only samples are ever used in training. This contradicts the contribution claim (Section 1) that ProLearn enables learning with both paired and image-only data. The image-only experiments in Table 2 are also internally inconsistent: the caption states that paired reports are excluded entirely, yet ProLearn rows use 1%, 5%, and 10% of paired reports. Please clarify the exact data available to each method and add experiments in which image-only samples are actually retained and used by ProLearn, since the current protocol does not test the proposed benefit.
  3. [Section 5.1 and Tables 1-2] All reported results are single runs without error bars or significance tests. Several margins versus the strongest baseline are very small (e.g., Table 1, QaTa-COV19 50% Dice: ProLearn 0.8667 vs SGSeg 0.8641), and the hyperparameter sensitivity analysis in Section 5.5 also reports no variance. Please report mean ± standard deviation over multiple seeds and provide significance tests for the key comparisons, particularly those involving small performance gaps.
minor comments (6)
  1. [Eq. (5)] The index j is used for both the K-means sub-cluster label and the paired-sample index (e.g., '⟨eI_j, eT_j⟩ ∈ Ci'); please rename one of them to avoid confusion.
  2. [Algorithm 2, lines 12-13] The summation and softmax notation is malformed: 'r∗ ← P (qij , rij )∈Q∗×R∗ wij rij' suggests a cross product of queries and responses and the softmax expression lacks a division sign. Please use the notation of Eq. (11).
  3. [Figure 5] The plot is labeled 'nIoU' but the metric defined and used elsewhere is mIoU; if this is intentional, define nIoU, otherwise correct the typo.
  4. [Reference [29]] Reference [29] is listed as a multilingual medical language model paper, but Section 5.3 cites it for Llama3; please verify that the reference is correct.
  5. [Appendix A.2] The token-selection threshold tau and the HDBSCAN hyperparameters (e.g., min_cluster_size) are not reported, so the initialization procedure is not reproducible from the manuscript alone.
  6. [Section 1 and Section 2.1] There are typos such as 'textural reliance' and 'release the textual reliance'; these should read 'textual reliance' and 'alleviate the textual reliance'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; ProLearn's prototype-driven pipeline is an empirical retrieval method with no step that reduces to its own inputs.

full rationale

Walking the derivation chain: PSA initialization (Sec. 3.1, Eqs. 1-7) builds a discrete prototype space from paired training data using BioMedCLIP features, attention-based token selection, HDBSCAN, and K-means; querying and responding (Sec. 3.2, Eqs. 8-11) retrieves and aggregates those prototypes from image features alone. No predicted quantity is used to define its own input: the response r* is a weighted sum over training text embeddings, which is a retrieval/approximation mechanism rather than a fit to the test target. The 'separately trained Language-guided U-Net' used for token relevance is independent of the PSA response parameters, and the claimed gains are empirical comparisons against external baselines. The only self-citation [42] (SGSeg) is contextual and used as a comparative baseline; it is not load-bearing for the PSA mechanism. The skeptic's 384-prototype cardinality concern is an internal consistency and feasibility issue about whether K-means can populate an N x M prototype space when only a small number of paired samples is available; it challenges whether the described algorithm can run as stated, not whether a claimed prediction is equivalent to an input by construction. Accordingly, no circular step is present and the circularity score is 0.

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

The method's central claim depends on several hand-set hyperparameters (N=6, M=64, k=10, tau) and on domain assumptions about the closed vocabulary of medical reports, BioMedCLIP's image-text alignment, and the sufficiency of top-k prototype aggregation. No new physical or conceptual entities are introduced; the prototype space is a learned representation.

free parameters (5)
  • Number of surrogate labels N = 6
    Set by hand in Appendix A.2; controls granularity of the prototype space; no sensitivity analysis for N is reported.
  • Number of sub-clusters M per surrogate label = 64
    Set to 64; sensitivity analysis in Fig 8b varies M but the selection is ad hoc.
  • Number of candidate responses k = 10
    Top-k prototype retrieval; sensitivity in Fig 8a but no principled selection.
  • Token selection threshold tau = not reported
    Eq. 2 retains tokens with cross-attention score above tau; the value is never given in the paper.
  • HDBSCAN clustering parameters = not reported
    Required to obtain N clusters from K semantic features; parameters such as min_cluster_size are not specified.
assumptions (4)
  • domain assumption The semantic space of segmentation-relevant medical report content is finite and can be represented by a small number of prototypes (N=6, M=64).
    Stated in Sec. 1: 'the semantic space of medical reports is inherently constrained... relatively closed vocabulary.' This underpins the feasibility of PSA and is not independently verified.
  • domain assumption BioMedCLIP image and text encoders produce aligned features in which cosine similarity reflects segmentation-relevant semantic similarity.
    Eqs. 1-8 build query and response spaces from BioMedCLIP embeddings and select prototypes by cosine similarity; no in-domain validation of this alignment is provided.
  • domain assumption Cross-attention scores from a separately trained language-guided U-Net identify which tokens are segmentation-relevant.
    Eq. 2 uses alpha_j > tau to extract T_selected; the separate U-Net is trained on the same K pairs, creating a potential leakage loop.
  • domain assumption Weighted aggregation of top-k prototype text features approximates the missing report semantics for an unseen image.
    Core query-and-respond mechanism, Sec. 3.2 and Eq. 11; no theoretical guarantee or ablation proving the approximation is faithful beyond segmentation metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation." pith.science (2026). https://pith.science/paper/VUIC7MNS

@misc{pith2026250711055,
  author       = {Pith},
  title        = {Pith review of: Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VUIC7MNS}},
  note         = {Machine review of arXiv:2507.11055}
}
read the original abstract

Medical language-guided segmentation, integrating textual clinical reports as auxiliary guidance to enhance image segmentation, has demonstrated significant improvements over unimodal approaches. However, its inherent reliance on paired image-text input, which we refer to as ``textual reliance", presents two fundamental limitations: 1) many medical segmentation datasets lack paired reports, leaving a substantial portion of image-only data underutilized for training; and 2) inference is limited to retrospective analysis of cases with paired reports, limiting its applicability in most clinical scenarios where segmentation typically precedes reporting. To address these limitations, we propose ProLearn, the first Prototype-driven Learning framework for language-guided segmentation that fundamentally alleviates textual reliance. At its core, we introduce a novel Prototype-driven Semantic Approximation (PSA) module to enable approximation of semantic guidance from textual input. PSA initializes a discrete and compact prototype space by distilling segmentation-relevant semantics from textual reports. Once initialized, it supports a query-and-respond mechanism which approximates semantic guidance for images without textual input, thereby alleviating textual reliance. Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms state-of-the-art language-guided methods when limited text is available.

Figures

Figures reproduced from arXiv: 2507.11055 by the authors.

Figure 1
Figure 1. Comparison of vision-language paradigms for medical [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Flow diagram comparison between SGSeg [42] and our proposed ProLearn. a) SGSeg: image → LLM → report → BERT → embedding. b) ProLearn: image → PSA → embedding. ever, these pretraining-based approaches fail to fully ex￾ploit the disease-specific information embedded in the tar￾get dataset’s reports, as the general knowledge learned dur￾ing pretraining often lacks the domain-specific details es￾sential for precise dise… view at source ↗
Figure 3
Figure 3. The overview of the ProLearn framework for alleviating textual reliance in medical language-guided segmentation. a) Language [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Attention-guided surrogate label extraction. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Performance degradation as the ratio of text availability [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Comparison between the proposed prototype-driven ap [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visualization comparison of language-guided segmentation methods under different text availabilities. The upper row shows [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Hyperparameter sensitivity analysis with varying candi [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Localization-Infused Vision-Language Semantic Fusion for Text-Guided Medical Image Segmentation

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A localization-infused vision-language fusion method converts textual location cues into multi-scale localization predictions and uses them to guide medical image segmentation, outperforming prior methods on three benchmarks.

Reference graph

Works this paper leans on

51 extracted references · 43 canonical work pages · cited by 1 Pith paper

  1. [1]

    Medical image segmentation review: The suc- cess of u-net

    Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The suc- cess of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10076–10095, 2024. 1

  2. [2]

    Sabuncu, John Guttag, and Adrian V

    Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R. Sabuncu, John Guttag, and Adrian V . Dalca. Uni- verseg: Universal medical image segmentation. In Proceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 21438–21451, 2023. 2

  3. [3]

    Density-based clustering based on hierarchical density esti- mates

    Ricardo JGB Campello, Davoud Moulavi, and J ¨org Sander. Density-based clustering based on hierarchical density esti- mates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 4

  4. [4]

    Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration

    Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE transac- tions on medical imaging, 33(2):577–590, 2013. 2

  5. [5]

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205–218. Springer, 2022. 2, 7

  6. [6]

    Vlp: A survey on vision-language pre-training

    Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Re- search, 20(1):38–56, 2023. 1

  7. [7]

    Deep learning with edge comput- ing: A review

    Jiasi Chen and Xukan Ran. Deep learning with edge comput- ing: A review. Proceedings of the IEEE, 107(8):1655–1674,

  8. [8]

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 7

Show all 51 references
  1. [9]

    Prior: Prototype representation joint learning from medical images and reports

    Pujin Cheng, Li Lin, Junyan Lyu, Yijin Huang, Wenhan Luo, and Xiaoying Tang. Prior: Prototype representation joint learning from medical images and reports. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 21361–21371, 2023. 4

  2. [10]

    Aysen Degerli, Serkan Kiranyaz, Muhammad E. H. Chowd- hury, and Moncef Gabbouj. Osegnet: Operational segmen- tation network for covid-19 detection using chest x-ray im- ages. In 2022 IEEE International Conference on Image Pro- cessing (ICIP), pages 2306–2310, 2022. 3, 5, 6

  3. [11]

    Measures of the amount of ecologic association between species

    Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297–302, 1945. 5

  4. [12]

    Value of multidetector computed tomogra- phy image segmentation for preoperative planning in general surgery

    Vincenzo Ferrari, Marina Carbone, Carla Cappelli, Luigi Boni, Franca Melfi, Mauro Ferrari, Franco Mosca, and An- drea Pietrabissa. Value of multidetector computed tomogra- phy image segmentation for preoperative planning in general surgery. Surgical endoscopy, 26:616–626, 2012. 2

  5. [13]

    Vision-and-language navigation: A survey of tasks, methods, and future directions

    Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers) , pages 7606–76...

  6. [14]

    Feasibility of real- time workflow segmentation for tracked needle interven- tions

    Matthew Stephen Holden, Tamas Ungi, Derek Sargent, Robert C McGraw, Elvis CS Chen, Sugantha Ganapathy, Terry M Peters, and Gabor Fichtinger. Feasibility of real- time workflow segmentation for tracked needle interven- tions. IEEE Transactions on Biomedical Engineering, 61(6): ...

  7. [15]

    Lga: A language guide adapter for advancing the sam model’s capabilities in medi- cal image segmentation

    Jihong Hu, Yinhao Li, Hao Sun, Yu Song, Chujie Zhang, Lanfen Lin, and Yen-Wei Chen. Lga: A language guide adapter for advancing the sam model’s capabilities in medi- cal image segmentation. page 610–620, Berlin, Heidelberg,

  8. [16]

    Lungren, and Serena Yeung

    Shih-Cheng Huang, Liyue Shen, Matthew P. Lungren, and Serena Yeung. Gloria: A multimodal global-local represen- tation learning framework for label-efficient medical image recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3922–3931, 2021. 6, 7

  9. [17]

    ´Etude comparative de la distribution florale dans une portion des alpes et des jura

    Paul Jaccard. ´Etude comparative de la distribution florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat, 37:547–579, 1901. 5

  10. [18]

    Kvasir-seg: A segmented polyp dataset

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, P ˚al Halvorsen, Thomas De Lange, Dag Johansen, and H˚avard D Johansen. Kvasir-seg: A segmented polyp dataset. In In- ternational conference on multimedia modeling, pages 451–

  11. [19]

    Registration and segmenta- tion for image-guided therapy

    Tina Kapur, Jan Egger, Jagadeesan Jayender, Matthew Toews, and William M Wells. Registration and segmenta- tion for image-guided therapy. Intraoperative Imaging and Image-Guided Therapy, pages 79–91, 2014. 2

  12. [20]

    Optical image-guided surgery—where do we stand? Molecular Imaging and Biology, 13:199–207, 2011

    Stijn Keereweer, Jeroen DF Kerrebijn, Pieter BAA Van Driel, Bangwen Xie, Eric L Kaijzel, Thomas JA Snoeks, Ivo Que, Merlijn Hutteman, Joost R Van Der V orst, J Sven D Mieog, et al. Optical image-guided surgery—where do we stand? Molecular Imaging and Biology, 13:199–207, 2011. 2

  13. [21]

    Lvit: language meets vision transformer in medical image seg- mentation

    Zihan Li, Yunxiang Li, Qingde Li, Puyang Wang, Dazhou Guo, Le Lu, Dakai Jin, You Zhang, and Qingqi Hong. Lvit: language meets vision transformer in medical image seg- mentation. IEEE transactions on medical imaging, 2023. 2, 5, 6

  14. [22]

    S. Lloyd. Least squares quantization in pcm. IEEE Transac- tions on Information Theory, 28(2):129–137, 1982. 4

  15. [23]

    Image segmenta- tion using text and image prompts

    Timo L ¨uddecke and Alexander Ecker. Image segmenta- tion using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086–7096, 2022. 1, 2

  16. [24]

    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(1):654, 2024. 1, 3, 7

  17. [25]

    Atten- tion res-unet with guided decoder for semantic segmentation of brain tumors

    Dhiraj Maji, Prarthana Sigedar, and Munendra Singh. Atten- tion res-unet with guided decoder for semantic segmentation of brain tumors. Biomedical Signal Processing and Control, 71:103077, 2022. 2

  18. [26]

    Breast cancer segmentation methods: current status and 9 future potentials

    Epimack Michael, He Ma, Hong Li, Frank Kulwa, and Jing Li. Breast cancer segmentation methods: current status and 9 future potentials. BioMed research international, 2021(1): 9962109, 2021. 2

  19. [27]

    Mosmeddata: Chest ct scans with covid-19 related findings

    Sergey Morozov, Anna Andreychenko, Nikolay Pavlov, Anton Vladzymyrskyy, Natalya Ledikhova, Victor Gom- bolevskiy, Ivan Blokhin, Pavel Gelezhe, Anna Gonchar, Va- leria Chernina, et al. Mosmeddata: Chest ct scans with covid-19 related findings. 2020. 3, 5, 6

  20. [28]

    Oktay, J

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y . Hammerla, B. Kainz, and et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 7

  21. [29]

    Towards building multilingual language model for medicine

    Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards building multilingual language model for medicine. Nature Communications, 15(1):8384, 2024. 6

  22. [30]

    Language models are unsuper- vised multitask learners

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsuper- vised multitask learners. 2019. 6

  23. [31]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on machine learning, ...

  24. [32]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolu- tional networks for biomedical image segmentation. InMed- ical Image Computing and Computer-Assisted Intervention– MICCAI 2015: 18th International Conference, Munich, Ger- many, October 5-9, 2015, Proceedings, Part III 1...

  25. [33]

    Siddique, S

    N. Siddique, S. Paheding, C. P. Elkin, and V . Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. IEEE Access, 9:82031– 82057, 2021. 1

  26. [34]

    Prototypical networks for few-shot learning

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural informa- tion processing systems, 30, 2017. 4

  27. [35]

    Large language models in medicine

    Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabi- lan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930–1940, 2023. 3

  28. [36]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information pro- cessing systems, 30, 2017. 4

  29. [37]

    Cris: Clip- driven referring image segmentation

    Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clip- driven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686–11695, 2022. 1, 2

  30. [38]

    Provision of automated step-by-step procedural guidance in virtual reality surgery simulation

    Sudanthi Wijewickrema, Yun Zhou, James Bailey, Gregor Kennedy, and Stephen O’Leary. Provision of automated step-by-step procedural guidance in virtual reality surgery simulation. In Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology, pages 69–72, 2016. 2

  31. [39]

    Bridging vision and language en- coders: Parameter-efficient tuning for referring image seg- mentation

    Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xi- ang Wan, and Guanbin Li. Bridging vision and language en- coders: Parameter-efficient tuning for referring image seg- mentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17503–17512,...

  32. [40]

    Robust classification with convolutional proto- type learning

    Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng- Lin Liu. Robust classification with convolutional proto- type learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474–3482,

  33. [41]

    Covid-19: im- munopathogenesis and immunotherapeutics

    Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: im- munopathogenesis and immunotherapeutics. Signal trans- duction and targeted therapy, 5(1):128, 2020. 5

  34. [42]

    Enabling text-free inference in language- guided segmentation of chest x-rays via self-guidance

    Shuchang Ye, Mingyuan Meng, Mingjian Li, Dagan Feng, and Jinman Kim. Enabling text-free inference in language- guided segmentation of chest x-rays via self-guidance. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 242–252. Springer,

  35. [43]

    [retracted] u-net-based medical image segmentation

    Xiao-Xia Yin, Le Sun, Yuhan Fu, Ruiliang Lu, and Yanchun Zhang. [retracted] u-net-based medical image segmentation. Journal of healthcare engineering, 2022(1):4189781, 2022. 1

  36. [44]

    Vision-language models for vision tasks: A survey

    Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence,

  37. [45]

    Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon

    Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Andrea Tupini, Yu Wang, Matt Mazzola, Swadheen Shukla, Lars Liden, Jianfeng Gao, Angela Crabtree, Brian Piening, Carlo Bifulco, Matthew P....

  38. [46]

    Madapter: A better interaction between image and language for medical image segmentation

    Xu Zhang, Bo Ni, Yang Yang, and Lefei Zhang. Madapter: A better interaction between image and language for medical image segmentation. In International Conference on Medi- cal Image Computing and Computer-Assisted Intervention , pages 425–434. Springer, 2024. 2, 3, 5

  39. [47]

    A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities

    Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Hin Lee, Ho Sid Kiblawi, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, Brian Piening, Carlo Bifulco, Mu Wei, Hoifung Poon, and Sheng Wang. A foundation model for joint segmentation, detection...

  40. [48]

    Ariadne’s thread: Using text prompts to improve segmentation of infected areas from chest x-ray images

    Yi Zhong, Mengqiu Xu, Kongming Liang, Kaixin Chen, and Ming Wu. Ariadne’s thread: Using text prompts to improve segmentation of infected areas from chest x-ray images. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 724–733. Springer,

  41. [49]

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical im- 10 age segmentation. In Deep Learning in Medical Image Anal- ysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and...

  42. [462]

    Springer, 2019. 5, 6

  43. [2024]

    Springer-Verlag. 2, 3, 5

Pith tools

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