Pith. sign in

REVIEW 4 major objections 5 minor 42 references

The same language-guidance adapter can ride on many vision and text backbones for medical lesion segmentation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 02:42 UTC pith:4D6F2VHT

load-bearing objection Solid engineering paper: reusable shape-preserving adapter + hierarchical losses that mostly transfer across backbones and beat strong VL baselines with modest cost; transferability is real but not absolute. the 4 major comments →

arxiv 2607.09481 v1 pith:4D6F2VHT submitted 2026-07-10 cs.CV cs.AI

Decoupling Language Guidance from Backbones for Text-Guided Medical Segmentation

Yungeng Liu , Xuanzi Fang , Haijin Zeng , Qi Dai , Yongyong Chen This is my paper
classification cs.CV cs.AI
keywords Medical Image SegmentationBackbone TransferabilityVision-Language ModelsHierarchical FrameworkCross-Modal AlignmentText-Guided SegmentationAdapter
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Most text-guided medical segmentation systems weld the language encoder, visual backbone, fusion block, and decoder into one architecture. Swap either encoder and the fusion path usually has to be redesigned. This paper argues that language guidance can instead sit on a stable feature interface: multi-scale visual maps plus a text embedding go in, fused maps of identical shape come out, and the existing decoder stays untouched. The reusable piece is a scale-adaptive gated adapter (SAGSG) that injects text with resolution-specific gates and channel recalibration, trained under a hierarchical coarse-to-fine loss that first aligns image and report globally, then localizes lesions at multiple scales, then refines boundaries. The same adapter and supervision recipe remains competitive across convolutional and transformer visual encoders and across several language models, and improves strong text-guided baselines on four public lesion datasets with only modest extra parameters and FLOPs.

Core claim

Language guidance for medical segmentation need not be co-designed with a particular backbone pair. A shape-preserving hierarchical adapter (BTHA) that accepts multi-scale visual features and a text representation, injects semantics via scale-adaptive gated fusion, and is trained with global image-text alignment plus multi-scale auxiliary and boundary-aware losses, transfers across heterogeneous vision and language encoders while improving strong baselines.

What carries the argument

BTHA’s shape-preserving SAGSG adapter: at each scale it runs masked text-to-vision cross-attention, dual residual gates (initialized near zero) for language injection and feed-forward refinement, and SE channel recalibration, then returns a tensor with the original spatial size and channel count so any decoder contract is preserved.

Load-bearing premise

Different vision and text backbones still produce multi-scale feature maps and text vectors that are similar enough in resolution, channel layout, and semantics for one fixed adapter topology and one shared hierarchical loss schedule to work without redesign.

What would settle it

Hold the SAGSG topology, hierarchical losses, and decoder interface fixed, swap in a vision or text encoder whose feature hierarchy or embedding distribution differs sharply from the tested set, and check whether Dice/mIoU collapses relative to a backbone-specific redesign on the same four datasets.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A laboratory can keep one language-guidance module while swapping visual encoders (CNN or transformer) or radiology text models without rewriting fusion or decoder code.
  • Training can be decomposed into global image-report alignment, coarse multi-scale localization, and boundary refinement, giving intermediate features explicit roles instead of a single final mask loss.
  • Text injection strength can be learned per scale and started near identity, reducing the risk that poorly aligned language features corrupt pre-trained visual anatomy early in training.
  • Reported gains hold with only a few extra million parameters and a small FLOP increase over the strongest comparable text-guided baselines.

Where Pith is reading between the lines

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

  • The same interface idea could let text-guided adapters ride on future foundation visual or language models without a full architectural rewrite each time a better encoder appears.
  • If the hierarchical losses are what stabilizes transfer, similar coarse-to-fine auxiliary heads may help other cross-modal medical tasks (detection, report grounding) when backbones change.
  • The residual-gate initialization near zero suggests a practical recipe for safe language injection into any frozen or lightly tuned medical visual backbone.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes BTHA, a backbone-transferable hierarchical adapter for text-guided medical image segmentation. The central claim is that language guidance can be decoupled from specific vision/text encoder pairs via a minimal feature-level interface: multi-scale visual features and a text representation are fused by a shape-preserving Scale-Adaptive Gated Semantic Guidance (SAGSG) adapter (masked cross-attention, dual zero-initialized tanh residual gates, SE recalibration), while a Hierarchical Coarse-to-Fine Supervision Strategy (image-text contrastive alignment, multi-scale auxiliary localization, boundary-aware hybrid final loss) regularizes learning without changing decoder tensor contracts. Transferability is evaluated by swapping text encoders (Table I) and vision encoders (Table II) on QaTa-COV19 with fixed adapter/supervision topology; SOTA comparisons on MosMedData+, QaTa-COV19, SIIM-ACR, and Kvasir-SEG (Table IV) and component ablations (Tables V–VI) report consistent gains over strong text-guided baselines with modest parameter/FLOP overhead.

Significance. If the transferability claim holds, the work is a useful systems contribution: modular language-guidance adapters would reduce redesign cost when swapping convolutional vs. transformer visual backbones or radiology-specific vs. general language encoders, which is a practical pain point in medical VL segmentation. Strengths include controlled backbone-swap tables, complementary ablations (hierarchical supervision alone helps; SAGSG alone hurts; both together improve), and multi-dataset comparison with reported params/FLOPs. The contribution is empirical and architectural rather than theoretical; significance depends on how convincingly the same fixed-topology interface generalizes beyond the default ConvNeXt+CXR-BERT pair and beyond a single transferability dataset.

major comments (4)
  1. [§III.B, Tables I–II vs Table IV] The paper’s primary claim is backbone transferability of the same SAGSG topology and hierarchical supervision (§II.A, Fig. 3). Tables I–II test this only on QaTa-COV19. Table IV’s four-dataset SOTA comparison uses a single default backbone pair. Without at least one additional dataset under the same swap protocol (or a clear limitation statement with multi-dataset partial swaps), the central transferability claim is only partially supported and risks being overstated relative to the evidence.
  2. [Tables I–II; §II.C Eqs. (4)–(5)] BTHA is second-best precisely on the clearest distribution mismatches: BioClinicalBERT (Table I, Dice 88.96 vs FMISeg 90.60) and ResNet50 (Table II, Dice 88.50 vs FMISeg 90.58). The manuscript notes residual competitiveness but provides no analysis of feature hierarchy/channel compatibility, gate activations (w^s_g, w^s_f), or attention-mask behavior under these backbones. For a transferability paper, these failure modes are load-bearing: either show that projections/gates absorb the mismatch in a controlled way, or qualify when the fixed {1/8,1/16,1/32} shape-preserving contract is insufficient.
  3. [Table V; §II.B Eq. (3)] Table V shows SAGSG alone drops Dice from 90.89% to 88.12%, recovering only when hierarchical supervision is added (91.88%). This is an important finding, but it means the reusable “adapter” is not independently transferable: success depends on co-transferring the three-level loss design (Eq. 3) and scale-dependent λ weights. The paper should treat supervision co-transfer as part of the interface contract and discuss implications for drop-in reuse on new backbones/datasets, rather than framing SAGSG as a standalone shape-preserving module.
  4. [§III.A, Table III] Text supervision construction is heterogeneous and under-specified for reproducibility of the four-dataset claim: SIIM-ACR uses manual lesion-containing annotations; Kvasir-SEG uses attribute-based generated prompts following TGA-Net (§III.A, Table III). Please specify annotation protocol, inter-annotator consistency (if any), prompt templates, and whether text quality correlates with gains on SIIM-ACR/Kvasir-SEG versus the radiology report-style sets. Without this, part of the multi-dataset improvement may be confounded by text quality rather than BTHA.
minor comments (5)
  1. [§II.B] Hyperparameters α_s, γ, and scale-dependent λ_d/λ_e are described qualitatively (“increase Dice weight for deep features… boundary loss for shallow features”) but not listed numerically. A short table or appendix would aid reproduction.
  2. [Fig. 4] Fig. 4 qualitative comparison is useful; adding failure cases (e.g., ResNet50 or BioClinicalBERT settings) would better support the transferability narrative.
  3. [§II.A; Table IV] Clarify whether intermediate auxiliary heads and ITC projections are removed at inference for all backbone settings, and whether FLOPs in Table IV exclude them consistently.
  4. [§II.B; Fig. 4] Minor wording: “Lov ´asz-hinge” spacing; ensure consistent naming of LanGuideMedSeg vs LanGuide in Fig. 4.
  5. [§I] Related work could briefly position against other adapter-style medical VL methods beyond those already compared, to sharpen novelty of the shape-preserving interface claim.

Circularity Check

0 steps flagged

No significant circularity: empirical adapter/supervision design evaluated on external public benchmarks, not a derivation that reduces to its inputs.

full rationale

BTHA is a methods paper whose central claims are experimental (same shape-preserving SAGSG topology + hierarchical losses remain usable across backbone pairs; mean Dice 81.97% on four public datasets with modest overhead). The training objectives (ITC contrastive loss Eq. 1 following the standard CLIP-style formulation [31]; hybrid Dice/Focal/Edge/Lovasz main loss Eq. 2; total loss Eq. 3) and the adapter (masked cross-attention, dual tanh gates zero-initialized, SE recalibration) are standard external constructions, not algebraic restatements of the evaluation metrics. There is no fitted parameter that is then reported as a prediction of a closely related quantity, no uniqueness theorem imported from the authors, and no load-bearing self-citation chain. Ablations and transferability tables (I–II, V–VI) compare against held-out splits and external baselines; success or second-place results under distribution mismatch (ResNet50, BioClinicalBERT) are reported rather than forced by construction. Minor design choices (logit upsampling, scale-specific gate weights) do not create circular proof structure. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

Load-bearing content is architectural and empirical, not theorem-driven. Free parameters are the usual ML loss/optimizer weights and scale-specific gates. Axioms are standard CV/ML and medical-imaging practice assumptions. Invented entities are the named modules that package known operators into a transferable interface; they have no independent physical existence outside the proposed network.

free parameters (4)
  • Hierarchical loss weights α_s (s∈{8,16,32}) and γ for ITC
    Control contribution of auxiliary scales and global contrastive term in L_total; chosen as hyperparameters, not derived.
  • Hybrid main/aux loss coefficients λ_d, λ_f, λ_e, λ_l (with scale-dependent emphasis)
    Dice/Focal/Edge/Lovasz mix and deeper-vs-shallower reweighting are hand-set design choices that affect the reported Dice/mIoU.
  • Scale-specific gate parameters w^s_g, w^s_f (tanh-gated residuals, zero-init)
    Learned per scale; initialization and capacity are design knobs that determine how aggressively text is injected.
  • AdamW learning rates 3e-4 (new heads/adapters) and 3e-5 (pretrained backbones) plus LambdaLR warmup
    Training schedule hyperparameters that influence final numbers across backbone tables.
axioms (5)
  • domain assumption Heterogeneous vision encoders expose multi-scale feature maps at roughly 1/8, 1/16, 1/32 that can share one shape-preserving fusion topology.
    Core interface assumption in §II.A; transferability claim collapses if hierarchies are incompatible.
  • domain assumption Image-text contrastive alignment on projected global features is a valid semantic anchor for dense medical segmentation.
    Uses standard ITC formulation (Eq. 1) following CLIP-style practice; assumed useful for lesion localization.
  • ad hoc to paper Upsampling intermediate logits to full resolution (instead of downsampling GT) preserves small-lesion supervision without dataset-specific mask preprocessing.
    Explicit design choice in §II.B that shapes the hierarchical supervision signal.
  • ad hoc to paper Zero-initialized tanh gates make text injection conservative enough not to destroy pretrained visual anatomy before alignment is learned.
    Stated motivation for dual-gated residual design (Eqs. 4–5); ablation shows SAGSG alone underperforms without hierarchical losses.
  • domain assumption Public dataset splits and evaluation protocols (Dice/mIoU) used by prior text-guided methods are adequate for fair comparison.
    Standard empirical CV practice; SIIM manual labels and Kvasir generated text are additional dataset-construction assumptions.
invented entities (3)
  • BTHA (backbone-transferable hierarchical adapter framework) no independent evidence
    purpose: Package a stable feature interface, hierarchical supervision, and SAGSG so language guidance can be reused across encoder pairs.
    Named system-level contribution; existence is definitional to the paper's method.
  • SAGSG (Scale-Adaptive Gated Semantic Guidance) adapter no independent evidence
    purpose: Shape-preserving multi-scale text injection via masked cross-attention, dual residual gates, and SE recalibration.
    New module composition; operators are standard but the packaged adapter is paper-specific.
  • Hierarchical Coarse-to-Fine Supervision Strategy no independent evidence
    purpose: Decompose training into ITC, multi-scale aux localization, and boundary-aware final refinement attachable without changing decoder topology.
    Training-side invention that the paper claims is required for the adapter to work well.

pith-pipeline@v1.1.0-grok45 · 17658 in / 3603 out tokens · 44035 ms · 2026-07-13T02:42:32.050634+00:00 · methodology

0 comments
read the original abstract

Text-guided medical image segmentation leverages clinical semantics to improve lesion delineation, yet many existing models bind cross-modal fusion, supervision, and decoder design into a task-specific architecture. Such tight coupling makes it difficult to reuse language guidance modules across heterogeneous vision and text backbones, and often requires redesigning the network when the encoder pair changes. This paper presents BTHA, a backbone-transferable hierarchical adapter framework for text-guided medical image segmentation. BTHA is built around a stable feature-level interface: given multi-scale visual features and a text representation, it injects semantic guidance through shape-preserving adapters while maintaining the decoder-side tensor contract. To make this interface effective, we introduce a Hierarchical Coarse-to-Fine Supervision Strategy that decomposes learning into global image-text alignment, multi-scale auxiliary localization, and boundary-aware final mask refinement. We further design a Scale-Adaptive Gated Semantic Guidance (SAGSG) adapter, where resolution-specific gates adaptively control textual injection and channel recalibration suppresses redundant cross-modal responses. Evaluations across diverse vision and text backbones show that the same adapter and supervision design remains effective across convolutional and transformer-based visual encoders as well as different language encoders. Experiments on four public datasets further demonstrate that BTHA improves strong text-guided baselines with modest computational overhead.

Figures

Figures reproduced from arXiv: 2607.09481 by Haijin Zeng, Qi Dai, Xuanzi Fang, Yongyong Chen, Yungeng Liu.

Figure 1
Figure 1. Figure 1: Comparison between existing text-guided segmentation paradigms and our proposed method. Existing methods require fusion redesigns when backbones change, leading to limited reuse. Conversely, our method introduces a shape-preserving interface, enabling module reuse across diverse backbones. but becomes fragile when the feature hierarchy or language representation changes. For example, replacing a convolutio… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of BTHA. (a) Overall framework of BTHA. Heterogeneous vision and text backbones provide multi-scale visual features and text representations. The SAGSG adapter injects textual semantics into visual features, while the hierarchical supervision strategy regularizes global image-text alignment, multi-scale localization, and boundary-aware refinement. (b) Detailed structure of the SAGSG adapter. SAGSG… view at source ↗
Figure 3
Figure 3. Figure 3: Motivation for backbone-transferable language guidance. (a) Existing text-guided segmentation methods tightly couple the backbone pair, fusion module, and decoder. (b) Replacing the backbone changes the feature hierarchy and often requires a tailored fusion redesign. (c) BTHA uses a unified shape-preserving SAGSG adapter, allowing the same language-guidance module to support diverse vision and text backbon… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of segmentation results on four datasets. Rows 1 to 4 correspond to MosMedData+, QaTa-COV19, SIIM-ACR, and Kvasir-SEG, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 2 linked inside Pith

  1. [1]

    BiomedParse-V: Scaling foundation model for universal text-guided volumetric biomedical image segmentation,

    T. Zhao, H. H. Lee, A. Santamaria-Pang, N. C. Codella, S. Kiblawi, Y . Guet al., “BiomedParse-V: Scaling foundation model for universal text-guided volumetric biomedical image segmentation,” inMedSegFM. Springer Nature Switzerland, 2026, pp. 109–138

    2026

  2. [2]

    U-Net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- works for biomedical image segmentation,” inMICCAI. Springer International Publishing, 2015, pp. 234–241

    2015

  3. [3]

    nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,”Nat. Methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

  4. [4]

    UCTransNet: rethinking the skip connections in U-Net from a channel-wise perspective with transformer,

    H. Wang, P. Cao, J. Wang, and O. R. Zaiane, “UCTransNet: rethinking the skip connections in U-Net from a channel-wise perspective with transformer,” inAAAI, vol. 36, no. 3, 2022, pp. 2441–2449

    2022

  5. [5]

    LViT: Language meets vision transformer in medical image segmentation,

    Z. Li, Y . Li, Q. Li, P. Wang, D. Guo, L. Luet al., “LViT: Language meets vision transformer in medical image segmentation,”IEEE Trans. Med. Imaging, vol. 43, no. 1, pp. 96–107, 2024

    2024

  6. [6]

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

    Y . Zhong, M. Xu, K. Liang, K. Chen, and M. Wu, “Ariadne’s thread: Using text prompts to improve segmentation of infected areas from chest X-ray images,” inMICCAI. Springer, 2023, pp. 724–733

    2023

  7. [7]

    DuSSS: dual semantic similarity-supervised vision-language model for semi-supervised medi- cal image segmentation,

    Q. Pan, W. Qiao, J. Lou, B. Ji, and S. Li, “DuSSS: dual semantic similarity-supervised vision-language model for semi-supervised medi- cal image segmentation,” inAAAI, vol. 39, no. 6, 2025, pp. 6299–6307

    2025

  8. [8]

    Harnessing text insights with visual alignment for medical image segmentation,

    Q. Zeng, H. Luo, Z. Lu, Y . Xie, Z. Wang, Y . Zhanget al., “Harnessing text insights with visual alignment for medical image segmentation,” IEEE Trans. Med. Imaging, vol. 45, no. 2, pp. 477–489, 2026

    2026

  9. [9]

    TGS-LGP: Text-guided medical image segmentation via local-global perception,

    B. Ji, J. Huang, Z. Xu, M. Ou, T. Liu, S. Zenget al., “TGS-LGP: Text-guided medical image segmentation via local-global perception,” inBIBM, 2025, pp. 993–998

    2025

  10. [10]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafsonet al., “Segment anything,” inICCV, 2023, pp. 4015–4026

    2023

  11. [11]

    SAM- Adapter: Adapting segment anything in underperformed scenes,

    T. Chen, L. Zhu, C. Ding, R. Cao, Y . Wang, S. Zhanget al., “SAM- Adapter: Adapting segment anything in underperformed scenes,” in ICCV Workshops, 2023, pp. 3359–3367

    2023

  12. [12]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,”Nat. Commun., vol. 15, no. 1, p. 654, 2024

    2024

  13. [13]

    SAM 3: Segment anything with concepts,

    N. Carion, L. Gustafson, Y .-T. Hu, S. Debnath, R. Hu, D. S. Coll-Vinent et al., “SAM 3: Segment anything with concepts,” inICLR, 2026

    2026

  14. [14]

    Gloria: A multimodal global-local representation learning framework for label- efficient medical image recognition,

    S.-C. Huang, L. Shen, M. P. Lungren, and S. Yeung, “Gloria: A multimodal global-local representation learning framework for label- efficient medical image recognition,” inICCV, 2021, pp. 3942–3951

    2021

  15. [15]

    TGANet: Text-guided attention for improved polyp segmentation,

    N. K. Tomar, D. Jha, U. Bagci, and S. Ali, “TGANet: Text-guided attention for improved polyp segmentation,” inMICCAI. Springer Nature Switzerland, 2022, pp. 151–160

    2022

  16. [16]

    Text-guided cross-position attention for segmentation: Case of medical image,

    G.-E. Lee, S. H. Kim, J. Cho, S. T. Choi, and S.-I. Choi, “Text-guided cross-position attention for segmentation: Case of medical image,” in MICCAI. Springer Nature Switzerland, 2023, pp. 537–546

    2023

  17. [17]

    Common vision-language attention for text-guided medical image segmentation of pneumonia,

    Y . Guo, X. Zeng, P. Zeng, Y . Fei, L. Wen, J. Zhouet al., “Common vision-language attention for text-guided medical image segmentation of pneumonia,” inMICCAI, vol. LNCS 15009. Springer Nature Switzerland, 2024, pp. 192 – 201

    2024

  18. [18]

    Frequency- domain multi-modal fusion for language-guided medical image segmen- tation,

    B. Yu, J. Yang, Z. Du, Y . Huang, C. Li, and L. Wang, “Frequency- domain multi-modal fusion for language-guided medical image segmen- tation,” inMICCAI. Springer, 2025, pp. 278–288

    2025

  19. [19]

    Bi-VLGM: Bi-level class-severity- aware vision-language graph matching for text guided medical image segmentation,

    W. Chen, J. Liu, T. Liu, and Y . Yuan, “Bi-VLGM: Bi-level class-severity- aware vision-language graph matching for text guided medical image segmentation,”Int. J. Comput. Vis., vol. 133, no. 3, pp. 1375–1391, 2025

    2025

  20. [20]

    LGA: A language guide adapter for advancing the SAM model’s capabilities in medical image segmentation,

    J. Hu, Y . Li, H. Sun, Y . Song, C. Zhang, L. Linet al., “LGA: A language guide adapter for advancing the SAM model’s capabilities in medical image segmentation,” inMICCAI. Springer Nature Switzerland, 2024, pp. 610–620

    2024

  21. [21]

    Learning to exploit temporal structure for biomedical vision- language processing,

    S. Bannur, S. Hyland, Q. Liu, F. P ´erez-Garc´ıa, M. Ilse, D. C. Castro et al., “Learning to exploit temporal structure for biomedical vision- language processing,” inCVPR, 2023, pp. 15 016–15 027

    2023

  22. [22]

    Publicly available clinical BERT embeddings,

    E. Alsentzer, J. Murphy, W. Boag, W.-H. Weng, D. Jindi, T. Naumann et al., “Publicly available clinical BERT embeddings,” inClin. Nat. Lang. Process. Workshop. Association for Computational Linguistics, 2019, pp. 72–78

    2019

  23. [23]

    A visual–language foundation model for pathology image analysis using medical Twitter,

    Z. Huang, F. Bianchi, M. Yuksekgonul, T. J. Montine, and J. Zou, “A visual–language foundation model for pathology image analysis using medical Twitter,”Nat. Med., vol. 29, no. 9, pp. 2307–2316, 2023

    2023

  24. [24]

    G2D: From global to dense radiography representation learning via vision- language pre-training,

    C. Liu, C. Ouyang, S. Cheng, A. Shah, W. Bai, and R. Arcucci, “G2D: From global to dense radiography representation learning via vision- language pre-training,” inNeurIPS, vol. 37. Curran Associates, Inc., 2024, pp. 14 751–14 773

    2024

  25. [25]

    EviVLM: When evidential learning meets vision language model for medical image segmentation,

    Q. Pan, Z. Li, G. Yang, Q. Yang, and B. Ji, “EviVLM: When evidential learning meets vision language model for medical image segmentation,” IEEE Trans. Med. Imaging, vol. 45, no. 4, pp. 1369–1382, 2026

    2026

  26. [26]

    MedKLIP: Medical knowledge enhanced language-image pre-training for x-ray diagnosis,

    C. Wu, X. Zhang, Y . Zhang, Y . Wang, and W. Xie, “MedKLIP: Medical knowledge enhanced language-image pre-training for x-ray diagnosis,” inICCV, 2023, pp. 21 315–21 326

    2023

  27. [27]

    CXR- CLIP: Toward large scale chest x-ray language-image pre-training,

    K. You, J. Gu, J. Ham, B. Park, J. Kim, E. K. Honget al., “CXR- CLIP: Toward large scale chest x-ray language-image pre-training,” in MICCAI. Springer Nature Switzerland, 2023, pp. 101–111

    2023

  28. [28]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” inCVPR, 2022, pp. 11 976–11 986

    2022

  29. [29]

    Making the most of text semantics to improve biomedical vision–language processing,

    B. Boecking, N. Usuyama, S. Bannur, D. C. Castro, A. Schwaighofer, S. Hylandet al., “Making the most of text semantics to improve biomedical vision–language processing,” inECCV. Springer Nature Switzerland, 2022, pp. 1–21

    2022

  30. [30]

    UNETR: Transformers for 3d medical image segmentation,

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Landman et al., “UNETR: Transformers for 3d medical image segmentation,” in WACV, 2022, pp. 1748–1758

    2022

  31. [31]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal et al., “Learning transferable visual models from natural language supervision,” inICML, vol. 139. PMLR, 2021, pp. 8748–8763

    2021

  32. [32]

    MosMedData: Chest CT scans with COVID-19 related findings dataset,

    S. P. Morozov, A. E. Andreychenko, N. A. Pavlov, A. Vladzymyrskyy, N. V . Ledikhova, V . A. Gombolevskiyet al., “MosMedData: Chest CT scans with COVID-19 related findings dataset,”arXiv preprint arXiv:2005.06465, 2020

    Pith/arXiv arXiv 2005

  33. [33]

    OSegNet: Operational segmentation network for COVID-19 detection using chest x-ray images,

    A. Degerli, S. Kiranyaz, M. E. H. Chowdhury, and M. Gabbouj, “OSegNet: Operational segmentation network for COVID-19 detection using chest x-ray images,” inICIP, 2022, pp. 2306–2310

    2022

  34. [34]

    SIIM-ACR pneumothorax segmentation 2019,

    A. Zawacki, C. Wu, G. Shih, J. Elliott, M. Fomitchev, M. Hussainet al., “SIIM-ACR pneumothorax segmentation 2019,” 2019

    2019

  35. [35]

    Kvasir-seg: A segmented polyp dataset,

    D. Jha, P. H. Smedsrud, M. A. Riegler, P. Halvorsen, T. De Lange, D. Johansenet al., “Kvasir-seg: A segmented polyp dataset,” inMMM. Springer, 2019, pp. 451–462

    2019

  36. [36]

    Swin Transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhanget al., “Swin Transformer: Hierarchical vision transformer using shifted windows,” inICCV, 2021, pp. 10 012–10 022

    2021

  37. [37]

    Training data-efficient image transformers & distillation through attention,

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jegou, “Training data-efficient image transformers & distillation through attention,” inICML, M. Meila and T. Zhang, Eds., vol. 139. PMLR, 2021, pp. 10 347–10 357

    2021

  38. [38]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016, pp. 770–778

    2016

  39. [39]

    MultiResUNet: Rethinking the U-Net architecture for multimodal biomedical image segmentation,

    N. Ibtehaz and M. S. Rahman, “MultiResUNet: Rethinking the U-Net architecture for multimodal biomedical image segmentation,”Neural Netw., vol. 121, pp. 74–87, 2020

    2020

  40. [40]

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

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tianet al., “Swin- Unet: Unet-like pure transformer for medical image segmentation,” in ECCV Workshops. Springer Nature Switzerland, 2023, pp. 205–218

    2023

  41. [41]

    SAM-Med2D,

    J. Cheng, J. Ye, Z. Deng, J. Chen, T. Li, H. Wanget al., “SAM-Med2D,” arXiv preprint arXiv:2308.16184, 2023

    Pith/arXiv arXiv 2023

  42. [42]

    Cross- modal conditioned reconstruction for language-guided medical image segmentation,

    X. Huang, H. Li, M. Cao, L. Chen, C. You, and D. An, “Cross- modal conditioned reconstruction for language-guided medical image segmentation,”IEEE Trans. Med. Imaging, vol. 44, no. 4, pp. 1821– 1835, 2025

    2025