Pith. sign in

REVIEW 3 major objections 4 minor 46 references

DistMedVL: Distributional Vision-Language Alignment for Uncertainty-Aware Medical Image Segmentation

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read DistMedVL claims that replacing deterministic cross-modal matching with variance-weighted probabilistic alignment makes medical image segmentation robust to uncertainty, and demonstrates consistent state-of-the-art results across eight…

desk verdict The mask-guided prompts for TN3K and PKTN leak the ground truth into the text encoder, so the headline SOTA claim is not supported by the current experiments. read the letter →

arxiv 2608.05683 v1 pith:GUXE23F3 submitted 2026-08-06 cs.CV cs.AIcs.MM

Jiaxuan Li , Qing Xu , Xiangjian He , Yue Li , Daokun Zhang , Fiseha B. Tesema , Rong Qu This is my paper · ORCID
classification cs.CVcs.AIcs.MM
keywords medicalimagesegmentationvision-languagemodelsuncertaintyestimationcross-modalalignmentMahalanobisdistanceprobabilisticembeddingsdomaingeneralizationdataefficiency
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

DistMedVL claims that the fragility of vision-language medical image segmentation under real-world uncertainty comes from deterministic cross-modal matching, and proposes replacing it with probabilistic alignment. The paper introduces a lightweight Probabilistic Cross-Modal Adapter on frozen encoders, where textual tokens are treated as Gaussian distributions and patch-text compatibility is scored by Mahalanobis distance with variance weighing. A second module estimates modality-wise confidence and performs reliability-gated vision-guided refinement of textual distributions. Across eight benchmarks, the method reports state-of-the-art Dice and NSD with only 6.3M trainable parameters, with gains that widen under data scarcity, domain shift, and input perturbations. A sympathetic reader would care because the method targets exactly the conditions that make clinical segmentation unreliable: ambiguous boundaries, limited annotations, and unseen imaging protocols.

What carries the argument

The central object is the Probabilistic Cross-Modal Adapter (PCM-Adapter), a bottleneck module inserted at several transformer layers of frozen CLIP vision and text encoders. Its first component, the Mahalanobis Alignment Module, estimates a mean and variance for each textual token and an image-level visual variance, fuses them into a bilateral variance, and computes a normalized Mahalanobis distance between each visual patch and each textual Gaussian; a softmax over these distances yields a probabilistic patch-text compatibility matrix. Its second component, the Distribution Flow Module, computes a global visual mean and inter-patch variance, maps that variance through a reliability gate, and uses a residual MLP to refine the textual distribution only with visual evidence that is internally consistent. The variance-conditioned distance is the mechanism that downweights uncertain feature dimensions, and the reliability gate is the mechanism that suppresses unreliable global visual signals.

What would settle it

Train DistMedVL on a benchmark but replace the estimated variances with fixed positive constants and disable the reliability gate, while keeping all other components; if Dice and NSD stay within a small margin of the full model, the claimed uncertainty mechanism is not the cause of the robustness. A stronger test would use corrupted inputs with known noise levels and check whether the estimated variances increase monotonically with the injected noise.

Watch

Extended reading notes

Core claim

The central claim is that explicitly modeling representational uncertainty in cross-modal alignment improves medical image segmentation. Instead of treating visual and textual features as fixed points, the Mahalanobis Alignment Module models each text token as a Gaussian and computes patch-text distance inversely weighted by per-dimension variance, so unreliable dimensions contribute less. The Distribution Flow Module then aggregates visual patches into a global semantic summary, measures inter-patch variance as a reliability signal, and gates how much visual evidence refines the textual distribution. The paper argues this accounts for both aleatoric uncertainty (noisy inputs, ambiguous boundaries) and epistemic uncertainty (limited data, domain shift), and demonstrates consistent improvements over state-of-the-art methods, especially when supervision is scarce or the test domain shifts.

Load-bearing premise

The argument depends on the learned variance parameters and reliability scores genuinely representing uncertainty, rather than being arbitrary per-dimension scalings that improve the training loss without carrying information about ambiguity or data scarcity.

Editorial extensions

If this is right

  • With limited training data, the method's advantage over deterministic baselines grows, suggesting probabilistic alignment is most valuable precisely when epistemic uncertainty is high.
  • In cross-domain transfer, the gains are largest for the most distant domain pair, implying variance-aware matching may help models recalibrate when encountering unseen imaging protocols.
  • Under both image blur and text perturbation, degradation is smaller than for the probabilistic baseline MedCLIPSeg, indicating the dual-module design captures more of the perturbation signal.
  • Because both encoders stay frozen and only 6.3M parameters are trained, the approach can be adapted to new tasks at relatively low computational cost.

Reading between the lines

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

  • Inference beyond the paper: the estimated variances may function as learned per-dimension scalings rather than calibrated uncertainty estimates, but the method could still transfer; a testable extension is to compare against a version with fixed constant variances.
  • Inference beyond the paper: the same PCM-Adapter design could apply to other dense prediction tasks such as referring expression segmentation or visual grounding, where sparse text tokens guide noisy visual inputs.
  • Inference beyond the paper: the reliability gate in the DFM could be replaced by other consistency signals (e.g., entropy of patch features), and performance differences would reveal whether inter-patch variance is the essential ingredient.
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 / 4 minor

Summary. The paper proposes DistMedVL, a probabilistic vision-language framework for medical image segmentation built on frozen CLIPSeg encoders with a lightweight trainable PCM-Adapter. The adapter contains two modules: a Mahalanobis Alignment Module (MAM) that models text tokens as Gaussian distributions and computes patch-text compatibility with variance-weighted Mahalanobis distances, and a Distribution Flow Module (DFM) that performs vision-guided, reliability-gated refinement of textual distributions. The method is evaluated on eight benchmarks under data-efficiency, domain-generalization, and perturbation scenarios, and reports consistent improvements over unimodal and cross-modal baselines with only 6.3M trainable parameters.

Significance. If the reported results are valid, DistMedVL is a credible lightweight alternative for uncertainty-aware cross-modal medical segmentation, and the combination of Mahalanobis matching with a reliability gate is a reasonable design. The paper provides broad empirical coverage: four data ratios, three domain-shift settings, two perturbation types, and an ablation study. However, the evaluation is compromised for two datasets whose text prompts are generated from ground-truth masks, and the interpretation of learned variance parameters as genuine uncertainty estimates is not validated against any external uncertainty benchmark. Both issues bear directly on the central claims, so the significance of the contribution as currently evidenced is limited pending correction.

major comments (3)
  1. [IV-A, Tables II and III] The text generation protocol for TN3K and PKTN is a load-bearing confound. Section IV-A states that for these datasets 'text descriptions are generated using Qwen2.5, guided by dataset descriptions and ground-truth masks.' Because the MAM converts text tokens into a dense semantic prior S (Eq. 7) that is gated into visual features, prompts derived from ground-truth masks can leak the segmentation target, and possibly location and shape information, into the model through the text modality. TN3K is one of the five datasets averaged in Table II, so the headline 87.63% Dice result and the data-efficiency comparisons are contaminated by mask-derived text. PKTN is the out-of-distribution target in Table III, so the reported 68.02% Dice and 77.48% NSD for TN3K-to-PKTN transfer do not measure cross-dataset generalization; they partly measure how well the text encoder decodes the mask. The paper does not release the prompt templates or state that the prompts are class-level generic rather than per-image mask-derived. The authors must either release the prompt generation pipeline and demonstrate that prompts do not encode mask information, or rerun the TN3K and PKTN experiments with prompts generated without ground-truth masks, before the claimed consistent state-of-the-art performance across eight benchmarks can be accepted.
  2. [III-C, III-D, Eq. (4), Eq. (8)] The central conceptual claim is that the variance parameters in the MAM (Eq. 4) and the reliability gate in the DFM (Eq. 8) reflect aleatoric and epistemic uncertainty. However, these parameters are learned by optimizing the same segmentation objective as the rest of the network, and no external validation is provided: there are no calibration experiments, no correlation with label noise or annotation ambiguity, no comparison with MC Dropout or deep ensembles beyond qualitative uncertainty maps in Fig. 5, and no analysis of how the learned variances change under domain shift. It remains possible that these are arbitrary learned scalings that improve training loss without carrying meaningful uncertainty information. The robustness claims in Sections IV-E and IV-F are therefore only partially supported. I recommend adding quantitative uncertainty evaluation, such as selective prediction with error-retention curves, calibration of predicted probabilities, or a direct comparison of learned variance magnitudes under clean versus perturbed inputs and under in-distribution versus out-of-distribution data.
  3. [Tables II and III] No error bars or statistical significance tests are reported, despite several margins being small relative to the number of seeds typically used in medical segmentation experiments. For example, in Table III the BUSI-to-BUID advantage over MedCLIPSeg is 0.77% Dice and the colonoscopy target advantage is 1.04% Dice; without multiple runs and standard deviations, the claim that DistMedVL 'consistently outperforms' all baselines is not established. The Table II data-efficiency results would also be more convincing if at least three seeds were reported for the proposed method and the strongest baselines. This is a load-bearing issue because the paper's central claim is based on ranking by averaged point estimates.
minor comments (4)
  1. [Fig. 2 and Table V] The module is called the Mahalanobis Alignment Module (MAM) in the text, but Fig. 2 uses 'MDM' and Table V uses 'MDM' throughout; please harmonize the acronym.
  2. [Eq. (1)] Equation (1) defines F_v^down and F_t^down, then defines tilde F_v and tilde F_t in terms of hat F_v and hat F_t, but hat F_v and hat F_t are not defined until Eq. (2); the notation flow should be clarified.
  3. [IV-F] The text-perturbation operation is described only as 'spatial orientation inversion for text prompts'; please specify exactly how this is implemented, since the interpretation of Table IV depends on it.
  4. [IV-A] The paper should release the exact text prompt templates for all datasets, especially TN3K and PKTN, and state whether prompts are per-image or per-class, to allow reproducibility and to resolve the leakage concern.

Circularity Check

1 steps flagged · score 6.0 of 10

TN3K and PKTN text prompts are generated from ground-truth masks and then injected into visual features (Eq. 7), so the claimed SOTA gains on those benchmarks are partly forced by construction.

  1. self definitional [Section IV-A (Datasets) with Section III-C, Eqs. (6)-(7)]
    "For TN3K and PKTN, text descriptions are generated using Qwen2.5 [42], guided by dataset descriptions and ground-truth masks. ... Based on this assignment, a dense semantic prior S=A T∈R N×H is generated for each visual patch through weighted aggregation over textual features. The semantic prior is then adaptively integrated into the visual representation via a gating mechanism: ˆF(ℓ) v =W o(V+g(V)⊙S)."

    The text input for TN3K and PKTN is constructed from the ground-truth mask, the task's target variable. MAM maps those text tokens to a dense semantic prior S=A T (Eq. 6) and fuses it into visual features via F_v=W_o(V+g(V)⊙S) (Eq. 7), so the mask-derived prompt injects the segmentation target into the decoder features. TN3K contributes to the averaged Table II data-efficiency results (headline 87.63% Dice / 92.36% NSD) and to Table IV; PKTN is the declared OOD target in Table III, so its 68.02% Dice / 77.48% NSD and the claimed 2.38% Dice / 4.20% NSD 'advantage' partly measure mask decoding rather than domain generalization. The paper does not state that prompts are class-level generic instead of per-image mask-derived, so the reduction is not ruled out.

full rationale

The circularity is genuine but localized to the evaluation protocol. The architectural derivation of the PCM-Adapter is self-contained: MAM (Eqs. 3-5), DFM (Eqs. 8-11), and the gating in Eq. 7 are defined by their own learnable projections and trained with the stated segmentation/contrastive objective; no load-bearing self-citation or imported uniqueness theorem is used to justify the design. However, Section IV-A states that for TN3K and PKTN the text prompt is generated 'guided by ... ground-truth masks,' and Eqs. (6)-(7) turn text tokens into a semantic prior that is gated into the visual features. Hence on these datasets the target mask is an input to the model, making the 'prediction' partly equivalent to its input by construction. This contaminates the headline data-efficiency average (Table II), the perturbation comparison (Table IV), and the out-of-distribution claim on TN3K→PKTN (Table III). The learned variance parameters being unvalidated as uncertainty estimates is a separate interpretability concern, not itself a circular reduction, so I do not score it as a step. Other benchmarks use external MedCLIPSeg splits and prompts and the ablation includes decoder-only baselines, giving the architecture independent content. Overall, this is partial circularity of the central evaluation claim, not of the method's derivation.

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

The method relies on learned variance parameters and gates that are interpreted as uncertainty, but they are optimized for the segmentation loss. Standard assumptions about the transferability of frozen CLIPSeg encoders and the sufficiency of diagonal Gaussian distributions are also load-bearing.

free parameters (3)
  • Temperature tau = learned, constrained to [tau_min, tau_max]
    Controls softmax sharpness in Eq. 6; fitted during training.
  • Loss weight lambda2 = 0.1
    Contrastive loss weight adopted from MedCLIPSeg, hand-chosen.
  • Variance fusion gate theta_lambda = learned
    Per-dimension interpolation weight in Eq. 4.
assumptions (3)
  • domain assumption Gaussian distributions with diagonal covariance adequately capture representational uncertainty
    Core to MAM; no justification beyond tractability.
  • domain assumption CLIPSeg frozen encoders provide transferable medical features
    Foundation of the method, not validated in paper.
  • ad hoc to paper Learned variances reflect genuine aleatoric and epistemic uncertainty
    Interpreted as uncertainty but trained solely to minimize segmentation loss.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DistMedVL: Distributional Vision-Language Alignment for Uncertainty-Aware Medical Image Segmentation." pith.science (2026). https://pith.science/paper/GUXE23F3

@misc{pith2026260805683,
  author       = {Pith},
  title        = {Pith review of: DistMedVL: Distributional Vision-Language Alignment for Uncertainty-Aware Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUXE23F3}},
  note         = {Machine review of arXiv:2608.05683}
}
read the original abstract

Cross-modal alignment of visual and textual representations is fundamental to multimodal medical image understanding, yet remains hindered by uncertainty in both modalities under real-world clinical conditions. Existing vision-language segmentation methods rely on deterministic cross-modal matching, which overlooks aleatoric uncertainty from ambiguous boundaries and epistemic uncertainty from limited training data, leading to fragile performance under domain shift. To address this issue, we propose DistMedVL, a probabilistic vision-language framework that introduces a lightweight Probabilistic Cross-Modal Adapter (PCM-Adapter) upon frozen encoders to explicitly model representational uncertainty. Specifically, the PCM-Adapter comprises two sequential modules for progressive probabilistic alignment. We first devise a Mahalanobis Alignment Module (MAM) that models textual tokens as Gaussian distributions and computes patch-text compatibility via Mahalanobis distance, yielding variance-conditioned matching that downweights unreliable feature dimensions. Moreover, we devise a Distribution Flow Module (DFM) that estimates modality-wise confidence parameters and performs vision-guided refinement of textual distributions, accommodating distributional variation across imaging modalities. Extensive experiments across eight medical segmentation benchmarks demonstrate that DistMedVL outperforms state-of-the-art methods with only 6.3M trainable parameters, exhibiting superior data efficiency, perturbation robustness and cross-dataset generalization.

Figures

Figures reproduced from arXiv: 2608.05683 by the authors.

Figure 1
Figure 1. (a) Overview of the DistMedVL architecture, where only the PCM adapter and decoder are trainable. (b) Detail of the PCM Adapter, which maps [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of the proposed method. (a) Mahalanobis Distance Module (MDM), which maps image features to text features. (b) Distribution [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of vision-language models on medical image segmentation under the 100% training data setting. White regions indicate [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of domain shift between source and target datasets via PCA projection of extracted features. The Maximum Mean Discrepancy (MMD) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of uncertainty maps. Warmer colors indicate [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 38 canonical work pages

  1. [1]

    A survey on deep learning in medical image analysis,

    G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I. S´anchez, “A survey on deep learning in medical image analysis,”Med. Image Anal., vol. 42, pp. 60–88, 2017

    work page 2017

  2. [2]

    Medical image analysis using convolutional neural net- works: a review,

    S. M. Anwar, M. Majid, A. Qayyum, M. Awais, M. Alnowami, and M. K. Khan, “Medical image analysis using convolutional neural net- works: a review,”J. Med. Syst., vol. 42, no. 11, p. 226, 2018

    work page 2018

  3. [3]

    Self-supervised learning methods and applications in medical imaging analysis: A survey,

    S. Shurrab and R. Duwairi, “Self-supervised learning methods and applications in medical imaging analysis: A survey,”PeerJ Comput. Sci., vol. 8, p. e1045, 2022

    work page 2022

  4. [4]

    Medclip: Contrastive learning from unpaired medical images and text,

    Z. Wang, Z. Wu, D. Agarwal, and J. Sun, “Medclip: Contrastive learning from unpaired medical images and text,” inEMNLP, 2022, pp. 3876– 3887

    work page 2022

  5. [5]

    U-net: Convolutional networks for biomedical image segmentation,

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

    2015

  6. [6]

    Fat-net: Feature adaptive transformers for automated skin lesion segmentation,

    H. Wu, S. Chen, G. Chen, W. Wang, B. Lei, and Z. Wen, “Fat-net: Feature adaptive transformers for automated skin lesion segmentation,” Med. Image Anal., vol. 76, p. 102327, 2022

    work page 2022

  7. [7]

    Swin-unet: Unet-like pure transformer for medical image segmenta- tion,

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-unet: Unet-like pure transformer for medical image segmenta- tion,” inECCV, 2022, pp. 205–218

    2022

  8. [8]

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

    R. Azad, M. Heidari, M. Shariatnia, E. K. Aghdam, S. Karimijafarbigloo, E. Adeli, and D. Merhof, “Transdeeplab: Convolution-free transformer- based deeplab v3+ for medical image segmentation,” inProc. PRIME Workshop, 2022, pp. 91–102

    work page 2022

Show all 46 references
  1. [9]

    Transunet: Transformers make strong encoders for medical image segmentation,

    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

    2021 arXiv

  2. [10]

    H2former: An efficient hierarchical hybrid transformer for medical image segmentation,

    A. He, K. Wang, T. Li, C. Du, S. Xia, and H. Fu, “H2former: An efficient hierarchical hybrid transformer for medical image segmentation,”IEEE Trans. Med. Imaging, vol. 42, no. 9, pp. 2763–2775, 2023

    2023

  3. [11]

    Learning transferable visual models from natural language supervision,

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

    2021

  4. [12]

    Vlsm-adapter: Finetuning vision-language segmentation efficiently with lightweight blocks,

    M. Dhakal, R. Adhikari, S. Thapaliya, and B. Khanal, “Vlsm-adapter: Finetuning vision-language segmentation efficiently with lightweight blocks,” inMICCAI, 2024, pp. 712–722

    2024

  5. [13]

    Lvit: language meets vision transformer in medical image segmentation,

    Z. Li, Y . Li, Q. Li, P. Wang, D. Guo, L. Lu, D. Jin, Y . Zhang, and Q. Hong, “Lvit: language meets vision transformer in medical image segmentation,”IEEE Trans. Med. Imaging, vol. 43, no. 1, pp. 96–107, 2023

    2023

  6. [14]

    Lavt: Language-aware vision transformer for referring image segmentation,

    Z. Yang, J. Wang, Y . Tang, K. Chen, H. Zhao, and P. H. Torr, “Lavt: Language-aware vision transformer for referring image segmentation,” inCVPR, 2022, pp. 18 155–18 165

    2022

  7. [15]

    Denseclip: Language-guided dense prediction with context-aware prompting,

    Y . Rao, W. Zhao, G. Chen, Y . Tang, Z. Zhu, G. Huang, J. Zhou, and J. Lu, “Denseclip: Language-guided dense prediction with context-aware prompting,” inCVPR, 2022, pp. 18 082–18 091

    2022

  8. [16]

    From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation,

    R. Wen, H. Yuan, D. Ni, W. Xiao, and Y . Wu, “From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation,” inWACV, 2024, pp. 464–474

    2024

  9. [17]

    Enhancing the vision–language foundation model with key semantic knowledge-emphasized report refinement,

    W. Huang, C. Li, H. Yang, J. Liu, Y . Liang, H. Zheng, and S. Wang, “Enhancing the vision–language foundation model with key semantic knowledge-emphasized report refinement,”Med. Image Anal., vol. 97, p. 103299, 2024

    2024

  10. [18]

    Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation,

    N. Tajbakhsh, L. Jeyaseelan, Q. Li, J. N. Chiang, Z. Wu, and X. Ding, “Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation,”Med. Image Anal., vol. 63, p. 101693, 2020

    2020

  11. [19]

    From deterministic to probabilistic: A novel perspective on domain generalization for medical image segmentation,

    Y . Xu and T. Zhang, “From deterministic to probabilistic: A novel perspective on domain generalization for medical image segmentation,” arXiv preprint arXiv:2412.05572, 2024

    2024 arXiv

  12. [20]

    Medical image segmentation review: The success of u-net,

    R. Azad, E. K. Aghdam, A. Rauland, Y . Jia, A. H. Avval, A. Bozorgpour, S. Karimijafarbigloo, J. P. Cohen, E. Adeli, and D. Merhof, “Medical image segmentation review: The success of u-net,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 12, pp. 10 076–10 095, 2024

    2024

  13. [21]

    Dcsau-net: A deeper and more compact split-attention u-net for medical image segmentation,

    Q. Xu, Z. Ma, W. Duanet al., “Dcsau-net: A deeper and more compact split-attention u-net for medical image segmentation,”Comput. Biol. Med., vol. 154, p. 106626, 2023

    2023

  14. [22]

    Cfformer: Cross cnn-transformer channel attention and spatial feature fusion for improved segmentation of heterogeneous medical images,

    J. Li, Q. Xu, X. He, Z. Liu, D. Zhang, R. Wang, R. Qu, and G. Qiu, “Cfformer: Cross cnn-transformer channel attention and spatial feature fusion for improved segmentation of heterogeneous medical images,” Expert Syst. Appl., vol. 295, p. 128835, 2026

    2026

  15. [23]

    Pubmedclip: How much does clip benefit visual question answering in the medical domain?

    S. Eslami, C. Meinel, and G. De Melo, “Pubmedclip: How much does clip benefit visual question answering in the medical domain?” inFindings Assoc. Comput. Linguist.: EACL, 2023, pp. 1181–1193

    2023

  16. [24]

    Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities,

    M. U. Khattak, S. Kunhimon, M. Naseer, S. Khan, and F. S. Khan, “Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities,”arXiv preprint arXiv:2412.10372, 2024

    2024 arXiv

  17. [25]

    Causalclipseg: Unlocking clip’s potential in referring medical image segmentation with causal intervention,

    Y . Chen, M. Wei, Z. Zheng, J. Hu, Y . Shi, S. Xiong, X. X. Zhu, and L. Mou, “Causalclipseg: Unlocking clip’s potential in referring medical image segmentation with causal intervention,” inMICCAI, 2024, pp. 77–87

    2024

  18. [26]

    Medclipseg: Probabilistic vision-language adaptation for data-efficient and generalizable medical image segmentation,

    T. Koleilat, H. Asgariandehkordi, O. Nejatimanzari, B. Barile, Y . Xiao, and H. Rivaz, “Medclipseg: Probabilistic vision-language adaptation for data-efficient and generalizable medical image segmentation,” inCVPR, June 2026, pp. 1406–1417

    2026

  19. [27]

    What uncertainties do we need in bayesian deep learning for computer vision?

    A. Kendall and Y . Gal, “What uncertainties do we need in bayesian deep learning for computer vision?”Adv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  20. [28]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning,

    Y . Gal and Z. Ghahramani, “Dropout as a bayesian approximation: Representing model uncertainty in deep learning,” inICML. PMLR, 2016, pp. 1050–1059

    2016

  21. [29]

    Simple and scalable predictive uncertainty estimation using deep ensembles,

    B. Lakshminarayanan, A. Pritzel, and C. Blundell, “Simple and scalable predictive uncertainty estimation using deep ensembles,”Adv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  22. [30]

    Deep deterministic uncertainty: A new simple baseline,

    J. Mukhoti, A. Kirsch, J. Van Amersfoort, P. H. Torr, and Y . Gal, “Deep deterministic uncertainty: A new simple baseline,” inCVPR, 2023, pp. 24 384–24 394

    2023

  23. [31]

    Probabilistic embeddings for cross-modal retrieval,

    S. Chun, S. J. Oh, R. S. De Rezende, Y . Kalantidis, and D. Larlus, “Probabilistic embeddings for cross-modal retrieval,” inCVPR, 2021, pp. 8415–8424

    2021

  24. [32]

    Probvlm: Proba- bilistic adapter for frozen vison-language models,

    U. Upadhyay, S. Karthik, M. Mancini, and Z. Akata, “Probvlm: Proba- bilistic adapter for frozen vison-language models,” inICCV, 2023, pp. 1899–1910

    2023

  25. [33]

    Image segmentation using text and image prompts,

    T. L ¨uddecke and A. Ecker, “Image segmentation using text and image prompts,” inCVPR, June 2022, pp. 7086–7096

    2022

  26. [34]

    Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules,

    H. Gong, J. Chen, G. Chen, H. Li, F. Chen, and G. Li, “Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules,” Comput. Biol. Med., vol. 106389, pp. 1–12, 2022

    2022

  27. [35]

    Dataset of breast ultrasound images,

    W. Al-Dhabyani, M. Gomaa, H. Khaled, and A. Fahmy, “Dataset of breast ultrasound images,”Data Brief, vol. 28, p. 104863, 2020

    2020

  28. [36]

    Gutman, N

    D. Gutman, N. C. Codella, E. Celebi, B. Helba, M. Marchetti, N. Mishra, and A. Halpern, “Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic),...

    2016 arXiv

  29. [37]

    Kvasir-seg: A segmented polyp dataset,

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

    2019

  30. [38]

    Osegnet: Operational segmentation network for covid-19 detection using chest x-ray images,

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

    2022

  31. [39]

    An open-access breast lesion ultrasound image database: Applicable in artificial intelligence studies,

    A. A. Ardakani, A. Mohammadi, M. Mirza-Aghazadeh-Attari, and U. R. Acharya, “An open-access breast lesion ultrasound image database: Applicable in artificial intelligence studies,”Comput. Biol. Med., vol. 152, p. 106438, 2023

    2023

  32. [40]

    Automated polyp detection in colonoscopy videos using shape and context information,

    N. Tajbakhsh, S. R. Gurudu, and J. Liang, “Automated polyp detection in colonoscopy videos using shape and context information,”IEEE Trans. Med. Imaging, vol. 35, no. 2, pp. 630–644, 2015

    2015

  33. [41]

    Clip-tnseg: A multi- modal hybrid framework for thyroid nodule segmentation in ultrasound images,

    X. Sun, B. Wei, Y . Jiang, L. Mao, and Q. Zhao, “Clip-tnseg: A multi- modal hybrid framework for thyroid nodule segmentation in ultrasound images,” 2024

    2024

  34. [42]

    Qwen2. 5-coder technical report,

    B. Hui, J. Yang, Z. Cui, J. Yang, D. Liu, L. Zhang, T. Liu, J. Zhang, B. Yu, K. Luet al., “Qwen2. 5-coder technical report,”arXiv preprint arXiv:2409.12186, 2024

    2024 arXiv

  35. [43]

    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, 2022, pp. 151–160

    2022

  36. [44]

    Image segmentation using text and image prompts,

    T. L ¨uddecke and A. Ecker, “Image segmentation using text and image prompts,” inCVPR, 2022, pp. 7086–7096

    2022

  37. [45]

    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, 2023, pp. 724–733

    2023

  38. [46]

    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, 2024

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.