Pith. sign in

REVIEW 4 major objections 5 minor 49 references

Multi-Sequence Parotid Gland Lesion Segmentation via Expert Text-Guided Segment Anything Model

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Structured expert diagnostic text, turned into point-and-box prompts, can guide a fine-tuned Segment Anything Model to segment parotid gland lesions across multiple MRI sequences and three clinical centers.

desk verdict A reasonable engineering paper with a useful dataset, but the central SOTA claim is confounded because the expert text carries target-derived location and size information that the baselines never see. read the letter →

arxiv 2508.09645 v1 pith:GUJ7MASS submitted 2025-08-13 cs.CV

classification cs.CV
keywords parotidglandlesionsegmentationSegmentAnythingModelexpertdiagnostictextcross-sequenceattentionmulti-sequenceMRIpromptgenerationmedicalimageLoRAfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes PG-SAM, a parotid-gland lesion segmentation model that uses structured expert diagnostic text—location and size of the lesion—to automatically generate the point and box prompts that SAM needs, instead of relying on a user to click on the lesion. It couples this prompt generator with a cross-sequence attention module that fuses four MRI sequences (T1, T1C, T2, ADC) into a shared embedding. On three clinical datasets, PG-SAM reports the best or second-best Dice similarity coefficient on most modality–center combinations, including internal and external test sets, with the largest gains on T1 and T1C sequences. The authors interpret this as evidence that radiological domain knowledge, encoded as text, improves SAM-based medical image segmentation without manual annotations.

What carries the argument

The load-bearing machinery is the expert-text-to-prompt pathway: MedCLIP encodes the report, an adapter transfers the text embedding to the medical domain, a text guide decoder fuses it with the fused multi-sequence image embedding, and the resulting coarse mask is masked by a pre-computed spatial prior $M$ before a prompt point (centroid) and prompt box (maximal border) are extracted. These prompts are fed through SAM's prompt encoder into a hierarchical mask decoder with class-balanced mask-guided self-attention and learnable mask cross-attention. The cross-sequence attention formula $A_i = \mathrm{Softmax}(x_f W_q (x_i W_k)^\top / \sqrt{d_k}) W_v$ is the mechanism that lets each sequence

What would settle it

On the Site1 internal test set, rerun the trained PG-SAM with expert text edited so that the size field [xx] is replaced by a random value from the dataset's size distribution and the location field [pos] is moved to the contralateral side, keeping images, masks, and all other inputs fixed. If DSC and HD95 remain at the reported levels, target-derived text is not the source of the gains; if they collapse, the reported advantage over text-free baselines is partly an artifact of privileged ground-truth information in the prompt.

Watch

Extended reading notes

Core claim

PG-SAM's central claim is that expert knowledge from diagnostic reports can replace manual prompts for SAM in parotid lesion segmentation. The expert text is embedded with MedCLIP, adapted to the medical domain, and fused with multi-sequence image features in a text guide decoder. A coarse mask is generated from that fusion, constrained by a spatial prior mask of where parotid lesions typically appear; the centroid and bounding box of the coarse mask become the prompt point and prompt box. Meanwhile, a cross-sequence attention module lets each sequence's embedding be corrected by information from the other sequences. The paper reports that this combined design outperforms the compared UNet-b

Load-bearing premise

The load-bearing premise is that the expert diagnostic text used to build the prompts does not carry measurements (lesion size and location) taken from the same ground-truth masks the model is trained and scored against, and that the four MRI sequences are aligned well enough for pixel-wise cross-sequence fusion.

Editorial extensions

If this is right

  • If PG-SAM is correct, structured radiology reports can serve as cheap, automatically derived prompts for SAM, removing the need for per-slice manual point/box annotations.
  • Fusing T1, T1C, T2 and ADC through cross-sequence attention should make parotid segmentation more consistent across modalities than single-sequence models.
  • The paper's few-sample experiments imply that with only 10–30% of labeled slices, PG-SAM stays above 0.7 DSC on internal T1 and above 0.6 on external sets, suggesting the method is usable in low-annotation clinical settings.
  • The spatial prior mask, built from where parotid lesions occur, makes prompt generation more stable and is a component that could be reused for other anatomically localized lesions.
  • Combining text-guided prompts with parameter-efficient LoRA fine-tuning keeps the method practical: the paper reports training time cut to one-third and roughly 40% deployment-efficiency improvement relative to the traditional comparison methods.

Reading between the lines

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

  • Because the template text includes the lesion's measured size (e.g., 29×25×35 mm) and its location, and because those numbers come from the same lesion that defines the ground-truth mask, part of the reported gain may be attributable to privileged target information rather than to the architecture; a blinded text condition would settle this.
  • A clean extension would randomize or swap only the size field while keeping location fixed, or vice versa, to identify which textual variable actually drives the prompt's value.
  • The cross-sequence attention fuses features pixel-wise without an explicit registration step; if clinical MRI sequences are misaligned, the benefit could flip to harm, so a test that shifts one sequence by a few pixels would quantify this sensitivity.
  • The same expert-text-to-prompt mechanism could transfer to other anatomically localized lesions (e.g., thyroid, lymph nodes) where structured radiology reports exist, but the spatial prior mask would need to be rebuilt for each anatomy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PG-SAM, a SAM-based framework for multi-sequence MRI segmentation of parotid gland lesions. A prompt-generation module uses expert diagnostic text (encoded with MedCLIP and adapted) together with multi-sequence image features to produce a coarse mask, from which point and bounding-box prompts are derived. A cross-sequence attention module fuses T1/T1C/T2/ADC embeddings, and a hierarchical mask decoder (with class-balanced mask-guided attention and learnable mask cross-attention) produces the final segmentation. The method is trained on Site1 and evaluated on internal Site1 and external Site2/Site3 datasets, reporting DSC, HD95, ACC, and REC against UNet-family, transformer, and SAM-based baselines. The authors claim state-of-the-art performance across three centers.

Significance. If the central claim were established, the paper would make a useful contribution to SAM-based medical image segmentation by showing how expert text can replace manual prompts. The proposed architecture is plausible and the reported internal results are competitive. However, the evaluation is confounded: the expert text explicitly contains the lesion's location and dimensions, which are measurements of the same target that defines the ground-truth mask. The baselines receive no such privileged input, so the Table 2 comparison does not demonstrate architectural superiority. External-site gains are small and sometimes negative, no error bars or significance tests are reported, and the ablation tables do not consistently support the claimed benefits of the two main modules. These issues are load-bearing for the paper's main claims; with a fair comparison and proper statistical analysis the method may still be valuable, but the current manuscript does not establish it.

major comments (4)
  1. [§3.4, Table 4, Fig. 2] The expert text template in Table 4 is 'This is an MRI image with a lesion located in the [pos] parotid gland, with size of [xx] mm.' The placeholders are filled with the actual location and size of the lesion (e.g., '29 × 25 × 35 mm' in Fig. 2). This information is derived from the same lesion that defines the ground-truth mask used for evaluation. Section 3.4 then uses the text embedding to generate a coarse mask and, from it, the prompt point and bounding box. Thus PG-SAM receives target-derived privileged input, whereas all baselines in Table 2 are text-free. The comparison therefore confounds method quality with input information and does not support 'state-of-the-art performance' against those baselines. Please add controlled comparisons: (i) baselines given equivalent ground-truth-derived bounding-box/point prompts, (ii) PG-SAM with text stripped of location/size fields, and (iii)
  2. [Table 2] The external-site results do not consistently support the SOTA claim. On Site2 T1, PG-SAM scores 0.676 DSC versus 0.697 for Med-SA; on Site2 T2, 0.678 versus 0.637; on Site3 T1, 0.717 versus 0.704. The differences are small, and no standard deviations, confidence intervals, or significance tests are reported. With single-seed results, the claim of 'state-of-the-art performance across three independent clinical centers' is not statistically supported. Please report results over multiple runs and perform paired significance tests (e.g., paired bootstrap or Wilcoxon signed-rank over patients/slices).
  3. [Tables 3 and 4] The ablation results do not consistently show that the proposed modules help. In Table 3, on Site2 T1C, adding TPM to CAM decreases DSC from 0.690 to 0.635; on Site1 T1, adding CAM increases from 0.760 to 0.765 only. In Table 4, expert text versus 'None' on Site2 T2 gives 0.650 vs 0.660, and on Site3 T1 0.666 vs 0.661. These differences are within the noise of a single run. The conclusion that expert text and cross-sequence attention 'significantly improve' segmentation needs statistical support and, if appropriate, a more careful per-modality analysis.
  4. [§3.3, Eq. (2)] The cross-sequence attention module operates pixel-wise across the four MRI sequences without any description of registration or alignment. Since T1, T1C, T2, and ADC are acquired separately, they may not be voxel-aligned; Eq. (2) computes attention between spatial feature locations, which assumes a one-to-one spatial correspondence. Please state the registration/alignment procedure used in preprocessing, or validate that the sequences are co-registered in the datasets. Without this, the cross-sequence fusion results are difficult to interpret.
minor comments (5)
  1. [Eq. (9)] The Accuracy formula appears to be (TP + FN) / (TP + TN + FP + FN), which is not accuracy. It should be (TP + TN) / total. The text around the equation also mislabels TN and FN definitions.
  2. [Throughout] There are numerous typos: 'Mathmatics', 'rencent', 'breifly', 'VMLs', 'donotes', 'adpater', 'H95' vs 'HD95' in Table 2, and inconsistent use of 'Inner Site'/'Inter Site'. Please proofread carefully.
  3. [Fig. 6 caption] The caption says 'The first row to the second-last row' but should likely be 'the first row to the last row' for the methods; the ground truth is listed separately. Please clarify.
  4. [§4.1.2] No code or data availability statement is included. Given the clinical dataset, at least a code release would aid reproducibility.
  5. [Fig. 3 and §3.4] The notation x_f is defined as both concatenated embeddings and fused embeddings with the same symbol. Please use distinct symbols (e.g., v_i and v_f) to avoid confusion.

Circularity Check

1 steps flagged · score 6.0 of 10

PG-SAM's expert text carries the target lesion's location and size; the SOTA comparison against text-free baselines is confounded by privileged, ground-truth-derived input.

  1. fitted input called prediction [Section 3.4 (prompt generation module), Table 4 (prompt template), Section 4.2 (comparison)]
    "Table 4: 'This is an MRI image with a lesion located in the [pos] parotid gland, with size of [xx] mm.' Caption: ''pos' and 'xx' indicates the location and size of the lesion.' Section 3.4: 'Finally, the mean value [xmean, ymean] is computed as prompt point in the prior region, and the maximal border [xmin, ymin, xmax, ymax] is computed as prompt bbox.'"

    The placeholders [pos] and [xx] are filled with the location and dimensions of the same lesion whose segmentation is the ground-truth target. In §3.4 these text embeddings are fused with image features to generate a coarse mask, and the prompt point and bounding box are computed from that mask; these prompts condition SAM's mask decoder. Thus the model is given a target-derived spatial prior (size and location), so the Table 2 comparison against text-free baselines (UNet, Med-SA, etc.) measures privileged input, not solely the proposed architecture. The paper itself concedes the leakage hazard of extracting prompts directly from ground truth (§3.4), but the expert text is itself a ground-truth-derived description of the lesion.

full rationale

The only load-bearing circular step is the expert-text prompt: it encodes the ground-truth lesion's location and size, and the prompt generation module converts this text into the point/box that conditions the final SAM decoder. The abstract's 'state-of-the-art' claim therefore depends on a comparison in which PG-SAM receives target-derived information unavailable to the baselines. This is not a full definitional equivalence—the model still refines boundaries and the text does not specify the exact mask—but the central performance claim is partly forced by input construction, giving score 6. The cross-sequence attention module and LoRA fine-tuning are independent architectural contributions; no load-bearing self-citation or uniqueness-imported-from-authors pattern was found. The unstated co-registration assumption for multi-sequence fusion is a correctness risk, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The most consequential free parameter is the expert text template itself: it contains lesion size and location, which are measurements of the ground truth. The spatial prior mask M and the CMAttn noise variances are additional fitted priors. There are no invented physical entities. The claim therefore rests on a fairly large set of domain assumptions about data alignment, label quality, and the availability and legitimacy of the text input.

free parameters (6)
  • Spatial prior mask M = not reported
    Computed by counting lesion locations in the training set (Section 3.4) and used to constrain all generated prompts; it is a data-derived prior.
  • Gaussian noise variance var(i) in CMAttn = not reported
    Formulated offline and stored (Eq. 4); derived from category sample sizes and thus fitted to training class balance.
  • Loss weights w and beta = not reported
    Used in Eq. 6 to balance the three losses, but their values are never given, so the training objective is underspecified.
  • LoRA rank r = 5
    Chosen hyperparameter for image encoder fine-tuning, adopted from SAMed.
  • Slice inclusion threshold = lesion pixel count > 25
    Slices with fewer than 25 lesion pixels are excluded from training and evaluation, a post hoc data-selection rule that affects the reported metrics.
  • Expert text template = "This is an MRI image with a lesion located in the [pos] parotid gland, with size of [xx] mm."
    The exact prompt template is authored by the investigators; it encodes location and size, which are target-derived measurements.
assumptions (5)
  • domain assumption Multi-sequence MRI images (T1, T1C, T2, ADC) are pixel-wise co-registered.
    The cross-sequence attention module (Eq. 2) fuses features across sequences as if they are spatially aligned, but no registration procedure is described in Section 3.3 or Section 4.1.
  • domain assumption Ground-truth lesion masks produced by two radiologists are correct and consistent.
    All supervised losses (Eq. 6) and all evaluation metrics depend on the expert consensus labels; the paper does not report inter-observer variability.
  • domain assumption SAM's pre-trained weights transfer to medical MRI after LoRA fine-tuning.
    The method freezes SAM's encoder and inserts LoRA, assuming the natural-image features are a useful starting point for parotid MRI; this is the standard but unproven transfer assumption of the paper.
  • domain assumption MedCLIP text embeddings, after adapter fine-tuning, capture clinically relevant spatial information (location and size) from the expert text.
    Section 3.4 relies on MedCLIP to map diagnostic phrases to embeddings from which the coarse mask is generated; no analysis of the text encoder's fidelity is provided.
  • domain assumption Expert diagnostic text is available for every test case, including external sites.
    The method's test-time input includes the text prompt, but Section 4.1 does not state how or whether written reports were available for the external Site2 and Site3 patients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Sequence Parotid Gland Lesion Segmentation via Expert Text-Guided Segment Anything Model." pith.science (2026). https://pith.science/paper/GUJ7MASS

@misc{pith2026250809645,
  author       = {Pith},
  title        = {Pith review of: Multi-Sequence Parotid Gland Lesion Segmentation via Expert Text-Guided Segment Anything Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUJ7MASS}},
  note         = {Machine review of arXiv:2508.09645}
}
read the original abstract

Parotid gland lesion segmentation is essential for the treatment of parotid gland diseases. However, due to the variable size and complex lesion boundaries, accurate parotid gland lesion segmentation remains challenging. Recently, the Segment Anything Model (SAM) fine-tuning has shown remarkable performance in the field of medical image segmentation. Nevertheless, SAM's interaction segmentation model relies heavily on precise lesion prompts (points, boxes, masks, etc.), which are very difficult to obtain in real-world applications. Besides, current medical image segmentation methods are automatically generated, ignoring the domain knowledge of medical experts when performing segmentation. To address these limitations, we propose the parotid gland segment anything model (PG-SAM), an expert diagnosis text-guided SAM incorporating expert domain knowledge for cross-sequence parotid gland lesion segmentation. Specifically, we first propose an expert diagnosis report guided prompt generation module that can automatically generate prompt information containing the prior domain knowledge to guide the subsequent lesion segmentation process. Then, we introduce a cross-sequence attention module, which integrates the complementary information of different modalities to enhance the segmentation effect. Finally, the multi-sequence image features and generated prompts are feed into the decoder to get segmentation result. Experimental results demonstrate that PG-SAM achieves state-of-the-art performance in parotid gland lesion segmentation across three independent clinical centers, validating its clinical applicability and the effectiveness of diagnostic text for enhancing image segmentation in real-world clinical settings.

Figures

Figures reproduced from arXiv: 2508.09645 by the authors.

Figure 1
Figure 1. Visualization of typical samples. Samples reflecting the three main challenges, namely, (a) multi-sequence [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Comparison of SAM learning with different prompts. (a) classical interactive learning methods relying on [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the proposed PG-SAM framework. The overall network contains a text guide decoder, an [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The architecture of the LoRA in SAM. LoRA layers are inserted into the transformer blocks of SAM’s [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Architecture of the Learnable Mask Cross-Attention and Class-Balanced Mask-Guided Self-Attention. (a) [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Segmentation prediction plots of different network models on Site1. The first row to the second-last row [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Segmentation prediction plots of different network models on Site2. The first row to the second-last row [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Segmentation prediction plots of different network models on Site3. The first row to the second-last row [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Segmentation performance of different training sample ratios across three sites. The horizontal axis [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 44 canonical work pages

  1. [1]

    M.-Q. Jia, M. Gao, P. Ye, Survival outcome of salivary gland carcinoma: a 50-year retrospective study with long-term follow-up, Journal of Oral and Maxillofacial Surgery 80 (12) (2022) 2003–2014

  2. [2]

    M. Quer, J. C. Hernandez-Prera, C. Silver, Current trends and controversies in the management of warthin tumor of the parotid gland, Diagnostics 11 (8) (2021) 1467

  3. [3]

    G.-X. Xu, C. Liu, J. Liu, Cross-site severity assessment of covid-19 from ct images via domain adaptation, IEEE Transactions on Medical Imaging 41 (1) (2022) 88–102

  4. [4]

    Y. Yang, H. Fu, A. I. Aviles-Rivero, Diffmic-v2: Medical image classification via improved diffusion network, IEEE Transactions on Medical Imaging 44 (5) (2025) 2244–2255

  5. [5]

    Ren, G.-X

    C.-X. Ren, G.-X. Xu, D.-Q. Dai, Cross-site prognosis prediction for nasopharyngeal carcinoma from incomplete multi-modal data, Medical Image Analysis 93 (2024) 103103

  6. [6]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, Segment anything, in: Proceedings of the IEEE/CVF International Conference on Computer Bision, 2023, pp. 4015–4026

  7. [7]

    Huang, X

    Y. Huang, X. Yang, L. Liu, Segment anything model for medical images?, Medical Image Analysis 92 (2024) 103061

  8. [8]

    S. N. Gowda, D. A. Clifton, Cc-sam: Sam with cross-feature attention and context for ultrasound image segmentation, in: European Conference on Computer Vision, 2024, pp. 108–124

Show all 49 references
  1. [9]

    Y. Sun, J. Chen, S. Zhang, Vrp-sam: Sam with visual reference prompt, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23565–23574

  2. [10]

    S. Kato, H. Mitsuoka, K. Hotta, Generalized sam: Efficient fine-tuning of sam for variable input image sizes, in: European Conference on Computer Vision, 2024, pp. 167–182

  3. [11]

    Z. Peng, Z. Xu, Z. Zeng, Sam-parser: Fine-tuning sam efficiently by parameter space reconstruction, in: Pro- ceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, 2024, pp. 4515–4523. 23

  4. [12]

    T. Chen, L. Zhu, C. Deng, Sam-adapter: Adapting segment anything in underperformed scenes, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 3367–3375

  5. [13]

    L. Ke, M. Ye, M. Danelljan, Segment anything in high quality, Advances in Neural Information Processing Systems 36 (2023) 29914–29934

  6. [14]

    Zhang, D

    K. Zhang, D. Liu, Customized segment anything model for medical image segmentation, arXiv preprint arXiv:2304.13785 (2023)

  7. [15]

    C. Chen, J. Miao, D. Wu, Ma-sam: Modality-agnostic sam adaptation for 3d medical image segmentation, Medical Image Analysis 98 (2024) 103310

  8. [16]

    C. Zhou, K. Ning, Q. Shen, Sam-sp: Self-prompting makes sam great again, arXiv preprint arXiv:2408.12364 (2024)

  9. [17]

    R. Xiao, S. Kim, M.-I. Georgescu, Flair: Vlm with fine-grained language-informed image representations, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 24884–24894

  10. [18]

    R. Yu, W. Yu, X. Wang, Attention prompting on image for large vision-language models, in: European Confer- ence on Computer Vision, 2024, pp. 251–268

  11. [19]

    Q. Lei, B. Wang, R. Tan, Ez-hoi: Vlm adaptation via guided prompt learning for zero-shot hoi detection, Advances in Neural Information Processing Systems 37 (2024) 55831–55857

  12. [20]

    Pham, T.-D

    T.-H. Pham, T.-D. Bui, M. L. Quang, Silvar-med: A speech-driven visual language model for explainable abnormality detection in medical imaging, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 2984–2994

  13. [21]

    Y. Zou, Z. Yin, Alignment, mining and fusion: Representation alignment with hard negative mining and selective knowledge fusion for medical visual question answering, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 29623–29633

  14. [22]

    V. Nath, W. Li, D. Yang, Vila-m3: Enhancing vision-language models with medical expert knowledge, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 14788–14798

  15. [23]

    Achiam, S

    J. Achiam, S. Adler, S. Agarwal, Gpt-4 technical report, arXiv preprint arXiv:2303.08774 (2023)

  16. [24]

    A. Liu, B. Feng, B. Xue, Deepseek-v3 technical report, arXiv preprint arXiv:2412.19437 (2024)

  17. [25]

    K. Yan, Q. Cai, F. Zhang, Sgtc: Semantic-guided triplet co-training for sparsely annotated semi-supervised medical image segmentation, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, pp. 9112–9120

  18. [26]

    Xu, C.-X

    G.-X. Xu, C.-X. Ren, Y. Sun, Domain knowledge-driven encoder–decoder for nasopharyngeal carcinoma seg- mentation, Expert Systems with Applications 258 (2024) 125208

  19. [27]

    Z. Wang, Z. Wu, D. Agarwal, Medclip: Contrastive learning from unpaired medical images and text, in: Pro- ceedings of the Conference on Empirical Methods in Natural Language Processing. Conference on Empirical Methods in Natural Language Processing, Vol. 2022, 2022, p. 3876

  20. [28]

    J. Ma, Y. He, F. Li, Segment anything in medical images, Nature Communications 15 (2024) 654

  21. [29]

    J. Wu, Z. Wang, M. Hong, Medical sam adapter: Adapting segment anything model for medical image segmen- tation, Medical Image Analysis 102 (2025) 103547

  22. [30]

    E. J. Hu, Y. Shen, P. Wallis, Lora: Low-rank adaptation of large language models., International Conference 24 on Learning Representations 1 (2) (2022) 3

  23. [31]

    Houlsby, A

    N. Houlsby, A. Giurgiu, S. Jastrzebski, Parameter-efficient transfer learning for nlp, in: International Conference on Machine Learning, 2019, pp. 2790–2799

  24. [32]

    Cheng, Q

    Z. Cheng, Q. Wei, H. Zhu, Unleashing the potential of sam for medical adaptation via hierarchical decoding, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3511–3522

  25. [33]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, Learning transferable visual models from natural language supervision, in: International Conference on Machine Learning, 2021, pp. 8748–8763

  26. [34]

    S. Liu, Z. Zeng, T. Ren, Grounding dino: Marrying dino with grounded pre-training for open-set object detection (2024) 38–55

  27. [35]

    Eslami, C

    S. Eslami, C. Meinel, G. De Melo, Pubmedclip: How much does clip benefit visual question answering in the medical domain?, in: Findings of the Association for Computational Linguistics: EACL 2023, 2023, pp. 1181–1193

  28. [36]

    Zhang, P

    B. Zhang, P. Zhang, X. Dong, Long-clip: Unlocking the long-text capability of clip, in: European Conference on Computer Vision, Springer, 2024, pp. 310–325

  29. [37]

    K. Yang, T. Gu, X. An, Clip-cid: Efficient clip distillation via cluster-instance discrimination, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, 2025, pp. 21974–21982

  30. [38]

    L. Zhu, X. Wang, J. Feng, Weakclip: Adapting clip for weakly-supervised semantic segmentation, International Journal of Computer Vision 133 (3) (2025) 1085–1105

  31. [39]

    Huang, L

    S.-C. Huang, L. Shen, M. P. Lungren, Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3942–3951

  32. [40]

    Wenting, L

    C. Wenting, L. Jie, Y. Yixuan, Bi-vlgm: Bi-level class-severity-aware vision-language graph matching for text guided medical image segmentation, arXiv preprint arXiv:2305.12231 (2023)

  33. [41]

    Zhang, D

    C. Zhang, D. Han, Y. Qiao, Faster segment anything: Towards lightweight sam for mobile applications, arXiv preprint arXiv:2306.14289 (2023)

  34. [42]

    Xiong, B

    Y. Xiong, B. Varadarajan, L. Wu, Efficientsam: Leveraged masked image pretraining for efficient segment anything, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16111–16121

  35. [43]

    Li, Y.-m

    M. Li, Y.-m. Cheung, Y. Lu, Long-tailed visual recognition via gaussian clouded logit adjustment, in: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 6929–6938

  36. [44]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, 2015, pp. 234–241

  37. [45]

    Z. Zhou, M. M. R. Siddiquee, N. Tajbakhsh, Unet++: Redesigning skip connections to exploit multiscale features in image segmentation, IEEE Transactions on Medical Imaging (2019)

  38. [46]

    J. M. J. Valanarasu, V. M. Patel, Unext: Mlp-based rapid medical image segmentation network, in: International Conference on Medical Image Computing and Computer-Assisted Intervention, 2022, pp. 23–33

  39. [47]

    H. Cao, Y. Wang, J. Chen, Swin-unet: Unet-like pure transformer for medical image segmentation, in: European Conference on Computer Vision, 2022, pp. 205–218. 25

  40. [48]

    J. Chen, Y. Lu, Q. Yu, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306 (2021)

  41. [49]

    M. M. Rahman, M. Munir, R. Marculescu, Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 11769–11779. 26

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.