Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Pre-Trained LLM is a Semantic-Aware and Generalizable Segmentation Booster

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that inserting a single frozen layer from a pre-trained large language model into a CNN-based medical image segmenter improves segmentation across ultrasound, dermoscopy, polypscopy, and CT, with only small linear…

desk verdict A modest empirical trick that deserves a referee but not the 'semantic-aware' label: the random-transformer control alone cannot support language-pretraining causation. read the letter →

arxiv 2506.18034 v1 pith:G5FHCQ3U submitted 2025-06-22 cs.CV cs.AIcs.MM

Fenghe Tang , Wenxin Ma , Zhiyang He , Xiaodong Tao , Zihang Jiang , S. Kevin Zhou This is my paper
classification cs.CVcs.AIcs.MM
keywords medicalimagesegmentationlargelanguagemodelsfrozenLLMlayerhybridCNN-transformersemanticawarenesstransferlearningglobalcontext
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

This paper claims that a single frozen layer from a pre-trained large language model can act as a plug-in global-context module for medical image segmentation. The authors insert one LLaMA or DeepSeek transformer layer, kept frozen, between a CNN encoder and decoder, with only small trainable linear projections on either side. On five datasets spanning ultrasound, dermoscopy, polypscopy, and CT, the boosted models generally beat their CNN baselines and several published segmenters while adding only about 0.8 to 4.2 million trainable parameters. The authors interpret the gain as a transfer of semantic awareness from language pretraining, visible as sharper, less noisy feature activations. If this holds, it offers a low-cost route to global context for CNN segmenters in label-scarce medical settings.

What carries the argument

The load-bearing object is the frozen LLM transformer layer inserted between encoder and decoder, flanked by two trainable linear projections. The projections map flattened CNN feature maps into the LLM's token embedding dimension and back again, so the frozen block behaves as a fixed, high-capacity global mixer whose weights were shaped by language pretraining. The paper's analysis identifies the layer's effect as semantic refinement: it suppresses background activation noise and sharpens foreground boundaries, while SVD-based statistics show a broader singular-value spectrum after the layer. The selection of a sufficiently deep layer (layer 15 of LLaMA3.2-1B, layer 28 of DeepSeek-R1-Distill-Qwen-1.5B) matters, but the benefit is reported across both models and across 2D and 3D inputs.

What would settle it

Retrain the UNet+LLaMA and CMUNeXt+LLaMA models and their baselines on the same five public datasets across at least five random seeds, and check whether the frozen-LLM advantage in intersection-over-union exceeds the seed-to-seed spread; if it does not, the improvement claim fails. A matched control in which the frozen LLM layer is replaced by a frozen vision-pretrained transformer of the same size would test whether the effect is language-specific.

Watch

Extended reading notes

Core claim

The central discovery is that a frozen, pre-trained LLM layer can process visual tokens for segmentation without any language input, and that doing so generally improves segmentation quality. The paper shows this by swapping the randomly initialized transformer block in a hybrid CNN-transformer segmenter for a frozen LLaMA3.2-1B or DeepSeek-R1-Distill-Qwen-1.5B layer, wrapped in linear projections. Across BUSI, TNSCUI, ISIC, Kvasir, and BTCV, the frozen-LLM variants outperform their CNN baselines and match or exceed prior hybrid methods, and activation, rank, and singular-value analyses indicate that the LLM layer concentrates features on foreground regions and enlarges the effective rank of the feature space. Because the LLM weights are fixed, the result is attributed to pretrained semantic knowledge rather than to the extra parameters.

Load-bearing premise

The load-bearing premise is that the reported accuracy advantages of the frozen LLM layer, several of which are under one percentage point and all measured on a single data split with no error bars, are larger than the normal variation between training runs.

Editorial extensions

If this is right

  • Any CNN encoder-decoder segmenter can be upgraded by inserting one frozen LLM layer and two trainable linear projections, so the added training burden is minimal.
  • The boost appears on 2D modalities (ultrasound, dermoscopy, polypscopy) and 3D CT, so the effect is not limited to a single imaging domain.
  • Because a randomly initialized transformer of the same shape does not reproduce the gain, the improvement is attributed to the pre-trained weights rather than to added parameters or to the transformer structure alone.
  • Using LLaMA or DeepSeek layers both helps, and any sufficiently deep layer works, indicating the effect is robust to which LLM supplies the frozen block.

Reading between the lines

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

  • Editorial inference: if the gain survives seed-level retesting, the same frozen-LLM insert could be reused for detection, classification, or anomaly localization at almost no extra training cost, because only the surrounding projections need to be fit.
  • Editorial inference: the paper's experiments do not separate the effect of language pretraining from the effect of any strong transformer pretraining; a frozen vision-pretrained transformer control would pin down the mechanism.
  • Editorial inference: a natural next experiment is sweeping the insertion position and the number of frozen LLM layers, since the paper only varies which layer of the LLM is used, not where the block sits inside the encoder-decoder.
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 / 5 minor

Summary. The paper proposes LLM4Seg, a hybrid CNN encoder-decoder segmentation model in which the usual transformer block is replaced by a single frozen layer from a pretrained LLM (LLaMA3.2-1B or DeepSeek-R1-Distill-Qwen-1.5B), wrapped by trainable linear projections. Experiments on four 2D datasets (BUSI, TNSCUI, ISIC, Kvasir) and one 3D dataset (BTCV) report improved IoU/F1/Dice over the corresponding CNN baselines and over a randomly initialized transformer control, with only a small increase in trainable parameters. The authors interpret the gains as evidence that language-pretrained representations transfer 'semantic awareness' to visual segmentation, and they support this interpretation with activation visualizations, activation-concentration IoU, and singular-value statistics.

Significance. If the central claim held, the result would be practically interesting: it offers a parameter-efficient way to inject global context into CNN-based medical segmentation by reusing an off-the-shelf frozen LLM block, and it is validated across two LLMs and multiple modalities. The paper includes useful controls: a randomly initialized transformer with identical structure, trainable versions of the LLM layer, and computational cost reporting. The main deficiency is that the random-transformer control cannot isolate LLM pretraining from generic pretrained initialization; a frozen pretrained vision transformer control is needed to support the 'semantic-aware' attribution. The absence of error bars and the use of a single split also make several sub-IoU gains unreliable. With additional controls and statistical reporting, the empirical contribution would be solid; the explanatory claim currently outruns the evidence.

major comments (3)
  1. [3.2, Table 2] The only non-LLM transformer baseline is '+Transformer', a randomly initialized transformer with the same structure. This control appropriately rules out parameter-count and architectural effects, but it does not rule out the possibility that any frozen pretrained transformer block, regardless of pretraining modality, provides a better-conditioned feature mixer. The activation and singular-value analyses in Sections 4.1-4.3 compare against the same random baseline and therefore inherit the same confound. To support the abstract and title claim that the effect comes from LLM 'semantic awareness,' the authors should compare with a frozen pretrained vision transformer (e.g., a ViT/MAE/DINO or Swin block of comparable size) inserted at the same location. If such a control matches LLaMA, the empirical booster result remains, but the language-pretraining attribution would need to be removed or substantially softened.
  2. [3.1-3.2, Tables 1 and 2] All quantitative results are point estimates from a single 7/3 split, with no seeds, error bars, or significance tests. Several reported gains are small: CMUNeXt+LLaMA improves on ISIC by 0.22 IoU (82.69 vs 82.47), and nnUNet+LLaMA reduces Kvasir IoU by 0.37 (84.60 vs 84.97), despite the text claiming that performance 'consistently improves.' These differences are likely within run-to-run noise for this regime. Please report results over at least three seeds or provide significance tests to support the general-improvement claim.
  3. [3.1 and 4.3] The frozen LLM layer index (15 for LLaMA, 28 for DeepSeek) is described as 'empirically employed' on the same benchmarks that produce the headline numbers. If the layer index was chosen by looking at validation/test performance on these datasets, the reported improvements include selection bias. The paper should state the selection protocol (e.g., a separate tuning set or an a priori rule) and report sensitivity of Tables 1 and 2 to layer index within the range shown in Fig. 3. Likewise, the 0.4 activation threshold used in Fig. 2 should be justified a priori rather than chosen post hoc.
minor comments (5)
  1. [3.1] The text says '11 recent SOTA models' but then lists 10 named architectures; please reconcile the count.
  2. [Abstract and Introduction] There are several typos: 'polypscopy' should likely be 'polyp colonoscopy' or 'polyp segmentation,' 'developping' should be 'developing,' and Section 4.3 contains 'knowlesge' instead of 'knowledge.'
  3. [2, Eq. (4)] The symbol \hat{t} is reused for the outputs of the first Linear, the Transformer, and the second Linear; please use distinct variables for clarity.
  4. [4.2] The channel-wise SVD description is imprecise: 'obtaining C singular matrix \Sigma=[\sigma_1,...,\sigma_k] \in R^{H\times W}' is not a standard way to write C decompositions; please clarify how the Effective Rank and LSVR are computed and aggregated across channels.
  5. [Table 2] The table formatting appears to omit row labels for the 3D baselines (MedNeXt and 3D UX-Net) and does not clearly indicate which backbone the DeepSeek rows apply to; please reformat for readability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claim is an empirical benchmark comparison against external datasets, not a derivation that reduces to its inputs.

full rationale

The paper makes no formal derivation; its central claim is that a frozen pretrained LLM layer, wrapped in linear projections, improves medical image segmentation across several external benchmarks (BUSI, TNSCUI, ISIC, Kvasir, BTCV). The LLM weights are external artifacts (LLaMA3.2-1B, DeepSeek-R1-Distill-Qwen-1.5B), and improvements are measured against publicly available datasets and baselines. The key control in Table 2, '+Transformer', is a randomly initialized transformer with identical structure, which addresses the parameter-count and architecture confound; although it does not rule out generic pretrained-initialization effects, a missing control is an experimental confound, not a circular step. The only mild concern is that the frozen layer index (15 for LLaMA, 28 for DeepSeek) is empirically selected on the same benchmarks, as stated in Sec. 3.1: 'By default, we empirically employ the frozen 15-th layer of LLaMA3.2-1B [11] or 28-th DeepSeek-R1-Distill-Qwen-1.5B [12] as the LLM layer.' But Fig. 3 reports consistent gains across a range of layers, so the headline result is not forced by the selection. Self-citations (e.g., CMUNeXt as a backbone) are used as baselines and are not load-bearing for the claim that the frozen LLM layer transfers semantic awareness. The mechanistic claim is interpretive, not derived from an equation that assumes the conclusion. Conclusion: no step exhibits the paper's own equations or citation chain reducing to its own inputs; any residual concerns are about statistical robustness and confound control, not circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper relies on standard deep learning machinery (CNN encoders, transformer layers, linear projections) and on the untested interpretive premise that language-pretrained semantic knowledge transfers to visual token processing. The only hand-chosen quantities are the frozen layer index and an activation threshold used in the analysis. There are no invented entities.

free parameters (2)
  • LLM layer index = 15 for LLaMA3.2-1B, 28 for DeepSeek-R1-Distill-Qwen-1.5B
    Selected empirically in Sec. 3.1 ('by default, we empirically employ the frozen 15-th layer'), and Sec. 4.3 shows performance depends on depth; this is a tuned hyperparameter.
  • Activation threshold = 0.4
    Used in Sec. 4.1 to threshold activations before computing IoU with ground truth; not derived, hand-chosen for the concentration analysis.
assumptions (4)
  • standard math SVD, effective rank, and largest singular value ratio faithfully reflect representation quality of segmentation features.
    Sec. 4.2 uses these spectral statistics to support the semantic-refinement claim; this is a standard but nontrivial interpretive step.
  • domain assumption Large language models pretrained on text encode semantic priors that transfer to visual tokens.
    This is the mechanism proposed in Sec. 1 and Sec. 5; it is not independently proven and is not isolated from generic frozen-transformer benefits.
  • domain assumption The benchmark datasets and metrics (IoU, F1, Dice) are accepted proxies for segmentation quality.
    Sec. 3.1 adopts prior evaluation conventions; the conclusion inherits whatever biases those benchmarks carry.
  • ad hoc to paper A single frozen layer from LLaMA or DeepSeek can be inserted into a CNN pipeline without catastrophic distribution shift.
    The feasibility of Linear to LLM block to Linear within a U-Net is assumed in Eq. (4) and validated only empirically on the chosen datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pre-Trained LLM is a Semantic-Aware and Generalizable Segmentation Booster." pith.science (2026). https://pith.science/paper/G5FHCQ3U

@misc{pith2026250618034,
  author       = {Pith},
  title        = {Pith review of: Pre-Trained LLM is a Semantic-Aware and Generalizable Segmentation Booster},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G5FHCQ3U}},
  note         = {Machine review of arXiv:2506.18034}
}
read the original abstract

With the advancement of Large Language Model (LLM) for natural language processing, this paper presents an intriguing finding: a frozen pre-trained LLM layer can process visual tokens for medical image segmentation tasks. Specifically, we propose a simple hybrid structure that integrates a pre-trained, frozen LLM layer within the CNN encoder-decoder segmentation framework (LLM4Seg). Surprisingly, this design improves segmentation performance with a minimal increase in trainable parameters across various modalities, including ultrasound, dermoscopy, polypscopy, and CT scans. Our in-depth analysis reveals the potential of transferring LLM's semantic awareness to enhance segmentation tasks, offering both improved global understanding and better local modeling capabilities. The improvement proves robust across different LLMs, validated using LLaMA and DeepSeek.

Figures

Figures reproduced from arXiv: 2506.18034 by the authors.

Figure 1
Figure 1. Comparison of current VLMs, hybrid segmentation models, and our novel hy￾brid segmentation model. Our LLM4Seg use a frozen LLM layer within CNN encoder￾decoder framework to boost global visual understanding. Abstract. With the advancement of Large Language Model (LLM) for natural language processing, this paper presents an intriguing finding: a frozen pre-trained LLM layer can process visual tokens for medical im￾ag… view at source ↗
Figure 2
Figure 2. (a) Activation visualization before and after the Transformer/LLaMA layer. (b) Concentration accuracy of activations. The activations are threshold by 0.4 and compared with ground truth segmentation masks to calculate IoU. (c) The distribu￾tion of Effective Rank (ER) and Largest Singular Value Ratio (LSVR) of activations extracted from TNSCUI dataset. Average value of each distribution is denoted by →. 4.2 Statistic… view at source ↗
Figure 3
Figure 3. Impact of different LLM Transformer layers on segmentation performance. 5 Conclusion and Discussion Our findings reveal a novel and unexpected generalization capability of LLMs: their semantic awareness in medical image segmentation. We propose to inte￾grate a frozen pre-trained LLM layer into a CNN architecture, demonstrating improved segmentation performance across various imaging modalities. Because the encoder i… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. DiffAttn: Diffusion-Based Drivers' Visual Attention Prediction with LLM-Enhanced Semantic Reasoning

    cs.CV 2026-03 unverdicted novelty 6.0 of 10

    DiffAttn formulates driver visual attention prediction as a conditional diffusion-denoising task with Swin Transformer encoding, multi-scale fusion, and LLM semantic reasoning, achieving SoTA results on four datasets.

  2. Large Language Model Evaluated Stand-alone Attention-Assisted Graph Neural Network with Spatial and Structural Information Interaction for Precise Endoscopic Image Segmentation

    cs.CV 2025-08 reject novelty 4.0 of 10

    FOCUS-Med reports state-of-the-art polyp segmentation scores by fusing graph, attention, and multi-scale fusion modules, but missing baseline details and an absent appendix undermine the claim.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [22]

    arXiv preprint arXiv:2310.12973 (2023)

    Pang, Z., Xie, Z., Man, Y., Wang, Y.X.: Frozen transformers in language models are effective visual encoder layers. arXiv preprint arXiv:2310.12973 (2023)

    arXiv 2023

  2. [1]

    Data in brief28, 104863 (2020)

    Al-Dhabyani, W., Gomaa, M., Khaled, H., Fahmy, A.: Dataset of breast ultrasound images. Data in brief28, 104863 (2020)

    2020

  3. [2]

    In: ECCV

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: ECCV. pp. 205–218. Springer (2022)

    work page 2022

  4. [3]

    arXiv preprint arXiv:2309.17002 (2023)

    Chen, H., Wang, J., Shah, A., Tao, R., Wei, H., Xie, X., Sugiyama, M., Raj, B.: Understanding and mitigating the label noise in pre-training on downstream tasks. arXiv preprint arXiv:2309.17002 (2023)

    arXiv 2023

  5. [4]

    arXiv preprint arXiv:2102.04306 (2021)

    Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.:Transunet:Transformersmakestrongencodersformedicalimagesegmentation. arXiv preprint arXiv:2102.04306 (2021)

    arXiv 2021

  6. [5]

    In: MICCAI

    Chen, J., Chen, R., Wang, W., Cheng, J., Zhang, L., Chen, L.: Tinyu-net: Lighter yet better u-net with cascaded multi-receptive fields. In: MICCAI. pp. 626–635. Springer (2024)

    2024

  7. [6]

    discriminability: Batch spectral penalization for adversarial domain adaptation

    Chen, X., Wang, S., Long, M., Wang, J.: Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In: ICML. pp. 1081–1090. PMLR (2019)

    work page 2019

  8. [7]

    In: CVPR

    Cheng, B., Misra, I., Schwing, A.G., Kirillov, A., Girdhar, R.: Masked-attention mask transformer for universal image segmentation. In: CVPR. pp. 1290–1299 (2022)

    work page 2022

Show all 40 references
  1. [8]

    arXiv preprint arXiv:1902.03368 (2019)

    Codella, N., Rotemberg, V., Tschandl, P., Celebi, M.E., Dusza, S., Gutman, D., Helba,B.,Kalloo,A.,Liopyris,K.,Marchetti,M.,et al.:Skinlesionanalysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:...

    2019 arXiv

  2. [9]

    In: CVPR

    Ding, X., Zhang, Y., Ge, Y., Zhao, S., Song, L., Yue, X., Shan, Y.: Unireplknet: A universal perception large-kernel convnet for audio video point cloud time-series and image recognition. In: CVPR. pp. 5513–5524 (2024)

    2024

  3. [10]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A.: An image is worth 16x16 words: Transformers for image recogni- tion at scale. arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  4. [11]

    arXiv preprint arXiv:2407.21783 (2024)

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  5. [12]

    arXiv preprint arXiv:2501.12948 (2025) 10 Fenghe Tang, Wenxin Ma, et al

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al.: Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025) 10 Fenghe Tang, Wenxin Ma, et al

    2025 arXiv

  6. [13]

    In: WACV

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: Unetr: Transformers for 3d medical image segmentation. In: WACV. pp. 574–584 (2022)

    2022

  7. [14]

    TMI42(5), 1484–1494 (2022)

    Huang, X., Deng, Z., Li, D., Yuan, X., Fu, Y.: Missformer: An effective transformer for 2d medical image segmentation. TMI42(5), 1484–1494 (2022)

    2022

  8. [15]

    Nature methods18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods18(2), 203–211 (2021)

    2021

  9. [16]

    In: Proc

    Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi- atlas labeling beyond the cranial vault–workshop and challenge. In: Proc. MICCAI Challenge. vol. 5, p. 12 (2015)

    2015

  10. [17]

    In: ICLR

    Lee, H.H., Bao, S., Huo, Y., Landman, B.A.: 3d ux-net: A large kernel volumetric convnet modernizing hierarchical transformer for medical image segmentation. In: ICLR

  11. [18]

    NeurIPS36(2024)

    Li, C., Wong, C., Zhang, S., Usuyama, N., Liu, H., Yang, J., Naumann, T., Poon, H., Gao, J.: Llava-med: Training a large language-and-vision assistant for biomedicine in one day. NeurIPS36(2024)

    2024

  12. [19]

    In: ICML

    Li, J., Li, D., Savarese, S., Hoi, S.: Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: ICML. pp. 19730–19742. PMLR (2023)

    2023

  13. [20]

    TIM71, 1–15 (2022)

    Lin, A., Chen, B., Xu, J., Zhang, Z., Lu, G., Zhang, D.: Ds-transunet: Dual swin transformer u-net for medical image segmentation. TIM71, 1–15 (2022)

    2022

  14. [21]

    arXiv preprint arXiv:2501.12295 (2025)

    Ma, W., Yao, Q., Zhang, X., Huang, Z., Jiang, Z., Zhou, S.K.: Towards accurate unified anomaly segmentation. arXiv preprint arXiv:2501.12295 (2025)

    2025 arXiv

  15. [23]

    In: CVPR

    Parkhi, O.M., Vedaldi, A., Zisserman, A., Jawahar, C.: Cats and dogs. In: CVPR. pp. 3498–3505. IEEE (2012)

    2012

  16. [24]

    In: 10th International symposium on medical information processing and analysis

    Pedraza, L., Vargas, C., Narváez, F., Durán, O., Muñoz, E., Romero, E.: An open access thyroid ultrasound image database. In: 10th International symposium on medical information processing and analysis. vol. 9287, pp. 188–193. SPIE (2015)

    2015

  17. [25]

    In: ACMMM

    Pogorelov, K., Randel, K.R., Griwodz, C., Eskeland, S.L., de Lange, T., Johansen, D., Spampinato, C., Dang-Nguyen, D.T., Lux, M., Schmidt, P.T., et al.: Kvasir: A multi-class image dataset for computer aided gastrointestinal disease detection. In: ACMMM. pp. 164–169 (2017)

    2017

  18. [26]

    In: MICCAI

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: MICCAI. pp. 234–241. Springer (2015)

    2015

  19. [27]

    In: 2007 15th European signal processing conference

    Roy, O., Vetterli, M.: The effective rank: A measure of effective dimensionality. In: 2007 15th European signal processing conference. pp. 606–610. IEEE (2007)

    2007

  20. [28]

    In: MICCAI

    Roy, S., Koehler, G., Ulrich, C., Baumgartner, M., Petersen, J., Isensee, F., Jaeger, P.F., Maier-Hein, K.H.: Mednext: transformer-driven scaling of convnets for med- ical image segmentation. In: MICCAI. pp. 405–415. Springer (2023)

    2023

  21. [29]

    Tang, F., Ding, J., Quan, Q., Wang, L., Ning, C., Zhou, S.K.: Cmunext: An efficient medicalimagesegmentationnetworkbasedonlargekernelandskipfusion.In:ISBI. pp. 1–5. IEEE (2024)

    2024

  22. [30]

    arXiv preprint arXiv:2312.01740 (2023)

    Tang, F., Nian, B., Ding, J., Quan, Q., Yang, J., Liu, W., Zhou, S.K.: Mobileutr: Revisiting the relationship between light-weight cnn and transformer for efficient medical image segmentation. arXiv preprint arXiv:2312.01740 (2023)

    2023 arXiv

  23. [31]

    Medical Image Analysis p

    Tang, F., Nian, B., Li, Y., Jiang, Z., Yang, J., Liu, W., Zhou, S.K.: Mambamim: Pre-training mamba with state space token interpolation and its application to medical image segmentation. Medical Image Analysis p. 103606 (2025) LLM4Seg 11

    2025

  24. [32]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Tang, F., Xu, R., Yao, Q., Fu, X., Quan, Q., Zhu, H., Liu, Z., Zhou, S.K.: Hyspark: Hybrid sparse masking for large scale medical image pre-training. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 330–340. Springer (2024)

    2024

  25. [33]

    In: MICCAI

    Tang, F., Xu, R., Yao, Q., Fu, X., Quan, Q., Zhu, H., Liu, Z., Zhou, S.K.: Hyspark: Hybrid sparse masking for large scale medical image pre-training. In: MICCAI. pp. 330–340. Springer (2024)

    2024

  26. [34]

    arXiv preprint arXiv:2502.08347 (2025)

    Tang, F., Yao, Q., Ma, W., Wu, C., Jiang, Z., Zhou, S.K.: Hi-end-mae: Hierarchical encoder-driven masked autoencoders are stronger vision learners for medical image segmentation. arXiv preprint arXiv:2502.08347 (2025)

    2025 arXiv

  27. [35]

    In: CVPR

    Tang, Y., Yang, D., Li, W., Roth, H.R., Landman, B., Xu, D., Nath, V., Hatamizadeh, A.: Self-supervised pre-training of swin transformers for 3d medi- cal image analysis. In: CVPR. pp. 20730–20740 (2022)

    2022

  28. [36]

    In: MICCAI

    Valanarasu, J.M.J., Patel, V.M.: Unext: Mlp-based rapid medical image segmen- tation network. In: MICCAI. pp. 23–33. Springer (2022)

    2022

  29. [37]

    In: AAAI

    Wang, H., Cao, P., Wang, J., Zaiane, O.R.: Uctransnet: rethinking the skip connec- tions in u-net from a channel-wise perspective with transformer. In: AAAI. vol. 36, pp. 2441–2449 (2022)

    2022

  30. [38]

    In: MICCAI

    Wenxuan, W., Chen, C., Meng, D., Hong, Y., Sen, Z., Jiangyun, L.: Transbts: Multimodal brain tumor segmentation using transformer. In: MICCAI. pp. 109– 119 (2021)

    2021

  31. [39]

    In: ICML

    Xue, Y., Whitecross, K., Mirzasoleiman, B.: Investigating why contrastive learning benefits robustness against label noise. In: ICML. pp. 24851–24871. PMLR (2022)

    2022

  32. [40]

    TMI39(6),1856– 1867 (2019)

    Zhou, Z., Siddiquee, M.M.R., Tajbakhsh, N., Liang, J.: Unet++: Redesigning skip connections to exploitmultiscale features in image segmentation. TMI39(6),1856– 1867 (2019)

    2019

Pith tools

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