Pith. sign in

REVIEW 4 major objections 6 minor 46 references

LangMamba: A Language-driven Mamba Framework for Low-dose CT Denoising with Vision-language Models

T0 review · 4 major / 6 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 LangMamba uses a frozen medical vision-language model's text tokens as semantic supervision for low-dose CT denoising, outperforming seven prior methods on Mayo-2016 and Mayo-2020 with no extra inference cost.

desk verdict Solid applied denoising paper with public code, but the 'language-driven' causal claim is not yet proven; worth a serious referee who pushes for the missing control. read the letter →

arxiv 2507.06140 v1 pith:OAPEO3AM submitted 2025-07-08 eess.IV cs.AIcs.CV

Zhihao Chen , Tao Chen , Chenhui Wang , Qi Gao , Huidong Xie , Chuang Niu , Ge Wang , Hongming Shan This is my paper
classification eess.IVcs.AIcs.CV
keywords CTdenoisingVision-languagemodelMambaStatespacemodelsVectorquantizationExplainabilityDual-spacealignment
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 argues that the missing ingredient in low-dose CT (LDCT) denoising is not more pixels but semantic supervision: a frozen vision-language model can tell the denoiser what anatomy it is looking at. It proposes LangMamba, a two-stage framework in which a language-guided autoencoder (LangAE) first maps normal-dose CT images to quantized text tokens that carry anatomical meaning, and a Mamba-based denoiser then uses those tokens to align denoised images with normal-dose images in both perceptual and semantic space. On the Mayo-2016 and Mayo-2020 datasets the authors report that LangMamba outperforms seven prior methods (CNN-, GAN-, transformer-, and diffusion-based) in PSNR, SSIM, and FSIM, and that LangAE trained only on Mayo-2016 transfers to Mayo-2020 without retraining. The practical payoff would be better detail preservation and visual fidelity at no extra inference cost, plus language-level explainability in the clinic.

What carries the argument

The central mechanism is a vector-quantized autoencoder whose codebook is not learned but frozen: it is the token-embedding layer of a medical vision-language model, so every quantized representation of a CT image becomes a set of word-like tokens. A pyramid semantic loss over three token layers (4, 64, and 1024 tokens) aligns these tokens with anatomical content at multiple scales while the decoder reconstructs the normal-dose image. The denoiser, SEED, freezes the LangAE encoder as a noise-robust feature extractor and adds an efficient Mamba attention block whose 2D selective scan skips every other patch, cutting the number of scanned tokens by a factor of four, to capture global context. The LangDA loss then aligns denoised and normal-dose images in continuous feature space and discrete token space, with the total denoising loss being a pixel-level MSE term plus a weighted LangDA term.

What would settle it

Train the identical denoiser and losses but replace the frozen medical-VLM codebook with a fixed, randomly initialized codebook of the same size and dimension, keeping the same pyramid semantic loss. If PSNR, SSIM, and FSIM on Mayo-2016 stay statistically indistinguishable from LangMamba's reported numbers, then the word-level semantics are not the active ingredient; as a complementary check, have radiologists judge whether the quantized tokens assigned to each image region match its anatomy, since the paper itself notes many tokens correlate poorly with CT structure.

Watch

Extended reading notes

Core claim

The central claim is that replacing the learned codebook of a vector-quantized autoencoder with frozen token embeddings from a medical vision-language model, and supervising the quantizer with a pyramid semantic loss, yields discrete tokens whose semantic content improves LDCT denoising. In the full framework, the frozen LangAE encoder supplies noise-robust local features to a U-shaped denoiser whose decoder uses efficient Mamba attention blocks, and a dual-space loss minimizes Euclidean distance between denoised and normal-dose images in both continuous features and discrete token embeddings. The authors report that LangMamba achieves the best PSNR, SSIM, and FSIM among all compared methods on both Mayo-2016 and Mayo-2020, for example PSNR 28.83 dB on Mayo-2016 versus 28.70 dB for the strongest diffusion baseline, with an inference time of 0.032 seconds per slice. They also claim that LangAE generalizes to an unseen dataset without retraining, and that the quantized tokens provide text-based explainability, such as tokens like 'liver', 'kidney', and 'lesion'.

Load-bearing premise

The method depends on the assumption that fixed word meanings from a general medical text-vision model actually match the anatomy visible in CT slices, so that pushing a denoised image toward those words teaches it something useful about anatomy.

Editorial extensions

If this is right

  • If the reported results hold, VLM-derived semantic supervision can improve LDCT denoising beyond pixel-level training, without adding any computational cost at inference time.
  • The pre-trained LangAE can be reused on new datasets without retraining, which would shorten deployment pipelines for new scanners or dose protocols.
  • The LangDA loss is plug-and-play: attaching it to an existing denoiser such as RED-CNN improves SSIM and FSIM, meaning the semantic alignment could be adopted by other backbones.
  • The quantized text tokens provide a route to explainability: radiologists could see which anatomical terms, such as liver, kidney, or lesion, the model associates with a denoised region.
  • Because the denoiser needs only a single forward pass, it offers diffusion-level quality with much shorter inference than multi-step generative baselines.

Reading between the lines

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

  • The paper's own ablation shows that adding the VLM codebook alone worsens reconstruction (FID 62.18 vs 58.17), so the active ingredient may be the pyramid semantic loss rather than word semantics per se; an editor would want to test whether a non-semantic fixed codebook trained with the same pyramid loss obtains the same denoising gain.
  • If the mechanism is semantic alignment rather than the specific vocabulary, the same two-stage recipe could transfer to other medical restoration tasks with paired high- and low-quality data, such as MRI denoising or CT deblurring, whenever a medical vision-language model supplies the token space.
  • The cross-dataset claim would be stronger if tested in the reverse direction (Mayo-2020 pretraining to Mayo-2016) and on anatomies outside the abdomen, since the current evidence is one-directional on abdominal CT only.
  • The explainability claim is partial: the paper displays only a subset of tokens from layers 2-3 because many candidate tokens correlate poorly with CT anatomy, so a quantitative token-to-structure correspondence study is needed to substantiate radiologist-facing explainability.
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 / 6 minor

Summary. LangMamba is a two-stage framework for low-dose CT (LDCT) denoising. In the first stage, a Language-guided AutoEncoder (LangAE) is pretrained on normal-dose CT (NDCT) by replacing the VQGAN codebook with frozen PubMedCLIP token embeddings and adding a SPAE-style pyramid semantic loss. In the second stage, the frozen LangAE encoder and multi-scale quantizer are used inside a Semantic-Enhanced Efficient Denoiser (SEED) with a Mamba-based decoder, and a Language-engaged Dual-space Alignment (LangDA) loss aligns denoised images with NDCT in both continuous perceptual and discrete semantic spaces. The paper reports superior PSNR, SSIM, and FSIM over seven existing methods on Mayo-2016 and Mayo-2020, demonstrates cross-dataset transfer of the pretrained LangAE, presents plug-and-play results of LangDA on RED-CNN, and illustrates token-based explainability.

Significance. If the language-semantics mechanism were convincingly established, this would be a useful contribution to medical image restoration: it proposes a resource-efficient way to inject frozen vision-language knowledge into an LDCT denoiser, with no extra inference cost, a linear-complexity Mamba decoder, public code, two public datasets, standard evaluation metrics, and a properly disclosed preliminary version. The authors should be credited for the breadth of the empirical study, including ablations of SEED, LangDA, and the LangAE components, as well as for the explicit discussion of limitations. However, the central claim that VLM-derived semantic supervision causes the improvement is not yet supported by the ablations, and several smaller issues in statistical reporting and notation need attention.

major comments (4)
  1. [Table V / Sec. IV-E3] The ablation in Table V does not isolate the contribution of VLM language semantics. The row '+semantic loss' simultaneously adds the pyramid semantic loss and improves the autoencoder's reconstruction FID from 62.18 to 51.79, and Sec. IV-E3 states that LangAE reconstruction quality directly affects denoising performance. Because no control with a learned or random codebook under an equally strong non-linguistic alignment or diversity loss is provided, the improved denoising PSNR/SSIM in the third row can be explained by better reconstruction quality or by the regularizing effect of the pyramid loss alone, without any causal role for the PubMedCLIP token embeddings. This is the load-bearing point for the 'language-driven' claim, so please add such controls or substantially soften the semantic-causality interpretation.
  2. [Eq. (5) / Sec. III-B] The semantic loss in Eq. (5) is written with exp(+||z_l - e(c)||^2) in the numerator and exp(-||z_l - e(k)||^2) in the denominator; this expression is not a normalized softmax and would make the log term negative. If the implementation uses exp(-distance) throughout, please correct the equation; if not, the loss is not what is described. This formula is central to LangAE training and must be reproducible.
  3. [Table I / Sec. IV-C2] The headline claim that LangMamba outperforms all previous methods rests on small margins relative to the reported variability: on Mayo-2016 the PSNR gain over the best baseline is 0.09 dB and the SSIM gain is 0.0015, and on Mayo-2020 the PSNR gain over CoreDiff is 0.20 dB. The table reports slice-level mean +/- std, not run-to-run variability, and no paired significance test or repeated-seed results are given. Please add paired tests across the test slices (or multiple training runs) to support the superiority claim, or explicitly frame the differences as descriptive.
  4. [Sec. IV-D / Fig. 6] The language-level explainability claim rests on a handful of selected tokens, while Sec. V acknowledges that many tokens in the candidate pools correlate poorly with CT anatomy. Without a human reader study, a quantitative token-anatomy alignment measure, or a comparison against a non-semantic tokenization, the explainability contribution is not demonstrated; at minimum the wording should be reduced to a qualitative illustration.
minor comments (6)
  1. [Sec. III-B] After Eq. (9), the hyperparameter list 'α, β, λ, and η are set to 0.3, 0.3, 0.1, and 0.1' should be 'α, β, γ, and η' because Eq. (9) contains γ, while λ is introduced in Eq. (13) as the LangDA weight.
  2. [Sec. IV-B / Table I] Please report the number of test slices and clarify how the HU window is applied; Sec. IV-B mentions a training window of [-1000, 2000] HU, while the metrics are computed in [-160, 240] HU.
  3. [Sec. IV-C1 / Figs. 4-5] The baseline name is spelled both 'ESAU-Net' and 'EASU-Net'; please standardize the spelling and verify that reference [36] is the correct source for the evaluated method.
  4. [Sec. IV-D / Fig. 6] The token lists contain 'skelet' and a duplicated 'intestinal'; if these are actual tokenizer outputs, please state so, otherwise correct the typos.
  5. [Abstract / Sec. III-A] The phrase 'LangAE requires no retraining on new datasets' should be qualified because the SEED denoiser is retrained on each target dataset; only the frozen LangAE transfers.
  6. [Table IV / Sec. IV-E2] The full LangDA yields lower SSIM and FSIM than LangDA-C on RED-CNN (0.8637 vs 0.8649 and 0.9389 vs 0.9421); the claim of a 'balanced improvement' should address this trade-off explicitly.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: LangMamba's claims are independently evaluated; the only self-citation is a transparent note about the preliminary LangAE paper.

full rationale

LangMamba's derivation chain does not reduce to its own inputs. The two-stage pipeline (LangAE pretraining with a frozen PubMedCLIP codebook and SPAE-style pyramid semantic loss, followed by SEED plus the LangDA loss) is a new training objective, not a fitted constant, and the reported gains are measured on held-out patients from two public datasets using PSNR, SSIM, and FSIM, which are external to the loss. Table V is an ablation showing that adding the semantic loss to the VLM codebook improves reconstruction (FID 51.79 vs 62.18) and downstream denoising (28.83 vs 28.63 PSNR); even if the causal role of language semantics is not fully isolated because no non-linguistic codebook control is tested, that is a confound or validity concern, not circularity, since the comparison is between training regimes rather than between a fitted input and a claimed prediction. The explicit note that a preliminary LangAE version appeared in the authors' BIBM 2024 paper [17] is transparent self-citation: the present paper re-runs the experiments and does not rely on [17]'s numbers as evidence, so it is not load-bearing. No uniqueness theorem from the authors is invoked, and the VQGAN, LQAE, SPAE, and PubMedCLIP components are cited from external sources. The central claim is therefore self-contained against external benchmarks, and the minor self-citation does not create circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The only invented entity is not an entity but a new loss function. The free parameters are hyperparameters, none of which are fitted to a held-out metric in a way that makes the reported results circular. The main burden is the assumption that VLM text embeddings align with CT anatomy.

free parameters (4)
  • lambda (LangDA weight) = 0.3
    Empirically set in Eq. (13); not derived from theory.
  • Semantic loss weights (alpha, beta, gamma, eta) = 0.3, 0.3, 0.1, 0.1
    Taken from SPAE [13], not tuned to LDCT data.
  • Pyramid thresholds rho_l for token pools = 0.95, 0.9, 0.8 for 3 layers
    Chosen by hand; directly controls which tokens enter the semantic loss.
  • ES2D skipping step size s = 2
    Chosen to reduce tokens by 4x; not justified by a search.
assumptions (3)
  • domain assumption PubMedCLIP frozen text embeddings provide a semantically meaningful codebook for CT image tokens.
    The whole framework depends on the VLM codebook carrying anatomical semantics. The paper's own Table V shows VLM codebook alone hurts performance, and the Discussion admits many tokens correlate poorly with CT anatomy.
  • domain assumption The SPAE pyramid semantic loss transfers to CT images.
    The semantic loss in Eq. (5) is copied from SPAE [13] with no CT-specific validation of whether its candidate token pools contain relevant anatomy.
  • domain assumption Mayo-2016 and Mayo-2020 simulated low-dose CT pairs are representative of clinical LDCT denoising.
    Both datasets are simulated-dose (quarter-dose/25% dose) abdominal CT; results may not transfer to other anatomies or real low-dose acquisitions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LangMamba: A Language-driven Mamba Framework for Low-dose CT Denoising with Vision-language Models." pith.science (2026). https://pith.science/paper/OAPEO3AM

@misc{pith2026250706140,
  author       = {Pith},
  title        = {Pith review of: LangMamba: A Language-driven Mamba Framework for Low-dose CT Denoising with Vision-language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OAPEO3AM}},
  note         = {Machine review of arXiv:2507.06140}
}
read the original abstract

Low-dose computed tomography (LDCT) reduces radiation exposure but often degrades image quality, potentially compromising diagnostic accuracy. Existing deep learning-based denoising methods focus primarily on pixel-level mappings, overlooking the potential benefits of high-level semantic guidance. Recent advances in vision-language models (VLMs) suggest that language can serve as a powerful tool for capturing structured semantic information, offering new opportunities to improve LDCT reconstruction. In this paper, we introduce LangMamba, a Language-driven Mamba framework for LDCT denoising that leverages VLM-derived representations to enhance supervision from normal-dose CT (NDCT). LangMamba follows a two-stage learning strategy. First, we pre-train a Language-guided AutoEncoder (LangAE) that leverages frozen VLMs to map NDCT images into a semantic space enriched with anatomical information. Second, we synergize LangAE with two key components to guide LDCT denoising: Semantic-Enhanced Efficient Denoiser (SEED), which enhances NDCT-relevant local semantic while capturing global features with efficient Mamba mechanism, and Language-engaged Dual-space Alignment (LangDA) Loss, which ensures that denoised images align with NDCT in both perceptual and semantic spaces. Extensive experiments on two public datasets demonstrate that LangMamba outperforms conventional state-of-the-art methods, significantly improving detail preservation and visual fidelity. Remarkably, LangAE exhibits strong generalizability to unseen datasets, thereby reducing training costs. Furthermore, LangDA loss improves explainability by integrating language-guided insights into image reconstruction and offers a plug-and-play fashion. Our findings shed new light on the potential of language as a supervisory signal to advance LDCT denoising. The code is publicly available on https://github.com/hao1635/LangMamba.

Figures

Figures reproduced from arXiv: 2507.06140 by the authors.

Figure 1
Figure 1. Overall architecture of proposed language-driven Mamba framework (LangMamba); (a) Architecture of language-guided autoencoder (LangAE); [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Feature similarity between NDCT and LDCT on different feature [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Architecture of ES2D, which skips scan patches with a step size [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Explainability provided by the text tokens in the 3-layer token [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative results of ablation studies on the EMA block and LangDA loss in LangMamba. (a) NDCT; (b) LDCT; (c) SEED w/o EMA; (d) SEED-R; [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results on the LangDA as a plug-and-play loss in RED-CNN. (a) NDCT; (b) LDCT; (c) RED-CNN; (d) RED-CNN+ [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 39 canonical work pages

  1. [1]

    CT image denoising and deblurring with deep learning: current status and perspectives,

    Y . Lei, C. Niu, J. Zhang, G. Wang, and H. Shan, “CT image denoising and deblurring with deep learning: current status and perspectives,”IEEE Trans. Radiat. Plasma Med. Sci. , vol. 8, no. 2, pp. 153–172, 2024

    work page 2024

  2. [2]

    Deep learning-based algorithms for low-dose CT imaging: A review,

    H. Chen, Q. Li, L. Zhou, and F. Li, “Deep learning-based algorithms for low-dose CT imaging: A review,” Eur. J. Radiol., p. 111355, 2024

    work page 2024

  3. [3]

    Low-dose CT lung cancer screening practices and attitudes among primary care providers at an academic medical center,

    J. A. Lewis, W. J. Petty, J. A. Tooze, D. P. Miller, C. Chiles, A. A. Miller, C. Bellinger, and K. E. Weaver, “Low-dose CT lung cancer screening practices and attitudes among primary care providers at an academic medical center,” Cancer Epidemiol. Biomarkers Prev. , vol. 24, no. 4, pp. 664–670, 2015

    work page 2015

  4. [4]

    PPFM: Image denoising in photon-counting CT using single-step posterior sampling Poisson flow generative mod- els,

    D. Hein, S. Holmin, T. Szczykutowicz, J. S. Maltz, M. Danielsson, G. Wang, and M. Persson, “PPFM: Image denoising in photon-counting CT using single-step posterior sampling Poisson flow generative mod- els,” IEEE Trans. Radiat. Plasma Med. Sci. , vol. 8, no. 7, pp. 788–799, 2024

    work page 2024

  5. [5]

    3-D convolutional encoder-decoder network for low-dose CT via transfer learning from a 2-D trained network,

    H. Shan, Y . Zhang, Q. Yang, U. Kruger, M. K. Kalra, L. Sun, W. Cong, and G. Wang, “3-D convolutional encoder-decoder network for low-dose CT via transfer learning from a 2-D trained network,” IEEE Trans. Med. Imag., vol. 37, no. 6, pp. 1522–1534, 2018

    work page 2018

  6. [6]

    Self-adaptive weight embedded lightweight network using semi- supervised learning for low-dose CT image denoising,

    J. Wang, H. Fan, Z. Wu, Q. Du, M. Li, J. Zheng, G. S. Mok, and B. M. Tsui, “Self-adaptive weight embedded lightweight network using semi- supervised learning for low-dose CT image denoising,” IEEE Trans. Radiat. Plasma Med. Sci. , 2025

    work page 2025

  7. [7]

    PrideDiff: Physics-regularized generalized diffusion model for CT reconstruction,

    Z. Lu, Q. Gao, T. Wang, Z. Yang, Z. Wang, H. Yu, H. Chen, J. Zhou, H. Shan, and Y . Zhang, “PrideDiff: Physics-regularized generalized diffusion model for CT reconstruction,” IEEE Trans. Radiat. Plasma Med. Sci., vol. 9, no. 2, pp. 157–168, 2025

    work page 2025

  8. [8]

    LIT-Former: Linking in-plane and through-plane transformers for simultaneous CT image denoising and deblurring,

    Z. Chen, C. Niu, Q. Gao, G. Wang, and H. Shan, “LIT-Former: Linking in-plane and through-plane transformers for simultaneous CT image denoising and deblurring,” IEEE Trans. Med. Imag. , vol. 43, no. 5, pp. 1880–1894, 2024

    work page 2024

Show all 46 references
  1. [9]

    CoSeR: Bridging image and language for cognitive super-resolution,

    H. Sun, W. Li, J. Liu, H. Chen, R. Pei, X. Zou, Y . Yan, and Y . Yang, “CoSeR: Bridging image and language for cognitive super-resolution,” in CVPR, 2024, pp. 25 868–25 878

    2024

  2. [10]

    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. Clark et al., “Learning transferable visual models from natural language supervision,” in ICML. PMLR, 2021, pp. 8748–8763

    2021

  3. [11]

    BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” in ICML. PMLR, 2022, pp. 12 888–12 900

    2022

  4. [12]

    IQAGPT: Computed tomography image quality assessment with vision- language and ChatGPT models,

    Z. Chen, B. Hu, C. Niu, T. Chen, Y . Li, H. Shan, and G. Wang, “IQAGPT: Computed tomography image quality assessment with vision- language and ChatGPT models,” Vis. Comput. Ind. Biomed. Art. , vol. 7, no. 1, p. 20, 2024

    2024

  5. [13]

    SPAE: Semantic pyramid autoencoder for multimodal generation with frozen LLMs,

    L. Yu, Y . Cheng, Z. Wang, V . Kumar, W. Macherey, Y . Huang, D. Ross, I. Essa, Y . Bisk, M.-H. Yang et al., “SPAE: Semantic pyramid autoencoder for multimodal generation with frozen LLMs,” in NeurIPS, vol. 36, 2023, pp. 52 692–52 704

    2023

  6. [14]

    LLM-Seg: Bridging image segmentation and large language model reasoning,

    J. Wang and L. Ke, “LLM-Seg: Bridging image segmentation and large language model reasoning,” in CVPR, 2024, pp. 1765–1774

    2024

  7. [15]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” in CVPR, 2021, pp. 12 873–12 883. IEEE TRANSACTIONS ON RADIATION AND PLASMA MEDICAL SCIENCES, VOL. XX, NO. X, 2025 11

    2021

  8. [16]

    Language quantized autoencoders: Towards unsupervised text-image alignment,

    H. Liu, W. Yan, and P. Abbeel, “Language quantized autoencoders: Towards unsupervised text-image alignment,” in NeurIPS, vol. 36, 2023, pp. 4382–4395

    2023

  9. [17]

    Low-dose CT denoising with language-engaged dual-space alignment,

    Z. Chen, T. Chen, C. Wang, C. Niu, G. Wang, and H. Shan, “Low-dose CT denoising with language-engaged dual-space alignment,” in BIBM, 2024, pp. 3088–3091

    2024

  10. [18]

    Neural discrete representation learning,

    A. Van Den Oord, O. Vinyals et al. , “Neural discrete representation learning,” in NeurIPS, vol. 30, 2017

    2017

  11. [19]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in CVPR, 2022, pp. 10 684–10 695

    2022

  12. [20]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. D. M.-W. C. Kenton and L. K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in NAACL, vol. 1. Minneapolis, Minnesota, 2019, p. 2

    2019

  13. [21]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. Ré, “Efficiently modeling long sequences with structured state spaces,” in ICLR, 2022

    2022

  14. [22]

    Simplified state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. Linderman, “Simplified state space layers for sequence modeling,” in ICLR, 2023

    2023

  15. [23]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” in COLM, 2024

    2024

  16. [24]

    VMamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “VMamba: Visual state space model,” inNeurIPS, vol. 37, 2024, pp. 103 031–103 063

    2024

  17. [25]

    MambaIR: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “MambaIR: A simple baseline for image restoration with state-space model,” in ECCV. Springer, 2025, pp. 222–241

    2025

  18. [26]

    EfficientVMamba: Atrous selective scan for light weight visual Mamba,

    X. Pei, T. Huang, and C. Xu, “EfficientVMamba: Atrous selective scan for light weight visual Mamba,” arXiv:2403.09977, 2024

    2024 arXiv

  19. [27]

    Does CLIP benefit visual question answering in the medical domain as much as it does in the general domain?

    S. Eslami, G. de Melo, and C. Meinel, “Does CLIP benefit visual question answering in the medical domain as much as it does in the general domain?” arXiv:2112.13906, 2021

    2021 arXiv

  20. [28]

    Perceptual losses for real-time style transfer and super-resolution,

    J. Johnson, A. Alahi, and L. Fei-Fei, “Perceptual losses for real-time style transfer and super-resolution,” in ECCV. Springer, 2016, pp. 694–711

    2016

  21. [29]

    CBAM: Convolutional block attention module,

    S. Woo, J. Park, J.-Y . Lee, and I. S. Kweon, “CBAM: Convolutional block attention module,” in ECCV, 2018, pp. 3–19

    2018

  22. [30]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

    S. Elfwing, E. Uchibe, and K. Doya, “Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,” Neural Netw., vol. 107, pp. 3–11, 2018

    2018

  23. [31]

    Low-dose CT with a residual encoder-decoder convolutional neural network,

    H. Chen, Y . Zhang, M. K. Kalra, F. Lin, Y . Chen, P. Liao, J. Zhou, and G. Wang, “Low-dose CT with a residual encoder-decoder convolutional neural network,” IEEE Trans. Med. Imag. , vol. 36, no. 12, pp. 2524– 2535, 2017

    2017

  24. [32]

    EDCNN: Edge enhancement- based densely connected network with compound loss for low-dose CT denoising,

    T. Liang, Y . Jin, Y . Li, and T. Wang, “EDCNN: Edge enhancement- based densely connected network with compound loss for low-dose CT denoising,” in ICSP, vol. 1. IEEE, 2020, pp. 193–198

    2020

  25. [33]

    Low-dose CT image denoising using a generative adversarial network with wasserstein distance and perceptual loss,

    Q. Yang, P. Yan, Y . Zhang, H. Yu, Y . Shi, X. Mou, M. K. Kalra, Y . Zhang, L. Sun, and G. Wang, “Low-dose CT image denoising using a generative adversarial network with wasserstein distance and perceptual loss,” IEEE Trans. Med. Imag. , vol. 37, no. 6, pp. 1348–1357, 2018

    2018

  26. [34]

    DU-GAN: Generative adversarial networks with dual-domain U-Net-based discriminators for low-dose CT denoising,

    Z. Huang, J. Zhang, Y . Zhang, and H. Shan, “DU-GAN: Generative adversarial networks with dual-domain U-Net-based discriminators for low-dose CT denoising,” IEEE Trans. Instrum. Meas., vol. 71, pp. 1–12, 2021

    2021

  27. [35]

    Hformer: highly efficient vision transformer for low-dose CT denoising,

    S. Zhang, Z. Wang, H. Yang, Y . Chen, Y . Li, Q. Pan, H. Wang, and C. Zhao, “Hformer: highly efficient vision transformer for low-dose CT denoising,” Nucl. Sci. Tech., vol. 34, no. 4, p. 61, 2023

    2023

  28. [36]

    ASCON: Anatomy-aware supervised contrastive learning framework for low-dose CT denoising,

    Z. Chen, Q. Gao, Y . Zhang, and H. Shan, “ASCON: Anatomy-aware supervised contrastive learning framework for low-dose CT denoising,” in MICCAI. Springer, 2023, pp. 355–365

    2023

  29. [37]

    CoreDiff: Contextual error-modulated generalized diffusion model for low-dose CT denoising and generalization,

    Q. Gao, Z. Li, J. Zhang, Y . Zhang, and H. Shan, “CoreDiff: Contextual error-modulated generalized diffusion model for low-dose CT denoising and generalization,” IEEE Trans. Med. Imag. , vol. 43, no. 2, pp. 745– 759, 2024

    2024

  30. [38]

    Low-dose CT for the detection and classification of metastatic liver lesions: Results of the 2016 low dose CT grand challenge,

    C. H. McCollough, A. C. Bartley, R. E. Carter, B. Chen, T. A. Drees, P. Edwards, D. R. Holmes III, A. E. Huang, F. Khan, S. Leng et al. , “Low-dose CT for the detection and classification of metastatic liver lesions: Results of the 2016 low dose CT grand challenge,” Med. Phys....

    2016

  31. [39]

    Low-dose CT image and projection dataset,

    T. R. Moen, B. Chen, D. R. Holmes III, X. Duan, Z. Yu, L. Yu, S. Leng, J. G. Fletcher, and C. H. McCollough, “Low-dose CT image and projection dataset,” Med. Phys., vol. 48, no. 2, pp. 902–911, 2021

    2021

  32. [40]

    Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose CT image reconstruction,

    H. Shan, A. Padole, F. Homayounieh, U. Kruger, R. D. Khera, C. Niti- warangkul, M. K. Kalra, and G. Wang, “Competitive performance of a modularized deep neural network compared to commercial algorithms for low-dose CT image reconstruction,” Nat. Mach. Intell., vol. 1, no. 6, p...

    2019

  33. [41]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in ICLR, 2019

    2019

  34. [42]

    SGDR: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” in ICLR, 2017

    2017

  35. [43]

    Deep residual learning for image recognition,

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

    2016

  36. [44]

    A perspective on deep imaging,

    G. Wang, “A perspective on deep imaging,” IEEE Access , vol. 4, pp. 8914–8924, 2016

    2016

  37. [45]

    Comparison of objective image quality metrics to expert radiologists’ scoring of diagnostic quality of MR images,

    A. Mason, J. Rioux, S. E. Clarke, A. Costa, M. Schmidt, V . Keough, T. Huynh, and S. Beyea, “Comparison of objective image quality metrics to expert radiologists’ scoring of diagnostic quality of MR images,” IEEE Trans. Med. Imag. , vol. 39, no. 4, pp. 1064–1072, 2020

    2020

  38. [46]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction,

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang, “Visual autoregressive modeling: Scalable image generation via next-scale prediction,” in NeurIPS, vol. 37, 2024, pp. 84 839–84 865

    2024

Pith tools

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