pith. sign in

arxiv: 2606.09378 · v1 · pith:VWOSYC6Anew · submitted 2026-06-08 · 💻 cs.CV

Echo-DM: Ultrasound Marker Removal via Conditional Latent Diffusion and Region-Aware Fusion

Zhiwei Wang , Tao Huang , Wentao Jiang , Muyi Li , Jianxin Liu , Jian Chen , Jie Zou , Yong Luo
show 2 more authors
Bo Du Jing Zhang
This is my paper

Pith reviewed 2026-06-27 17:27 UTC · model grok-4.3

classification 💻 cs.CV
keywords ultrasound marker removalconditional latent diffusionregion-aware fusionmedical image restorationEcho-PAIR datasetmask-free inferenceDiT diffusion
0
0 comments X

The pith

Echo-DM removes ultrasound markers via conditional latent diffusion and region-aware fusion without requiring masks at inference time.

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

The paper presents Echo-DM to eliminate artificial markers such as calipers and text from clinical ultrasound images that can bias automated downstream analysis. It encodes images into a latent space, applies a DiT-based conditional diffusion network for global restoration, and uses a region-aware fusion module for local refinement that preserves unaffected areas. This produces an end-to-end mask-free pipeline instantiated in two variants with different latent encoders. On the Echo-PAIR paired dataset the approach removes markers more effectively than two-stage baselines while maintaining anatomical fidelity and offering practical speed-quality balances.

Core claim

Echo-DM follows an encoder-diffusion-decoder design in which a conditional latent diffusion network performs global marker removal and a region-aware fusion module enforces preservation-aware refinement in image space, enabling mask-free inference that avoids both error propagation from explicit masks and over-smoothing from deterministic restorers.

What carries the argument

The region-aware fusion module, which performs preservation-aware image-space refinement after latent diffusion to maintain background consistency under mask-free operation.

If this is right

  • Downstream diagnostic models trained on cleaned images should rely less on marker shortcuts and more on anatomical features.
  • The architecture works with both VAE-based and RAE-based latent modules, indicating flexibility across encoder choices.
  • The method supplies favorable quality-efficiency operating points for different clinical deployment constraints.
  • Marker removal quality exceeds that of representative two-stage baselines on the Echo-PAIR dataset.

Where Pith is reading between the lines

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

  • The same diffusion-plus-fusion pattern could be tested on other imaging modalities that carry similar overlay artifacts.
  • Releasing the paired dataset would allow direct measurement of how much marker bias affects current ultrasound analysis models.
  • End-to-end mask-free operation removes a source of annotation cost that mask-dependent methods incur at scale.

Load-bearing premise

The region-aware fusion module can enforce preservation of unaffected regions during end-to-end mask-free inference without introducing artifacts or inconsistencies that affect anatomical fidelity.

What would settle it

A side-by-side comparison on Echo-PAIR test pairs in which the output differs visibly from the clean ground-truth image in any region outside the original marker locations would falsify the preservation claim.

Figures

Figures reproduced from arXiv: 2606.09378 by Bo Du, Jian Chen, Jianxin Liu, Jie Zou, Jing Zhang, Muyi Li, Tao Huang, Wentao Jiang, Yong Luo, Zhiwei Wang.

Figure 1
Figure 1. Figure 1: Motivating evidence for marker-induced train–deployment mismatch in ultrasound analysis. (a) Qualitative example on a shared clean test image: models trained on marked images show localization drift, whereas using de-marked training images improves detection quality. (b) Quantitative comparison of different training-data constructions evaluated on the same clean test set. Models trained on images processed… view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of Echo-DM. (a) Echo-PAIR: A Large-scale Paired Ultrasound Dataset. Echo-PAIR provides about 20K paired clean-marked ultrasound images covering multi-vendor, multi-organ, and multi-marker scenarios, supporting end-to-end mask-free inference in a codec-flexible diffusion-fusion framework. (b) Conditional Latent Diffusion: Global Marker Removal. The marked input 𝑥𝑚 is encoded into latent sp… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of marker-region reconstruction across methods. Four representative marker-affected regions from Echo-PAIR are shown, with columns corresponding to the marked input, clean GT, Echo-DM-V, Echo-DM-R, the DiT baseline, and SD v1.5 inpainting. Echo-DM-V and Echo-DM-R remove markers while better preserving local ultrasound texture and structural continuity. By comparison, the DiT baseline… view at source ↗
Figure 4
Figure 4. Figure 4: Multiscale feature, soft-mask, and output analysis in Echo-DM-V. Feature stage: encoder and decoder features at selected levels are interpolated to a unified size and concatenated as the input to mask prediction; encoder responses emphasize marker-corrupted regions, while decoder responses shift after restoration, and their discrepancy provides implicit localization cues. Mask stage: the predicted soft mas… view at source ↗
read the original abstract

Clinical ultrasound images often contain artificial markers, such as measurement calipers and text, to assist diagnostic interpretation and comparison. However, these markers can introduce shortcut bias in downstream automated analysis, encouraging deep learning models to rely on marker-related cues rather than clinically meaningful anatomy. Existing marker removal methods are either mask-dependent and vulnerable to error propagation, or mask-free deterministic restorers that may over-smooth ultrasound texture and perturb unaffected background regions. To address these challenges, we present Echo-DM, a framework for ultrasound marker removal via conditional latent diffusion and region-aware fusion. Echo-DM follows a common encoder-diffusion-decoder pipeline, where a DiT-based conditional latent diffusion network performs global restoration and a region-aware fusion module enforces preservation-aware image-space refinement under end-to-end mask-free inference. Building on this fixed core design, we further instantiate Echo-DM-V and Echo-DM-R with VAE-based and RAE-based latent modules, respectively, which demonstrates that the Echo-DM architecture is compatible with diverse latent-module instantiations. Extensive experiments on Echo-PAIR, a large-scale paired clinical ultrasound dataset, demonstrate superior marker removal and strong anatomical fidelity compared with representative two-stage baselines, while providing favorable quality--efficiency trade-offs across deployment settings. Data, code and models will be released at https://github.com/MiliLab/Echo-DM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper presents Echo-DM, a conditional latent diffusion framework for removing artificial markers (e.g., calipers, text) from clinical ultrasound images. It uses a DiT-based diffusion network for global restoration followed by a region-aware fusion module for image-space refinement in an end-to-end mask-free setting. Two latent-module variants (VAE-based Echo-DM-V and RAE-based Echo-DM-R) are instantiated, and experiments on the Echo-PAIR paired dataset claim superior marker removal, anatomical fidelity, and quality-efficiency trade-offs versus two-stage baselines, with public release of data, code, and models planned.

Significance. If the central empirical claims hold under verification, the work offers a practical mask-free alternative to existing marker-removal pipelines that could reduce shortcut bias in downstream ultrasound analysis models. The planned artifact release is a clear strength that supports reproducibility and extension by the community.

major comments (2)
  1. [§3.2] §3.2 (region-aware fusion module): The description of how region awareness is implemented during mask-free inference (e.g., via learned attention, implicit masking, or auxiliary preservation losses) is insufficiently detailed to evaluate whether unaffected background regions are reliably protected from perturbation. This mechanism is load-bearing for the superiority claim over mask-dependent baselines and the assertion of strong anatomical fidelity.
  2. [§4] §4 (experiments on Echo-PAIR): The abstract asserts quantitative superiority in marker removal and fidelity, yet the manuscript provides no error bars, statistical significance tests, or per-region breakdown (marker vs. background) that would confirm the fusion module avoids introducing inconsistencies in unaffected areas. Without these, the cross-method comparison cannot be fully assessed.
minor comments (2)
  1. [Abstract, §1] The acronyms Echo-DM-V and Echo-DM-R are introduced without immediate expansion on first use in the abstract and §1.
  2. [Figures 2-3] Figure captions and method diagrams should explicitly label the region-aware fusion block to clarify its placement relative to the diffusion decoder.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the region-aware fusion module and experimental reporting. We address each major comment below.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (region-aware fusion module): The description of how region awareness is implemented during mask-free inference (e.g., via learned attention, implicit masking, or auxiliary preservation losses) is insufficiently detailed to evaluate whether unaffected background regions are reliably protected from perturbation. This mechanism is load-bearing for the superiority claim over mask-dependent baselines and the assertion of strong anatomical fidelity.

    Authors: We agree that §3.2 requires additional detail on the implementation of region awareness under mask-free inference. In the revised manuscript we will expand this section to specify the learned attention formulation, the implicit protection of background regions, and the role of any auxiliary preservation losses, thereby clarifying how the fusion module safeguards unaffected areas. revision: yes

  2. Referee: [§4] §4 (experiments on Echo-PAIR): The abstract asserts quantitative superiority in marker removal and fidelity, yet the manuscript provides no error bars, statistical significance tests, or per-region breakdown (marker vs. background) that would confirm the fusion module avoids introducing inconsistencies in unaffected areas. Without these, the cross-method comparison cannot be fully assessed.

    Authors: We acknowledge that the current experimental section lacks error bars, statistical significance testing, and per-region (marker vs. background) breakdowns. We will revise §4 to report standard deviations across runs, include appropriate statistical tests, and add separate quantitative results for marker and background regions to strengthen the evaluation of the fusion module. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical engineering contribution

full rationale

The paper presents Echo-DM as an encoder-diffusion-decoder architecture with a region-aware fusion module, instantiated in variants (Echo-DM-V, Echo-DM-R) and validated empirically on the Echo-PAIR paired dataset. No equations, derivations, or claims reduce by construction to fitted inputs or self-citations; performance claims rest on external experimental comparisons rather than internal redefinitions or load-bearing self-references. The central differentiator (mask-free inference with region preservation) is asserted via design and results, not via any tautological step.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no details on free parameters, axioms, or invented entities; insufficient information available for ledger population.

pith-pipeline@v0.9.1-grok · 5797 in / 1046 out tokens · 24415 ms · 2026-06-27T17:27:11.844209+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

36 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Mededit: Counterfactual diffusion-based image editing on brain mri, in: International Workshop on Simulation and Synthesis in Medical Imaging, Springer

    Alaya, M.B., Lang, D.M., Wiestler, B., Schnabel, J.A., Bercea, C.I., 2024. Mededit: Counterfactual diffusion-based image editing on brain mri, in: International Workshop on Simulation and Synthesis in Medical Imaging, Springer. pp. 167–176

    2024

  2. [2]

    Pet image denoising based on denoising diffusion probabilistic model

    Gong, K., Johnson, K., El Fakhri, G., Li, Q., Pan, T., 2024. Pet image denoising based on denoising diffusion probabilistic model. European Journal of Nuclear Medicine and Molecular Imaging 51, 358–368

    2024

  3. [3]

    Blindinpaintingwithobject-awarediscriminationforartificialmarkerremoval,in: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE

    Guo,X.,Hu,W.,Ni,C.,Chai,W.,Li,S.,Wang,G.,2024. Blindinpaintingwithobject-awarediscriminationforartificialmarkerremoval,in: ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1516–1520

    2024

  4. [4]

    Inpainting pathology in lumbar spine mri with latent diffusion

    Hansen, C., Glinskis, S., Raju, A., Kornreich, M., Park, J., Pawar, J., Herzog, R., Zhang, L., Odry, B., 2024. Inpainting pathology in lumbar spine mri with latent diffusion. arXiv preprint arXiv:2406.02477

    work page arXiv 2024

  5. [5]

    Masked autoencoders are scalable vision learners, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R., 2022. Masked autoencoders are scalable vision learners, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 16000–16009

    2022

  6. [6]

    Ho,J.,Jain,A.,Abbeel,P.,2020.Denoisingdiffusionprobabilisticmodels.Advancesinneuralinformationprocessingsystems33,6840–6851

    2020

  7. [7]

    Scope of validity of psnr in image/video quality assessment

    Huynh-Thu, Q., Ghanbari, M., 2008. Scope of validity of psnr in image/video quality assessment. Electronics letters 44, 800–801

    2008

  8. [8]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H., 2021. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18, 203–211

    2021

  9. [9]

    Denoising diffusion restoration models

    Kawar, B., Elad, M., Ermon, S., Song, J., 2022. Denoising diffusion restoration models. Advances in neural information processing systems 35, 23593–23606

    2022

  10. [10]

    Auto-encoding variational bayes

    Kingma, D.P., Welling, M., 2014. Auto-encoding variational bayes. stat 1050, 1

    2014

  11. [11]

    4681–4690

    Ledig,C.,Theis,L.,Huszár,F.,Caballero,J.,Cunningham,A.,Acosta,A.,Aitken,A.,Tejani,A.,Totz,J.,Wang,Z.,etal.,2017.Photo-realistic single image super-resolution using a generative adversarial network, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4681–4690

    2017

  12. [12]

    Noise2noise: Learning image restoration without clean data, in: International Conference on Machine Learning, PMLR

    Lehtinen, J., Munkberg, J., Hasselgren, J., Laine, S., Karras, T., Aittala, M., Aila, T., 2018. Noise2noise: Learning image restoration without clean data, in: International Conference on Machine Learning, PMLR. pp. 2965–2974

    2018

  13. [13]

    Mat: Mask-aware transformer for large hole image inpainting, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Li, W., Lin, Z., Zhou, K., Qi, L., Wang, Y., Jia, J., 2022. Mat: Mask-aware transformer for large hole image inpainting, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10758–10768

    2022

  14. [14]

    Ultrasound in Medicine & Biology 50, 509–519

    Li,X.,Fu,C.,Xu,S.,Sham,C.W.,2024.Thyroidultrasoundimagedatabaseandmarkermaskinpaintingmethodforresearchanddevelopment. Ultrasound in Medicine & Biology 50, 509–519

    2024

  15. [15]

    Image inpainting for irregular holes using partial convolutions, in: Proceedings of the European conference on computer vision (ECCV), pp

    Liu, G., Reda, F.A., Shih, K.J., Wang, T.C., Tao, A., Catanzaro, B., 2018. Image inpainting for irregular holes using partial convolutions, in: Proceedings of the European conference on computer vision (ECCV), pp. 85–100

    2018

  16. [16]

    Liu, H., Wang, Y., Qian, B., Wang, M., Rui, Y., 2024. Structure matters: Tackling the semantic discrepancy in diffusion models for image inpainting, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8038–8047

    2024

  17. [17]

    Repaint: Inpainting using denoising diffusion probabilistic models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Lugmayr, A., Danelljan, M., Romero, A., Yu, F., Timofte, R., Van Gool, L., 2022. Repaint: Inpainting using denoising diffusion probabilistic models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11461–11471

    2022

  18. [18]

    Dfcl: Dual-pathway fusion contrastive learning for blind single-image visible watermark removal

    Meng, B., Zhou, J., Yang, H., Liu, J., Pu, Y., 2025. Dfcl: Dual-pathway fusion contrastive learning for blind single-image visible watermark removal. Neural Networks 184, 107077

    2025

  19. [19]

    4195–4205

    Peebles,W.,Xie,S.,2023.Scalablediffusionmodelswithtransformers,in:ProceedingsoftheIEEE/CVFinternationalconferenceoncomputer vision, pp. 4195–4205

    2023

  20. [20]

    Domain adaptation of stable diffusion for ultrasound inpainting: a synthetic data approach for enhanced thyroid nodule segmentation

    Prochazka, A., Zeman, J., 2025. Domain adaptation of stable diffusion for ultrasound inpainting: a synthetic data approach for enhanced thyroid nodule segmentation. Journal of Biomedical Informatics , 104963

    2025

  21. [21]

    High-resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B., 2022. High-resolution image synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 10684–10695

    2022

  22. [22]

    Imagesuper-resolutionviaiterativerefinement

    Saharia,C.,Ho,J.,Chan,W.,Salimans,T.,Fleet,D.J.,Norouzi,M.,2022. Imagesuper-resolutionviaiterativerefinement. IEEEtransactions on pattern analysis and machine intelligence 45, 4713–4726. Wang et al.:Preprint submitted to ElsevierPage 17 of 18 Echo-DM: Ultrasound Marker Removal via Conditional Latent Diffusion and Region-Aware Fusion

    2022

  23. [23]

    Removal of manually induced artifacts in ultrasound images of thyroid nodulesbasedonedge-connectionandcriminisiimagerestorationalgorithm

    Sun, M., Meng, Q., Wang, T., Liu, T., Zhu, Y., Qiu, J., Lu, W., 2021. Removal of manually induced artifacts in ultrasound images of thyroid nodulesbasedonedge-connectionandcriminisiimagerestorationalgorithm. ComputerMethodsandProgramsinBiomedicine200,105868

    2021

  24. [24]

    Narrowing the semantic gaps in u-net with learnable skip connections: The case of medical image segmentation

    Wang, H., Cao, P., Yang, J., Zaiane, O., 2024. Narrowing the semantic gaps in u-net with learnable skip connections: The case of medical image segmentation. Neural Networks 178, 106546

    2024

  25. [25]

    Vcnet: A robust approach to blind image inpainting, in: European Conference on Computer Vision, Springer

    Wang, Y., Chen, Y.C., Tao, X., Jia, J., 2020. Vcnet: A robust approach to blind image inpainting, in: European Conference on Computer Vision, Springer. pp. 752–768

    2020

  26. [26]

    Meansquarederror:Loveitorleaveit?anewlookatsignalfidelitymeasures

    Wang,Z.,Bovik,A.C.,2009. Meansquarederror:Loveitorleaveit?anewlookatsignalfidelitymeasures. IEEEsignalprocessingmagazine 26, 98–117

    2009

  27. [27]

    Image quality assessment: from error visibility to structural similarity

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P., 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13, 600–612

    2004

  28. [28]

    Diffcnn: A collaborative framework of diffusion model and cnn for semi-supervised medical image segmentation

    Xu, S., Tian, L., 2025. Diffcnn: A collaborative framework of diffusion model and cnn for semi-supervised medical image segmentation. Neural Networks 191, 107813

    2025

  29. [29]

    Frontiers in Bioengineering and Biotechnology 8, 599

    Yao,S.,Yan,J.,Wu,M.,Yang,X.,Zhang,W.,Lu,H.,Qian,B.,2020.Texturesynthesisbasedthyroidnoduledetectionfrommedicalultrasound images: interpreting and suppressing the adversarial effect of in-place manual annotation. Frontiers in Bioengineering and Biotechnology 8, 599

    2020

  30. [30]

    Cascademarkerremovalalgorithmfor thyroid ultrasound images

    Ying,X.,Zhang,Y.,Yu,M.,Wei,X.,Zhu,J.,Gao,J.,Liu,Z.,Shen,H.,Zhang,R.,Li,X.,etal.,2020. Cascademarkerremovalalgorithmfor thyroid ultrasound images. Medical & Biological Engineering & Computing 58, 2641–2656

    2020

  31. [31]

    Free-form image inpainting with gated convolution, in: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Yu, J., Lin, Z., Yang, J., Shen, X., Lu, X., Huang, T.S., 2019. Free-form image inpainting with gated convolution, in: Proceedings of the IEEE/CVF international conference on computer vision, pp. 4471–4480

    2019

  32. [32]

    Medienet:medicalimageenhancementnetworkbasedonconditional latent diffusion model

    Yuan,W.,Feng,Y.,Wen,T.,Luo,G.,Liang,J.,Sun,Q.,Liang,S.,2025. Medienet:medicalimageenhancementnetworkbasedonconditional latent diffusion model. BMC Medical Imaging 25, 372

    2025

  33. [33]

    Adding conditional control to text-to-image diffusion models, in: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Zhang, L., Rao, A., Agrawala, M., 2023a. Adding conditional control to text-to-image diffusion models, in: Proceedings of the IEEE/CVF international conference on computer vision, pp. 3836–3847

  34. [34]

    The unreasonable effectiveness of deep features as a perceptual metric, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O., 2018. The unreasonable effectiveness of deep features as a perceptual metric, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 586–595

    2018

  35. [35]

    Ultrasonicimage’sannotationremoval:Aself-supervisednoise2noiseapproach

    Zhang,Y.,Jiang,N.,Xie,Z.,Cao,J.,Teng,Y.,2023b. Ultrasonicimage’sannotationremoval:Aself-supervisednoise2noiseapproach. arXiv preprint arXiv:2307.04133

    work page arXiv

  36. [36]

    Diffusion Transformers with Representation Autoencoders

    Zheng, B., Ma, N., Tong, S., Xie, S., 2025. Diffusion transformers with representation autoencoders. arXiv preprint arXiv:2510.11690 . Wang et al.:Preprint submitted to ElsevierPage 18 of 18

    work page internal anchor Pith review Pith/arXiv arXiv 2025