Pith. sign in

REVIEW 3 major objections 6 minor 29 references

DiSC-Med: Diffusion-based Semantic Communications for Robust Medical Image Transmission

T0 review · 3 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 This paper proposes DiSC-Med, which sends only compressed anatomical maps—segmentation and edges—of a CT scan and uses a conditional diffusion model at the receiver to regenerate the image.

desk verdict The downstream segmentation evaluation is circular—the receiver conditions on the ground-truth segmentation map—so the paper's main claim does not hold, though the engineering is a reasonable first step. read the letter →

arxiv 2508.00172 v1 pith:7N27NKAP submitted 2025-07-31 cs.LG eess.IV

Fupei Guo , Hao Zheng , Xiang Zhang , Li Chen , Yue Wang , Songyang Zhang This is my paper · ORCID
classification cs.LGeess.IV
keywords semanticcommunicationsmedicalimagetransmissiondiffusionmodelsCTimagingchanneldenoisingreconstructiondownstreamsegmentationbandwidthefficiency
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

DiSC-Med is a semantic communication system for medical images: instead of sending CT pixels, the transmitter extracts and compresses a segmentation map and an edge map, and the receiver regenerates the 3D CT volume from those cues. The authors argue that for telehealth, what matters is diagnostic structure, so transmitting coarse anatomy rather than exact pixel values can cut bandwidth dramatically. A channel-aware denoising step cleans the received semantic maps, and a conditional diffusion model synthesizes the final image. Experiments report better perceptual quality and anatomical segmentation accuracy than autoencoder baselines at similar compression ratios, with the advantage holding under additive Gaussian noise and bit errors.

What carries the argument

The load-bearing mechanism is the conditional reverse diffusion process, where each slice's conditioning input $c_k = \mathrm{Concat}(\hat{x}'_{\mathrm{seg}}(k), \hat{x}'_{\mathrm{edge}}(k))$ concatenates the denoised segmentation and edge maps. The channel-aware denoising U-Net injects an embedding of the channel state (SNR or BER) through feature-wise affine transformations, so the receiver adapts its semantic restoration to noise severity. The 3D-to-2D decomposition keeps the diffusion stage tractable while still exploiting volumetric context by stacking each target slice with its two neighbors.

What would settle it

Apply a downstream segmentation model directly to the transmitted ground-truth segmentation map, or compare DiSC-Med against a receiver that receives the same segmentation and edge maps but reconstructs with a non-diffusion decoder; if the Dice advantage over the baselines disappears or the oracle already achieves the reported scores, the claimed communication superiority is an artifact of information leakage.

Watch

Extended reading notes

Core claim

DiSC-Med claims that a diffusion model can carry most of the communication burden in medical image transmission: the transmitter sends a downsampled segmentation volume and edge volume, and the receiver interpolates them, denoises them with a U-Net conditioned on channel state, and feeds them slice-by-slice into a DDPM that reconstructs the CT scan. The authors report average Dice of 0.8742 versus 0.8335 for the best autoencoder baseline on anatomical structure segmentation from regenerated images, and lower FID under both AWGN and bit-wise error. The key claim is that semantic content, not pixels, is the right payload for bandwidth-limited medical links.

Load-bearing premise

The evaluation assumes that measuring downstream segmentation on an image regenerated from the transmitted ground-truth segmentation map is a valid test of transmission quality, rather than a test of how much privileged label information the receiver already received.

Editorial extensions

If this is right

  • Sending only compressed anatomical maps, rather than pixels, can drastically reduce the data rate for CT transmission while preserving downstream segmentation accuracy.
  • A channel-aware denoising stage that knows the current SNR or BER can maintain reconstruction quality as channel conditions worsen.
  • Conditional diffusion can regenerate diagnostically useful CT volumes from sparse geometric and semantic cues, outperforming learned autoencoder compression on perceptual metrics.
  • The same pipeline could carry additional semantic conditions such as patient metadata, diagnosis records, or k-space information for other medical modalities.

Reading between the lines

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

  • An implication the authors leave implicit: the reported downstream segmentation gain may partly reflect that the receiver is given the exact label map that the evaluation later compares against, so part of the apparent superiority is privileged information rather than recovered image content.
  • A testable extension is to evaluate the marginal contribution of each transmitted channel—for instance, sending only the edge map or only the segmentation map—to see how much of the reconstruction quality each semantic cue actually provides.
  • The compression ratio counts only the downsampled semantic tensors sent over the channel; a fuller accounting would include the cost of running the large pre-trained extraction models at the transmitter and the diffusion model at the receiver, which matters for practical telehealth deployments.
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 / 6 minor

Summary. The paper proposes DiSC-Med, a semantic communication framework for transmitting 3D CT images over noisy, bandwidth-limited channels. At the transmitter, a pretrained TotalSegmentator produces a segmentation volume and a MedSAM/Canny pipeline produces an edge volume; these are downsampled, transmitted under AWGN or bit-error noise, interpolated, and denoised by a channel-aware U-Net, then used as conditioning for a slice-wise DDPM that reconstructs the CT volume. Experiments compare LPIPS/FID and downstream anatomical segmentation Dice/mIoU against DCAE, ResVAE, VQ-VAE-2, and CAE, claiming superior reconstruction and downstream-task performance.

Significance. If the empirical evaluation were valid, the idea of transmitting a compact semantic representation and reconstructing the image with a conditional diffusion model would be a plausible contribution to bandwidth-limited telehealth. The paper does contain useful engineering components, including 3D-to-2D decomposition, channel-aware denoising, and explicit handling of both AWGN and bit-error noise. However, the central downstream-segmentation claim is undermined by an information-leakage flaw: the transmitted representation includes the exact segmentation map that the downstream evaluation treats as ground truth. This makes the reported Dice/mIoU superiority an artifact of the evaluation design rather than evidence of better communication, and it invalidates the headline claim of superior performance in downstream tasks.

major comments (3)
  1. [Section IV-C, Eq. (4)] The downstream segmentation evaluation is circular in effect. The receiver's diffusion conditioning input c_k in Eq. (4) is Concat(seg, edge), where seg is the segmentation map generated by TotalSegmentator from the original image x (Section III-B1). Section IV-C then measures Dice and mIoU by comparing a blind segmentation model's output on the reconstructed image with the same anatomical structures. Because the reconstructed image is generated from the answer key of the segmentation task, the reported Dice primarily measures how faithfully the diffusion model renders the transmitted semantic label, not whether the communication system preserved diagnostically relevant information. The AE baselines in Table II receive no comparable semantic side information, so the comparison is unfair. This invalidates the central claim of 'superior performance in ... specific downstream tasks' and the average Dice advantage in Table II (0.8742 vs 0.8335).
  2. [Section IV-B, Table I] The statement that the proposed method 'consistently outperformed the baselines' is contradicted by Table I: DiSC-Med's LPIPS (0.1499) is worse than VQ-VAE-2's (0.1218). Only FID is better across the board. The reconstruction-quality claim is therefore only partially supported, and the abstract's phrase 'superior reconstruction performance' is too strong as written.
  3. [Section IV-D, Figs. 5-6] The noise-robustness results inherit the same information leakage as the noiseless downstream evaluation, since the Dice curves are computed on images reconstructed from the transmitted (possibly noisy) segmentation map. In addition, the figures report no error bars, confidence intervals, or significance tests, so the visual gaps in FID and Dice are not established as statistically reliable. The robustness claim needs a non-circular evaluation or a baseline that also receives the semantic side information before it can support the paper's conclusions.
minor comments (6)
  1. [Fig. 4] The figure legend uses 'Diff-Med' while the method is named DiSC-Med throughout the paper; please use a consistent name.
  2. [Reference [25]] Reference [25] is incomplete: it lacks author names and a full title. Please provide complete bibliographic information.
  3. [Section III-C2] The symbol c is used both for the channel noise condition embedding and for the denoised semantic volume {x'_seg, x'_edge}; the later use of c_k for the diffusion condition adds to the ambiguity. Please use distinct notations.
  4. [Section III-B2] The 'compression' step is implemented by fixed strided downsampling, but Section III-A describes it as a 'learned compression module.' Please clarify whether any parameters are learned in this stage.
  5. [Section IV-A] It is unclear whether the 186 3D CT volumes used to train the Denoising Module come from the same AMOS subset, a different cohort, or another dataset; please specify the overlap with the 16-patient evaluation set.
  6. [Section IV-C1] The metric 'mIoU w' is listed as 'weighted mean IoU' but the weighting scheme is not defined; please state the exact formula.

Circularity Check

1 steps flagged · score 6.0 of 10

Downstream segmentation evaluation is circular: the receiver conditions diffusion reconstruction on the segmentation map that the downstream test then re-segments, so Table II's Dice superiority reflects the transmitted labels rather than communication quality.

  1. self definitional [Section III-C3 (Eq. 4) and Section IV-C]
    "The restored 3D semantic volumes are converted into 2D slice-wise maps, i.e., ˆx′ seg(k) and ˆx′ edge(k) for the kth slice, which are concatenated to form the conditioning input: ck = Concat(ˆx′ seg(k), ˆx′ edge(k)). (4) ... To further evaluate the performance of data reconstruction at the receiver end, we tested the performance of regenerated CT images in anatomical structure segmentation."

    The transmitter extracts x_seg = f_seg(x) with TotalSegmentator (Sec. III-B1), transmits it, denoises it, and feeds it directly into the diffusion reconstruction as c_k = Concat(ˆx'_seg(k), ˆx'_edge(k)) (Eq. 4). Section IV-C then measures anatomical segmentation of the reconstructed image against ground-truth labels. Since the image is generated from the very segmentation labels the downstream task re-segments, Dice/mIoU quantify how well the blind model recovers the transmitted TotalSegmentator content, not communication/reconstruction quality. AE baselines get no such semantic side information, so the comparison is forced. Noisy-label Dice curves likewise merely propagate the channel's corruption of the conditioning map into the regenerated image.

full rationale

The image-fidelity comparisons (LPIPS/FID in Table I and Figures 5-6) are not circular; they compare reconstructed CT volumes pixel-wise and are independent of the conditioning-label leakage. However, the paper's headline claim of 'superior performance in ... specific downstream tasks' relies almost entirely on Table II and Figures 5-6, where the downstream task is anatomical segmentation. Because Eq. (4) makes the transmitted/denoised segmentation map the conditioning variable of the generative reconstruction, the segmentation evaluation reduces in large part to an agreement test between TotalSegmentator (the label source) and the blind segmentation model. This is information leakage by construction, not a fair comparison of transmission schemes. There is no load-bearing self-citation or imported uniqueness theorem in the paper, so the circularity is localized to the downstream evaluation; hence a score of 6 rather than 8 or 10.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The system relies on multiple pre-trained components (TotalSegmentator, MedSAM, a DDPM, and a segmentation evaluation model) whose assumptions are not tested independently. The main hand-chosen parameters are the strided-sampling rates, the CT intensity clipping range, and the diffusion architecture. No new entities are introduced.

free parameters (3)
  • Downsampling stride factors for semantic volumes = (2 along channel, 4 along height, 4 along width)
    Hand-chosen to set compression ratio and information loss; the paper reports compression ratio ~50.69x but does not optimize or justify these factors.
  • CT intensity clipping range = [-400, 400] HU
    Hand-chosen normalization before feeding to diffusion model; affects reconstruction quality and is not justified.
  • Diffusion model architecture size = U-Net with 5 down/up levels, channels 256-1024
    Hand-chosen capacity; no ablation.
assumptions (5)
  • domain assumption TotalSegmentator provides accurate anatomical segmentation of CT slices
    The transmitter's semantic representation relies on this pre-trained model; errors in x_seg are propagated to the receiver.
  • domain assumption MedSAM plus Canny edge detector produces edge maps that are useful conditions for a diffusion model
    No analysis shows the contribution of the edge stream.
  • domain assumption The blind segmentation model [29] provides an unbiased evaluation of the reconstructed images
    The model is pre-trained and claimed to have no prior exposure, but the transmitted segmentation map leaks the answer.
  • domain assumption The diffusion model can faithfully regenerate CT images from segmentation and edge conditions
    The method's success depends on the pre-trained DDPM's conditional generation quality, which is not independently verified.
  • domain assumption The test patients and the separate 186 volumes used to train the denoising module are representative and disjoint
    The paper does not specify the origin of the 186 volumes or verify the absence of data leakage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiSC-Med: Diffusion-based Semantic Communications for Robust Medical Image Transmission." pith.science (2026). https://pith.science/paper/7N27NKAP

@misc{pith2026250800172,
  author       = {Pith},
  title        = {Pith review of: DiSC-Med: Diffusion-based Semantic Communications for Robust Medical Image Transmission},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7N27NKAP}},
  note         = {Machine review of arXiv:2508.00172}
}
read the original abstract

The rapid development of artificial intelligence has driven smart health with next-generation wireless communication technologies, stimulating exciting applications in remote diagnosis and intervention. To enable a timely and effective response for remote healthcare, efficient transmission of medical data through noisy channels with limited bandwidth emerges as a critical challenge. In this work, we propose a novel diffusion-based semantic communication framework, namely DiSC-Med, for the medical image transmission, where medical-enhanced compression and denoising blocks are developed for bandwidth efficiency and robustness, respectively. Unlike conventional pixel-wise communication framework, our proposed DiSC-Med is able to capture the key semantic information and achieve superior reconstruction performance with ultra-high bandwidth efficiency against noisy channels. Extensive experiments on real-world medical datasets validate the effectiveness of our framework, demonstrating its potential for robust and efficient telehealth applications.

Figures

Figures reproduced from arXiv: 2508.00172 by the authors.

Figure 1
Figure 1. Overall framework of DiSC-Med: 1) Transmitter – The transmitter extracts the semantic representations from the original images, which are compressed to a transmission-friendly low-dimensional representation; 2) Receiver – The receiver interpolate and denoise the semantic representations, which serve as the conditions of the diffusion models for image reconstruction. communications. Utilizing the extracted semantic e… view at source ↗
Figure 2
Figure 2. Structure of the Denoising Module: DB stands for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Structure of the reconstruction module. slice-wise 2D U-Net as shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of reconstructed CT images. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance under AWGN channels: FID ↓ vs. SNR (left) and Dice ↑ vs. SNR (right). 10 −1 10 −1.310 −1.610 −1.910 −2.2 0 BER 100 150 200 250 300 350 400 FID DiSC-Med ResVAE DCAE CAE VQ-VAE-2 10 −1 10 −1.310 −1.610 −1.910 −2.2 0 BER 0.0 0.5 1.0 1.5 Dice DiSC-Med ResVAE DC…
Figure 6
Figure 6. Figure 6: Performance under bit-wise errors: FID ↓ vs. BER (left) and Dice ↑ vs. BER (right). V. CONCLUSION In this work, we proposed a semantic communication framework, DiSC-Med, for efficient medical image transmis￾sion and reconstruction using diffusion models. By leverag￾ing…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

  1. [1]

    6G technology in intelligent healthcare: smart health and its security and privacy perspectives,

    N. Arastouei and M. A. Khan, “6G technology in intelligent healthcare: smart health and its security and privacy perspectives,” IEEE Wireless Communications, vol. 32, no. 1, pp. 116–121, 2025

    work page 2025

  2. [2]

    Smart health in the 6g era: bringing security to future smart health services,

    E. Batista, P. L ´opez-Aguilar, and A. Solanas, “Smart health in the 6g era: bringing security to future smart health services,” IEEE Commu- nications Magazine , vol. 62, no. 6, pp. 74–80, 2024

    work page 2024

  3. [3]

    Domain knowledge driven se- mantic communication for image transmission over wireless channels,

    A. Li, X. Liu, G. Wang, and P. Zhang, “Domain knowledge driven se- mantic communication for image transmission over wireless channels,” IEEE Wireless Communications Letters, vol. 12, no. 1, pp. 55–59, 2023

    work page 2023

  4. [4]

    A review on medical image data compression techniques,

    G. Patidar, S. Kumar, and D. Kumar, “A review on medical image data compression techniques,” in 2nd International Conference on Data, Engineering and Applications (IDEA) , Bhopal, India, 2020, pp. 1–6

    work page 2020

  5. [5]

    Versatile approaches for medical image compression: a review,

    B. Patel and N. Doshi, “Versatile approaches for medical image compression: a review,” Journal of King Saud University, Computer and Information Sciences , vol. 32, no. 3, pp. 300–317, 2020

    work page 2020

  6. [6]

    Medical image compression based on variational autoencoder,

    X. Liu, L. Zhang, Z. Guo, T. Han, M. Ju, B. Xu, and H. Liu, “Medical image compression based on variational autoencoder,” Mathematical Problems in Engineering , vol. 2022, pp. 1–12, Dec. 2022

    work page 2022

  7. [7]

    X-ray image compression using variational auto-encoder,

    H. Xu, Y . Xue, J. Yang, and Z. Wu, “X-ray image compression using variational auto-encoder,” in 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI) , 2021, pp. 134–137

    work page 2021

  8. [8]

    Vari- ational image compression with a scale hyperprior,

    J. Ball ´e, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, “Vari- ational image compression with a scale hyperprior,” arXiv Preprint arXiv:1802.01436, 2018

    arXiv 2018

Show all 29 references
  1. [9]

    Fully convolutional model for variable bit length and lossy high density compression of mammograms,

    A. Kar, S. Phani Krishna Karri, N. Ghosh, R. Sethuraman, and D. Sheet, “Fully convolutional model for variable bit length and lossy high density compression of mammograms,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops , Salt Lake C...

    2018

  2. [10]

    Neuromorphologically- preserving volumetric data encoding using VQ-V AE,

    Y . Zhou, Y . Gao, R. Huang, and X. Zhang, “Neuromorphologically- preserving volumetric data encoding using VQ-V AE,” Medical Image Analysis, vol. 73, p. 102186, 2021

    2021

  3. [11]

    TransCS-net: a hybrid transformer-based privacy-protecting network using compressed sensing for medical image segmentation,

    S. Tang, C. F. Cheang, X. Yu, Y . Liang, Q. Feng, and Z. Chen, “TransCS-net: a hybrid transformer-based privacy-protecting network using compressed sensing for medical image segmentation,”Biomedical Signal Processing and Control , vol. 86, p. 105131, 2023

    2023

  4. [12]

    Diff-GO+: an efficient diffusion goal-oriented communication system with local feedback,

    A. Wijesinghe, S. Zhang, S. Wanninayaka, W. Wang, and Z. Ding, “Diff-GO+: an efficient diffusion goal-oriented communication system with local feedback,” IEEE Transactions on Wireless Communications , pp. 1–1, 2025

    2025

  5. [13]

    The long- term consequence of anterior cruciate ligament and meniscus injuries: osteoarthritis,

    L. S. Lohmander, M. Englund, L. Dahl, and E. Roos, “The long- term consequence of anterior cruciate ligament and meniscus injuries: osteoarthritis,” American Journal of Sports Medicine , vol. 35, pp. 1756– 69, Nov. 2007

    2007

  6. [14]

    Semantic communications: overview, open issues, and future research directions,

    X. Luo, H.-H. Chen, and Q. Guo, “Semantic communications: overview, open issues, and future research directions,” IEEE Wireless Communications, vol. 29, no. 1, pp. 210–219, 2022

    2022

  7. [15]

    Deep learning enabled semantic communication systems,

    H. Xie, Z. Qin, G. Y . Li, and B.-H. Juang, “Deep learning enabled semantic communication systems,” IEEE Transactions on Signal Pro- cessing, vol. 69, pp. 2663–2675, 2021

    2021

  8. [16]

    Robust semantic communications with masked VQ-V AE enabled codebook,

    Q. Hu, G. Zhang, Z. Qin, Y . Cai, G. Yu, and G. Y . Li, “Robust semantic communications with masked VQ-V AE enabled codebook,” IEEE Transactions on Wireless Communications , vol. 22, no. 12, pp. 8707–8722, 2023

    2023

  9. [17]

    Generative semantic communication: diffusion models beyond bit recovery,

    E. Grassucci, S. Barbarossa, and D. Comminiello, “Generative semantic communication: diffusion models beyond bit recovery,” arXiv Preprint arXiv:2306.04321, 2023

    2023 arXiv

  10. [18]

    Diff-GO: diffusion goal-oriented communications with ultra-high spectrum efficiency,

    A. Wijesinghe, S. Zhang, S. Wanninayaka, W. Wang, and Z. Ding, “Diff-GO: diffusion goal-oriented communications with ultra-high spectrum efficiency,” in 2024 IEEE International Conference on Com- munications Workshops, Denver, CO, USA, 2024, pp. 1079–1084

    2024

  11. [19]

    Token communications: a unified framework for cross-modal context-aware semantic communications,

    L. Qiao, M. B. Mashhadi, Z. Gao, R. Tafazolli, M. Bennis, and D. Niyato, “Token communications: a unified framework for cross-modal context-aware semantic communications,” arXiv Preprint arXiv:2502.12096, 2025

    2025 arXiv

  12. [20]

    TotalSegmen- tator: robust segmentation of 104 anatomical structures in CT images,

    J. M. Wasserthal, D. Zimmerer, L. Fidon, and Others, “TotalSegmen- tator: robust segmentation of 104 anatomical structures in CT images,” Nature Communications, vol. 14, no. 1, p. 3552, 2023

    2023

  13. [21]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,” Nature Communications , vol. 15, no. 1, p. 654, Jan. 2024

    2024

  14. [22]

    Denoising diffusion probabilistic mod- els,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic mod- els,” in Advances in Neural Information Processing Systems , vol. 33, Virtual, 2020, pp. 6840–6851

    2020

  15. [23]

    Amos: a large-scale abdominal multi- organ benchmark for versatile medical image segmentation,

    Y . Ji, H. Bai, C. Ge, J. Yang, Y . Zhu, R. Zhang, Z. Li, L. Zhanng, W. Ma, X. Wan, and Others, “Amos: a large-scale abdominal multi- organ benchmark for versatile medical image segmentation,” in Ad- vances in Neural Information Processing Systems , vol. 35, New Or- leans, LA,...

    2022

  16. [24]

    Convolutional autoencoder-based medical image compression using a novel annotated medical X-ray imaging dataset,

    A. Fettah, R. Menassel, A. Gattal, and A. Gattal, “Convolutional autoencoder-based medical image compression using a novel annotated medical X-ray imaging dataset,” Biomedical Signal Processing and Control, 2024

    2024

  17. [25]

    Generating diverse high-fidelity images with vq-vae-2,

    “Generating diverse high-fidelity images with vq-vae-2,” in Proceedings of the 33rd Neurips

  18. [26]

    Reducing the dimensionality of data with neural networks,

    G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006

    2006

  19. [27]

    Auto-encoding variational bayes,

    D. P. Kingma, M. Welling, and Others, “Auto-encoding variational bayes,” arXiv Prepreint arXiv:1312.6114 , 2014

    2014 arXiv

  20. [28]

    Neural discrete representa- tion learning,

    A. Van Den Oord, O. Vinyals, and Others, “Neural discrete representa- tion learning,” in Advances in Neural Information Processing Systems , vol. 30, Long Beach, CA, USA, 2017

    2017

  21. [29]

    Versatile medical image segmentation learned from multi-source datasets via model self-disambiguation,

    X. Chen, H. Zheng, Y . Li, Y . Ma, L. Ma, H. Li, and Y . Fan, “Versatile medical image segmentation learned from multi-source datasets via model self-disambiguation,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , Seattle W A, USA, 202...

    2024

Pith tools

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