Pith. sign in

REVIEW 3 major objections 7 minor 59 references

Gating U-Net fusion from the difference between decoder and encoder features beats standard attention.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Difference-driven gating (absolute feature difference or signed entropy difference) produces coupled maps that modulate both global and local U-Net features and outperform prior attention fusion on three modalities.

T0 review reviewed 2026-07-14 challenge →

load-bearing objection Solid engineering paper: difference-driven dual-stream gating beats selective/cross-attention on three modalities; the entropy-as-certainty story is under-supported but not load-bearing for the empirical claim. the 3 major comments →

arxiv 2607.11096 v1 pith:52JPWHGU submitted 2026-07-13 cs.CV cs.SDstat.ML

Difference-Driven Gating: Adaptive Feature Fusion for U-Net Decoder

Kai Li , Xuechao Zou , Jiashen Fu , Zijun Yan , Xintong Wang , Xiaolin Hu This is my paper
classification cs.CV cs.SDstat.ML
keywords U-Netfeature fusionentropy-difference gatingdifference-driven attentionmedical image segmentationcloud removalspeech separationmulti-scale fusion
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

U-Net-style networks rebuild fine detail by fusing high-level decoder features with low-level encoder features. Most attention methods either gate the encoder stream from the decoder alone or from their correlation. This paper instead derives the gates from the difference between the two streams, producing coupled maps that modulate both sides at once. Feature-difference gating uses absolute feature disagreement; entropy-difference gating uses the signed difference of Shannon entropy so the more certain stream wins. Across medical segmentation, multi-temporal cloud removal, and speech separation, both modules improve over selective and cross attention, with entropy-difference gating strongest, at only modest extra cost. The claim is that difference-driven, dual-stream gating is a more effective multi-scale fusion paradigm for encoder-decoder models.

Core claim

Deriving coupled gating maps from the difference between global decoder and local encoder streams—either absolute feature difference or signed entropy difference—and modulating both streams simultaneously yields higher accuracy than selective-attention or cross-attention fusion on medical image segmentation, multi-temporal cloud removal, and speech separation, establishing difference-driven gating as a practical multi-scale fusion paradigm for U-Net-style architectures.

What carries the argument

Entropy-difference gating (EDG): Shannon entropy of softmax-normalized intermediate activations measures each stream's representational certainty; the signed entropy difference supplies channel-wise and spatiotemporal gates that bias fusion toward the lower-entropy stream while the global and local content features are both scaled and added.

Load-bearing premise

The paper treats Shannon entropy of softmax-normalized intermediate features as a faithful, task-agnostic measure of which stream is more certain and should therefore dominate the fusion.

What would settle it

Replace entropy with mean, variance, L2, or a learnable projection while keeping the rest of the dual-stream gating fixed; if those alternatives match or beat entropy-difference gating on the same three tasks, the claim that entropy specifically captures useful certainty collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Existing U-Net, TransAttUNet, PMAA, TDANet and TIGER models can replace their fusion blocks with EDG and expect measurable gains on Synapse/ACDC, Sen2MTC and LRS2/EchoSet without large parameter increases.
  • Dual-stream modulation (gating both G and L) is shown to be superior to the common single-stream practice of only modulating the encoder features.
  • Channel-wise and spatiotemporal certainty supply complementary signals whose relative value is task-dependent (spatial for vision, channel for speech).
  • The same lightweight difference-driven module is portable across vision and audio U-Net variants, supporting a modality-agnostic fusion recipe.

Where Pith is reading between the lines

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

  • If entropy truly tracks reliability, the same signed-difference construction could be tried as a drop-in fusion block in non-U-Net multi-scale designs such as FPN or feature-pyramid Transformers.
  • The neuroscience-inspired predictive-coding framing suggests testing whether the magnitude of the difference itself could serve as an auxiliary training signal or uncertainty map, not only as a gate.
  • A natural next measurement is whether EDG still helps when both streams are already heavily regularized or when one stream is deliberately degraded (e.g., heavy cloud or high reverberation).
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes a difference-driven paradigm for multi-scale feature fusion in U-Net decoders. Instead of deriving attention from global features alone (selective attention) or from pairwise correlations (cross-attention), it derives coupled gating maps from the difference between top-down global (G) and bottom-up local (L) streams. Two modules are introduced: Feature-Difference Gating (FDG), which uses absolute differences of mean-pooled features, and Entropy-Difference Gating (EDG), which uses signed Shannon-entropy differences of softmax-normalized activations as a proxy for relative representational certainty. Both modules modulate G and L simultaneously. Controlled drop-in replacements of the fusion block, with mean±std over multiple runs on Synapse and evaluations on ACDC, Sen2MTC cloud removal, and LRS2-2Mix/EchoSet speech separation, show consistent gains over summation, concatenation, selective attention, and cross-attention, with EDG strongest.

Significance. If the empirical pattern holds, the work offers a lightweight, architecture-agnostic fusion primitive that improves three distinct dense-prediction modalities with modest parameter/MAC overhead. Strengths include: (i) controlled replacement of only the fusion module while freezing the host architecture; (ii) multi-run statistics on Synapse; (iii) ablations of gating dimensions, dual- vs single-stream modulation, and aggregation functions (Table 11); and (iv) demonstration that even the simpler FDG already beats standard attention fusions. These elements make the contribution more than a single-task attention tweak and support the claim of a reusable multi-scale fusion design pattern for U-Net-style models.

major comments (3)
  1. Sec. 3.2.2 (Eqs. 4–7, 12–13) and the EDG narrative treat Shannon entropy of softmax-normalized intermediate activations as a measure of “representational certainty,” so that signed Δ = ĦG − ĦL correctly indicates which stream should dominate. The only direct support is the qualitative boundary heatmap in Fig. 10 and the fact that entropy wins the aggregator bake-off in Table 11. No quantitative correlation is reported between per-location (or per-channel) entropy and any reliability proxy (local Dice residual, reconstruction error, SI-SNR contribution, or agreement with ground truth). Without that link, the claimed inductive bias remains an interpretation of a useful non-linearity. Please either (a) add a quantitative reliability analysis, or (b) reframe EDG as an empirically superior aggregator within the difference-driven framework and tone down certainty language in the abstract, intr
  2. Sec. 4.1.2 and the subsequent Remark assert that FDG “consistently outperformed existing methods” on medical segmentation, cloud removal, and speech separation, yet full FDG numbers appear only for U-Net on Synapse (Tables 1–2). All later tables report only EDG. Because the paper’s central paradigm claim is difference-driven gating (of which FDG is the purest, entropy-free instance), FDG results (mean metrics and preferably mean±std) should be reported for at least one cloud-removal and one speech-separation setting under the same controlled fusion-replacement protocol. Otherwise the multi-task support for the paradigm rests on an unreported claim.
  3. Sec. 3.3.2 (Eqs. 14–15) hard-codes an FDG inductive bias that large absolute disagreement biases fusion toward the local stream (σ(Δ) ≥ 0.5). This prior is never ablated against the opposite bias or a symmetric signed feature difference. Given that EDG’s advantage is partly attributed to data-driven signed gating, a short ablation (e.g., reverse-FDG or signed feature difference without entropy) on Synapse would clarify whether the local-preferring prior is necessary or incidental to FDG’s gains in Table 1.
minor comments (7)
  1. Fig. 3 caption and Sec. 3.1: the FE/GG pipeline is clear, but the role of the third content branch (Xd) is easy to miss; a one-sentence reminder that content features are refined by 1×1+BN+GELU and are not used to form Δ would help.
  2. Table 5: TransAttUNet-EDG reduces parameters from 25.97M to 14.39M and MACs from 68.01G to 14.14G by replacing heavy fusion layers. Please state explicitly in the text that part of the efficiency gain is architectural simplification, not only EDG’s own cost, so readers do not attribute the full reduction to the gating module.
  3. Eqs. (4) and (6): ε is added only in the denominator of the softmax; standard practice also clamps the argument of log. Clarify numerical handling of log(PX) when PX is near zero.
  4. Fig. 7(b): dashed arrows indicate a shared top-level global feature fed to every EDG instance; this multi-scale reuse differs from the per-stage (G,L) pairing in Fig. 7(a) and should be stated in the main text of Sec. 3.4.1, not only in the caption.
  5. Related work (Sec. 2) omits several recent U-Net fusion variants (e.g., gated skip connections, bi-directional attention U-Nets, and predictive-coding-inspired vision models beyond the cited [19–21]). A short paragraph situating difference-driven gating against these would strengthen positioning.
  6. Typos / polish: “Zijun Y an” in the author list; “T raining” with a space in several subsection headings; “Fréechet” / encoding glitches in the FID discussion; inconsistent hyphenation of “Sen2 MTC” vs “Sen2MTC”.
  7. Code availability is promised “upon acceptance”; for reproducibility of the multi-run Synapse protocol and the three-task suite, a public repository (or anonymized supplement) at revision time would be valuable.

Circularity Check

0 steps flagged

No circularity: modules are independently defined architectural components whose superiority is measured empirically on held-out public benchmarks, not derived by construction from their own inputs.

full rationale

The paper defines FDG (absolute feature difference, Eqs. 10-11) and EDG (signed Shannon entropy difference after softmax, Eqs. 4-7 and 12-13) as drop-in fusion blocks that produce coupled gates (Eqs. 14-17). These definitions do not refer to the evaluation metrics (DSC, HD95, PSNR, SSIM, SI-SDRi) or to any fitted free parameters that later reappear as 'predictions.' Performance claims rest on controlled replacements inside existing U-Net hosts (U-Net, TransAttUNet, PMAA, TDANet, TIGER) evaluated on public held-out splits (Synapse, ACDC, Sen2MTC, LRS2-2Mix, EchoSet). Self-citations are limited to the authors' prior host architectures that serve as experimental substrates; they do not supply uniqueness theorems, inductive biases, or numerical values that force the reported gains. Ablations (Tables 10-11, Fig. 13) vary aggregation functions and gating dimensions while keeping the dual-stream fusion fixed, confirming that the gains are not tautological. The neuroscience inspiration (predictive coding) is motivational only and is not used as a load-bearing derivation step. Consequently the claimed 'new paradigm' is an empirical architectural claim, not a circular reduction of outputs to inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The paper is an empirical architecture paper. Its central claim rests on standard deep-learning practice plus two domain-level modeling choices (difference as a useful gating signal; entropy as certainty) rather than on free parameters fitted to the target metrics or on newly postulated physical entities.

free parameters (2)
  • numerical stability epsilon
    Small constant added inside softmax and normalization (Eqs. 4, 6, 16); value not stated but conventional and does not drive the performance claim.
  • loss weighting 0.6 Dice + 0.4 CE
    Fixed joint objective for medical segmentation; chosen by convention rather than grid-searched on the test set.
axioms (4)
  • ad hoc to paper Shannon entropy of a softmax distribution over channels or positions quantifies representational certainty of a feature stream.
    Introduced in Sec. 3.2.2 without independent theoretical derivation; treated as self-evident for gating design.
  • ad hoc to paper Absolute or signed difference between compressed global and local features is a more informative fusion cue than raw features or pairwise correlation.
    Core design hypothesis of Sec. 3; justified by predictive-coding analogy and by empirical ablation, not by prior theorem.
  • domain assumption Standard U-Net encoder-decoder with skip connections is an appropriate host architecture for testing fusion modules.
    Taken from the large body of U-Net literature cited in Sec. 1-2.
  • domain assumption Nearest-neighbor upsampling and 1x1 convolutions suffice for resolution and channel alignment before gating.
    Common engineering practice restated in Sec. 3.3.
invented entities (2)
  • Feature-difference gating (FDG) module no independent evidence
    purpose: Generate dual-stream gates from absolute difference of mean-pooled features.
    New architectural block defined in Sec. 3; no independent existence outside the paper.
  • Entropy-difference gating (EDG) module no independent evidence
    purpose: Generate dual-stream gates from signed entropy difference of the two streams.
    New architectural block defined in Sec. 3; no independent existence outside the paper.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Difference-Driven Gating: Adaptive Feature Fusion for U-Net Decoder." pith.science (2026). https://pith.science/paper/52JPWHGU

@misc{pith2026260711096,
  author       = {Pith},
  title        = {Pith review of: Difference-Driven Gating: Adaptive Feature Fusion for U-Net Decoder},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/52JPWHGU}},
  note         = {Machine review of arXiv:2607.11096}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The U-Net style models have been widely used in many applications. A critical step in these models is to reconstruct the lower-level features using a top-down decoder. This reconstruction requires precise fusion of high-level semantics and low-level details. Existing attention-based fusion methods typically derive attention weights from the top-down decoder features (global) alone or the correlation between the top-down decoder features and the bottom-up encoder features (local), then modulate the encoder features using these weights. In this work, we explore a different paradigm: deriving attention weights from the difference between the two feature streams. To this end, we propose two difference-based gating approaches: Feature-difference gating (FDG), which directly uses the absolute difference between global and local features to generate adaptive gating maps, and Entropy-difference gating (EDG), which measures the representational certainty of each stream via information entropy and uses their signed entropy difference to derive the attention weights. Both methods produce coupled gating maps that simultaneously modulate the global and local features. Experiments on different tasks including medical image segmentation, remote sensing image cloud removal and speech separation showed that both methods outperformed existing attention-based fusion methods, and EDG performed better. The results suggested a new paradigm for multi-scale feature fusion in the U-Net style structures.

Figures

Figures reproduced from arXiv: 2607.11096 by Jiashen Fu, Kai Li, Xiaolin Hu, Xintong Wang, Xuechao Zou, Zijun Yan.

Figure 1
Figure 1. Figure 1: Overview of U-Net architecture with integrated fusion modules. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic illustration of four representative decoder fusion methods: (a) summation; (b) concatenation; (c) selective attention; (d) cross [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overall pipeline of the FE block. spatiotemporal dimensions to derive adaptive gating maps. Finally, the upsampled global content feature Gˆ d and the local content feature Lˆ d are modulated by the resulting gates and fused to yield F ∈ R C×P ′ . 3.2 Feature Extraction Block The FE block operates on a single feature stream at a time. It produces two complementary compressed features for the subsequent GG … view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of the GG block. In the FDG module, [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overall processing pipelines for the medical image segmentation [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Network architectures for two vision tasks. (a) Medical image segmentation based on U-Net [ [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: The diagram illustrates the separator structure of TDANet [ [PITH_FULL_IMAGE:figures/full_fig_p006_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual analysis of the per-position entropy ( [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of different methods on the Synapse dataset. From left to right: (a) ground truth, (b) U-Net-EDG (ours), (c) U-Net [ [PITH_FULL_IMAGE:figures/full_fig_p009_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visual comparison of cloud-removal results produced by different methods on two benchmarks: (a) Sen2 [PITH_FULL_IMAGE:figures/full_fig_p010_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Ablation of channel-only (EDG-ch), spatiotemporal-only (EDG [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗

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

59 extracted references · 1 linked inside Pith

  1. [1]

    U-Net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted In- tervention. Springer, 2015, pp. 234–241

    2015

  2. [2]

    Attention U-Net: learning where to look for the pancreas,

    O. Oktay, J. Schlemper, L. Le Folgoc, M. C. H. Lee, M. P . Heinrich, K. Misawa, K. Mori, S. G. McDonagh, N. Y. Hammerla, B. Kainz, B. Glocker, and D. Rueckert, “Attention U-Net: learning where to look for the pancreas,” inMedical Imaging with Deep Learning, 2018

    2018

  3. [3]

    Medical image segmentation review: the success of U-Net,

    R. Azad, E. K. Aghdam, A. Rauland, Y. Jia, A. H. Avval, A. Bozorg- pour, S. Karimijafarbigloo, J. P . Cohen, E. Adeli, and D. Merhof, “Medical image segmentation review: the success of U-Net,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  4. [4]

    MA-Unet: an improved version of Unet based on multi-scale and attention mechanism for medical image segmentation,

    Y. Cai and Y. Wang, “MA-Unet: an improved version of Unet based on multi-scale and attention mechanism for medical image segmentation,” inThird International Conference on Electronics and Communication; Network and Computer Technology, 2020

    2020

  5. [5]

    PMAA: a progressive multi-scale attention autoencoder model for high-performance cloud removal from multi-temporal satellite imagery,

    X. Zou, K. Li, J. Xing, P . Tao, and Y. Cui, “PMAA: a progressive multi-scale attention autoencoder model for high-performance cloud removal from multi-temporal satellite imagery,” inEuropean Conference on Artificial Intelligence, 2023, pp. 3165–3172

    2023

  6. [6]

    An effective thin cloud removal procedure for visible remote sensing images,

    H. Shen, H. Li, Y. Qian, L. Zhang, and Q. Yuan, “An effective thin cloud removal procedure for visible remote sensing images,” ISPRS Journal of Photogrammetry and Remote Sensing, vol. 96, pp. 224–235, 2014

    2014

  7. [7]

    An efficient encoder-decoder ar- chitecture with top-down attention for speech separation,

    K. Li, R. Yang, and X. Hu, “An efficient encoder-decoder ar- chitecture with top-down attention for speech separation,” in International Conference on Learning Representations, 2022

    2022

  8. [8]

    TIGER: Time-frequency interleaved gain extraction and reconstruction for efficient speech separation,

    M. Xu, K. Li, G. Chen, and X. Hu, “TIGER: Time-frequency interleaved gain extraction and reconstruction for efficient speech separation,” inInternational Conference on Learning Representations, 2024

    2024

  9. [9]

    Hierarchical models in the brain,

    K. Friston, “Hierarchical models in the brain,”PLoS Computational Biology, vol. 4, no. 11, p. e1000211, 2008

    2008

  10. [10]

    K. H. Pribram,Brain and perception: holonomy and structure in figural processing. Psychology Press, 2013

    2013

  11. [11]

    Auto-encoders in deep learning—a review with new perspectives,

    S. Chen and W. Guo, “Auto-encoders in deep learning—a review with new perspectives,”Mathematics, vol. 11, no. 8, p. 1777, 2023

    2023

  12. [12]

    Contributions of low- and high-level properties to neural processing of visual scenes in the human brain,

    I. I. Groen, E. H. Silson, and C. I. Baker, “Contributions of low- and high-level properties to neural processing of visual scenes in the human brain,”Philosophical Transactions of the Royal Society B: Biological Sciences, vol. 372, no. 1714, p. 20160102, 2017

    2017

  13. [13]

    Auditory brain stem response to complex sounds: a tutorial,

    E. Skoe and N. Kraus, “Auditory brain stem response to complex sounds: a tutorial,”Ear and Hearing, vol. 31, no. 3, pp. 302–324, 2010

    2010

  14. [14]

    Speech separation using an asynchronous fully recurrent convo- lutional neural network,

    X. Hu, K. Li, W. Zhang, Y. Luo, J.-M. Lemercier, and T. Gerkmann, “Speech separation using an asynchronous fully recurrent convo- lutional neural network,”Advances in Neural Information Processing Systems, vol. 34, pp. 22 509–22 522, 2021. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 14

    2021

  15. [15]

    Sudo RM -RF: efficient networks for universal audio source separation,

    E. Tzinis, Z. Wang, and P . Smaragdis, “Sudo RM -RF: efficient networks for universal audio source separation,” in2020 IEEE 30th International Workshop on Machine Learning for Signal Processing. IEEE, 2020, pp. 1–6

    2020

  16. [16]

    Swin-Unet: Unet-like pure Transformer for medical image segmentation,

    H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-Unet: Unet-like pure Transformer for medical image segmentation,” inProceedings of the European Conference on Computer Vision Workshops, 2022

    2022

  17. [17]

    The cortical organization of speech processing,

    G. Hickok and D. Poeppel, “The cortical organization of speech processing,” vol. 8, no. 5, pp. 393–402. [Online]. Available: https://www.nature.com/articles/nrn2113

  18. [18]

    Canonical microcircuits for predictive coding,

    A. M. Bastos, W. M. Usrey, R. A. Adams, G. R. Mangun, P . Fries, and K. J. Friston, “Canonical microcircuits for predictive coding,” Neuron, vol. 76, no. 4, pp. 695–711, Nov. 2012

    2012

  19. [19]

    Predictive coding in the visual cor- tex: a functional interpretation of some extra-classical receptive- field effects,

    R. P . N. Rao and D. H. Ballard, “Predictive coding in the visual cor- tex: a functional interpretation of some extra-classical receptive- field effects,”Nature Neuroscience, vol. 2, no. 1, pp. 79–87, Jan. 1999

    1999

  20. [20]

    Predictive coding as a model of response prop- erties in cortical area V1,

    M. W. Spratling, “Predictive coding as a model of response prop- erties in cortical area V1,”The Journal of Neuroscience, vol. 30, no. 9, pp. 3531–3543, Mar. 2010

    2010

  21. [21]

    Deep predictive coding network with local recurrent processing for object recognition,

    K. Han, H. Wen, Y. Zhang, D. Fu, E. Culurciello, and Z. Liu, “Deep predictive coding network with local recurrent processing for object recognition,” inAdvances in Neural Information Processing Systems, May 2018

    2018

  22. [22]

    A mathematical theory of communication,

    C. E. Shannon, “A mathematical theory of communication,”The Bell System Technical Journal, vol. 27, no. 3, pp. 379–423, 1948

    1948

  23. [23]

    UNet++: A nested U-Net architecture for medical image segmen- tation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: A nested U-Net architecture for medical image segmen- tation,” inDeep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support. Springer, 2018, pp. 3–11

    2018

  24. [24]

    ResUNet-a: a deep learning framework for semantic segmentation of remotely sensed data,

    F. I. Diakogiannis, F. Waldner, P . Caccetta, and C. Wu, “ResUNet-a: a deep learning framework for semantic segmentation of remotely sensed data,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 162, pp. 94–114, 2020

    2020

  25. [25]

    MA-Net: a multi-scale attention network for liver and tumor segmentation,

    T. Fan, G. Wang, Y. Li, and H. Wang, “MA-Net: a multi-scale attention network for liver and tumor segmentation,”IEEE Access, vol. 8, pp. 179 656–179 665, 2020

    2020

  26. [26]

    Batch normalization: accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: accelerating deep network training by reducing internal covariate shift,” inPro- ceedings of the 32nd International Conference on Machine Learning. PMLR, 2015, pp. 448–456

    2015

  27. [27]

    Gaussian error linear units (GELUs),

    D. Hendrycks and K. Gimpel, “Gaussian error linear units (GELUs),” 2016

    2016

  28. [28]

    TransAt- tUnet: multi-level attention-guided U-Net with Transformer for medical image segmentation,

    B. Chen, Y. Liu, Z. Zhang, G. Lu, and A. W. K. Kong, “TransAt- tUnet: multi-level attention-guided U-Net with Transformer for medical image segmentation,”IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 8, no. 1, pp. 55–68, 2024

    2024

  29. [29]

    Exfuse: enhanc- ing feature fusion for semantic segmentation,

    Z. Zhang, X. Zhang, C. Peng, X. Xue, and J. Sun, “Exfuse: enhanc- ing feature fusion for semantic segmentation,” inProceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 269–284

    2018

  30. [30]

    UNeXt: MLP-based rapid medical image segmentation network,

    J. M. J. Valanarasu and V . M. Patel, “UNeXt: MLP-based rapid medical image segmentation network,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 2022

    2022

  31. [31]

    UNet++: a nested U-Net architecture for medical image segmen- tation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “UNet++: a nested U-Net architecture for medical image segmen- tation,” 2018

    2018

  32. [32]

    Filmy cloud removal on satel- lite imagery with multispectral conditional generative adversarial nets,

    K. Enomoto, K. Sakurada, W. Wang, H. Fukui, M. Matsuoka, R. Nakamura, and N. Kawaguchi, “Filmy cloud removal on satel- lite imagery with multispectral conditional generative adversarial nets,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2017, pp. 48–56

    2017

  33. [33]

    Image-to-image trans- lation with conditional adversarial networks,

    P . Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image trans- lation with conditional adversarial networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017, pp. 1125–1134

    2017

  34. [34]

    A multi-temporal convolutional au- toencoder neural network for cloud removal in remote sens- ing images,

    W. Sintarasirikulchai, T. Kasetkasem, T. Isshiki, T. Chanwimalu- ang, and P . Rakwatin, “A multi-temporal convolutional au- toencoder neural network for cloud removal in remote sens- ing images,” inInternational Conference on Electrical Engineer- ing/Electronics, Computer, Telecommunications and Information Tech- nology, 2018, pp. 360–363

    2018

  35. [35]

    Thick clouds removing from multitemporal Landsat images using spa- tiotemporal neural networks,

    Y. Chen, Q. Weng, L. Tang, X. Zhang, M. Bilal, and Q. Li, “Thick clouds removing from multitemporal Landsat images using spa- tiotemporal neural networks,”IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2020

    2020

  36. [36]

    Cloud removal in Sentinel-2 imagery using a deep residual neural network and SAR- optical data fusion,

    A. Meraner, P . Ebel, X. X. Zhu, and M. Schmitt, “Cloud removal in Sentinel-2 imagery using a deep residual neural network and SAR- optical data fusion,”ISPRS Journal of Photogrammetry and Remote Sensing, vol. 166, pp. 333–346, 2020

    2020

  37. [37]

    Cloud removal in satellite images using spatiotemporal generative networks,

    V . Sarukkai, A. Jain, B. Uzkent, and S. Ermon, “Cloud removal in satellite images using spatiotemporal generative networks,” in Winter Conference on Applications of Computer Vision, 2020, pp. 1796– 1805

    2020

  38. [38]

    CTGAN: cloud Transformer gener- ative adversarial network,

    G.-L. Huang and P .-Y. Wu, “CTGAN: cloud Transformer gener- ative adversarial network,” inInternational Conference on Image Processing. IEEE, 2022, pp. 511–515

    2022

  39. [39]

    SEN12MS-CR-TS: a remote-sensing data set for multimodal multitemporal cloud re- moval,

    P . Ebel, Y. Xu, M. Schmitt, and X. X. Zhu, “SEN12MS-CR-TS: a remote-sensing data set for multimodal multitemporal cloud re- moval,”IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2022

    2022

  40. [40]

    Un- CRtainTS: uncertainty quantification for cloud removal in optical satellite time series,

    P . Ebel, V . S. F. Garnot, M. Schmitt, J. Wegner, and X. X. Zhu, “Un- CRtainTS: uncertainty quantification for cloud removal in optical satellite time series,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023, pp. 2086– 2096

    2023

  41. [41]

    Denoising diffusion probabilistic feature-based network for cloud removal in Sentinel-2 imagery,

    R. Jing, F. Duan, F. Lu, M. Zhang, and W. Zhao, “Denoising diffusion probabilistic feature-based network for cloud removal in Sentinel-2 imagery,”Remote Sensing, vol. 15, no. 9, 2023

    2023

  42. [42]

    Deep learning techniques for auto- matic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?

    O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P .-A. Heng, I. Cetin, K. Lekadir, O. Camara, M. A. Gonzalez Ballester, G. Sanroma, S. Napel, S. Petersen, G. Tziritas, E. Grinias, M. Khened, V . A. Kollerathu, G. Krishnamurthi, M.-M. Roh ´e, X. Pennec, M. Sermesant, F. Isensee, P . J ¨ager, K. H. Maier-Hein, P . M. Full, I. Wolf, S. Engelhardt,...

    2018

  43. [43]

    TransUNet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, and Y. Zhou, “TransUNet: Transformers make strong encoders for medical image segmentation,”Computing Research Repository, vol. abs/2102.04306, 2021

    Pith/arXiv arXiv 2021

  44. [44]

    Decoupled weight decay regulariza- tion,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regulariza- tion,” inInternational Conference on Learning Representations, 2018, pp. 1–18

    2018

  45. [45]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P . Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595

    2018

  46. [46]

    The Fr ´echet distance between multivariate normal distributions,

    D. C. Dowson and B. V . Landau, “The Fr ´echet distance between multivariate normal distributions,”Journal of Multivariate Analysis, vol. 12, no. 3, pp. 450–455, 1982

    1982

  47. [47]

    Deep audio-visual speech recognition,

    T. Afouras, J. S. Chung, A. Senior, O. Vinyals, and A. Zisserman, “Deep audio-visual speech recognition,”IEEE Transactions on Pat- tern Analysis and Machine Intelligence, vol. 44, no. 12, pp. 8717–8727, 2022

    2022

  48. [48]

    CSR-I (WSJ0) complete,

    J. S. Garofolo, D. Graff, J. M. Baker, D. Paul, and D. Pallett, “CSR-I (WSJ0) complete,” 2007

    2007

  49. [49]

    Librispeech: an ASR corpus based on public domain audio books,

    V . Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: an ASR corpus based on public domain audio books,” inIEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2015, pp. 5206–5210

    2015

  50. [50]

    Deep cluster- ing: discriminative embeddings for segmentation and separation,

    J. R. Hershey, Z. Chen, J. Le Roux, and S. Watanabe, “Deep cluster- ing: discriminative embeddings for segmentation and separation,” inIEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2016, pp. 31–35

    2016

  51. [51]

    SonicSim: a customizable simulation platform for speech processing in moving sound source scenarios,

    K. Li, W. Sang, C. Zeng, R. Yang, G. Chen, and X. Hu, “SonicSim: a customizable simulation platform for speech processing in moving sound source scenarios,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  52. [52]

    Matterport3D: learning from RGB-D data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niebner, M. Savva, S. Song, A. Zeng, and Y. Zhang, “Matterport3D: learning from RGB-D data in indoor environments,” inInternational Conference on 3D Vision, 2017

    2017

  53. [53]

    Adam: a method for stochastic optimiza- tion,

    D. P . Kingma and J. Ba, “Adam: a method for stochastic optimiza- tion,” inInternational Conference on Learning Representations, 2015

    2015

  54. [54]

    SDR — half-baked or well done?

    J. Le Roux, S. Wisdom, H. Erdogan, and J. R. Hershey, “SDR — half-baked or well done?” inIEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2019, pp. 626–630

    2019

  55. [55]

    Performance measure- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 15 ment in blind audio source separation,

    E. Vincent, R. Gribonval, and C. F ´evotte, “Performance measure- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 15 ment in blind audio source separation,”IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 4, pp. 1462–1469, 2006

    2015

  56. [56]

    Conv-TasNet: surpassing ideal time- frequency magnitude masking for speech separation,

    Y. Luo and N. Mesgarani, “Conv-TasNet: surpassing ideal time- frequency magnitude masking for speech separation,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 8, pp. 1256–1266, 2019

    2019

  57. [57]

    Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation,

    Y. Luo, Z. Chen, and T. Yoshioka, “Dual-path RNN: efficient long sequence modeling for time-domain single-channel speech separation,” inIEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2020, pp. 46–50

    2020

  58. [58]

    Music source separation with band-split RNN,

    Y. Luo and J. Yu, “Music source separation with band-split RNN,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 31, pp. 1893–1901, 2023

    1901

  59. [59]

    TF-GRIDNET: making time-frequency domain models great again for monaural speaker separation,

    Z.-Q. Wang, S. Cornell, S. Choi, Y. Lee, B.-Y. Kim, and S. Watan- abe, “TF-GRIDNET: making time-frequency domain models great again for monaural speaker separation,” inIEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2023, pp. 1–5. Kai Li(Student Member, IEEE) received the B.S. degree from the Department of Computer Technolo...

    2023

This paper was first reviewed by grok-4.5 on July 14, 2026.