Pith. sign in

REVIEW 4 major objections 7 minor 66 references

A training-free channel-fusion method, MIS-HCC, compresses medical segmentation models to 12.5% of their size while keeping segmentation accuracy within 0.14% of the uncompressed model on 3D CT data and slightly improving it on some 2D ultr

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 18:17 UTC pith:WCIOZ5QG

load-bearing objection Plausible training-free channel pruning with near-lossless 87.5% compression, but the central claims hinge on single runs and an underspecified Wasserstein distance. the 4 major comments →

arxiv 2607.17329 v1 pith:WCIOZ5QG submitted 2026-07-19 cs.CV

MIS-HCC: Hierarchical Channel Clustering for Efficient Medical Image Segmentation

Bo Zhao , Haoran Yu , Lifei Liu , Zongcheng Chu , Yining Liu , Chang Liu , Szu-Yu Chen , Zequn Xie This is my paper
classification cs.CV
keywords medical image segmentationmodel compressionchannel pruninghierarchical clusteringWasserstein distanceparameter fusionU-NetMedSAM2
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.

MIS-HCC claims that the right way to compress medical segmentation networks is not to delete 'redundant' channels but to identify groups of near-duplicate channels, via Wasserstein distance and hierarchical clustering, and replace each group by its average. At an 87.5% reduction in parameters and FLOPs, the method keeps 3D abdominal-organ segmentation (MedSAM2 on FLARE21) within 0.14% Dice of the uncompressed model, and on several 2D datasets it slightly beats the uncompressed U-Net. This matters because medical models must fit on resource-limited devices, and existing pruning methods lose fine anatomical detail under aggressive compression. The paper argues that medical image features are highly redundant, so fusion preserves what deletion throws away.

Core claim

The paper's central claim is that channel-level redundancy in pretrained medical segmentation models is structured enough that hierarchical clustering followed by arithmetic weight averaging yields a compressed network that retains near-baseline accuracy without any retraining. The evidence: on the 3D FLARE21 benchmark with a MedSAM2 backbone, MIS-HCC scores Dice 0.8471 and IoU 0.7789 at 87.5% pruning, against 0.8483/0.7807 for the uncompressed model; on BUSI, DSB, and ISIC with U-Net and U-Net++, it matches or slightly exceeds the original in most configurations. The authors interpret this as showing that representing and exploiting inter-channel similarity, rather than keeping or dropping

What carries the argument

The machinery is a pairwise channel similarity matrix built from the Wasserstein distance between flattened channel weight vectors (Equation 1), converted to similarities (Equation 2). Agglomerative hierarchical clustering with average linkage (Equation 3) partitions the Nin channels of each layer into K clusters, and Equation 4 replaces each cluster by the arithmetic mean of its member weight vectors. What it does: it converts a discrete keep-or-drop decision into a continuous fusion decision, so compression is closed-form and training-free, and it applies to any layer whose weights have an input-channel dimension.

Load-bearing premise

The method works only if channels that the clustering groups together really are near-duplicates, so that averaging them loses little information; if a cluster accidentally mixes complementary filters, the fused model's accuracy collapses.

What would settle it

Take a pretrained U-Net layer, feed a batch of real ultrasound images, and collect the activation maps of the channels that MIS-HCC places in a single cluster. If the arithmetic mean of those activation maps is far (in, say, L2 distance) from several of the individual maps, then the 'near-duplicate' premise fails and fusion would be destroying information; if the mean is close to all of them, the method's core assumption is confirmed.

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

If this is right

  • At 87.5% compression, medical segmentation accuracy can be preserved within 0.14% Dice on 3D CT and even improved on some 2D ultrasound datasets, without retraining.
  • The method works across CNN (U-Net, U-Net++) and Transformer-style (MedSAM2) backbones in both 2D and 3D, suggesting the redundancy is architectural, not dataset-specific.
  • Because fusion only averages existing weights, compressed models can be produced on the fly from any pretrained model, with no training data or gradient computation.
  • At compression rates of 50%, 75%, and 87.5%, the fused model consistently beats the trainability-preserving pruning baseline TPP in Dice while achieving equal or better FLOP/parameter reduction at the highest rate.

Where Pith is reading between the lines

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

  • If weight-space similarity implies functional redundancy, then the same fusion trick should transfer to other dense prediction tasks (e.g., natural-image segmentation, depth estimation) where internal channels are also redundant; that is a testable extension the paper does not make.
  • The 2D results where compression slightly improves Dice suggest averaging may act as a mild regularizer; one could test this by measuring whether the improvement disappears when the original model is trained with stronger regularization.
  • The paper never specifies how the channel weight vectors induce the probability measures used in the Wasserstein distance; pinning down that choice would be the first reproducibility hurdle and could change which channels are fused.

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

4 major / 7 minor

Summary. The paper proposes MIS-HCC, a training-free structured compression method for medical image segmentation models. It represents each channel of a pretrained layer by a flattened weight vector, computes a pairwise Wasserstein distance between channels (Eq. 1), converts distances to similarities, applies agglomerative hierarchical clustering with average linkage, and replaces each cluster of channels by their arithmetic mean. The authors claim that this channel-fusion compression preserves near-baseline accuracy at an 87.5% pruning rate on BUSI, DSB, and ISIC with U-Net/U-Net++, and on FLARE21 with MedSAM2, outperforming prior pruning methods. Additional experiments report compression ratios across pruning rates and an ablation of the Wasserstein and hierarchical-clustering components.

Significance. If the central claim is correct, the paper would make a useful contribution: it suggests that a simple, training-free operation—averaging weight-space-similar channels, chosen by optimal-transport distance and hierarchical clustering—can retain near-baseline segmentation accuracy in both 2D and 3D medical imaging, with hardware-friendly structured compression. The paper evaluates on public benchmarks and includes multiple pruning rates and an ablation. However, the current evidence is not yet convincing because the main similarity measure is underspecified, the structural coupling of input/output channel compression is not formalized, and the reported accuracy improvements are single-run differences with no error bars or significance tests.

major comments (4)
  1. [Section III-A, Eq. (1)] The Wasserstein distance is never defined as a computable quantity. Eq. (1) leaves µ_p, µ_q, the metric space X, and the ground metric d unspecified; no statement tells how the vector v_p in R^{D_i} induces a probability measure, nor which OT solver is used. Since the similarity matrix M2 is the only input to hierarchical clustering, every cluster and every fused weight depends on these unspecified choices. Different natural implementations (empirical measure on the D_i entries with Euclidean ground cost, a Gaussian assumption with fixed covariance, or a Sinkhorn approximation) produce different distances and potentially different compressed models. The experiments are therefore not reproducible and the central comparison cannot be verified. Please specify the full construction and release code or pseudo-code.
  2. [Section III-A/III-C] The paper says 'the same compression method is applied to the output channel dimension as well in practice' but gives no formalization. Output-channel compression of layer i changes the input-channel dimension of layer i+1; unless the output-channel clusters of layer i are consistent with the input-channel clusters of layer i+1, the fused tensors will not be structurally aligned. The manuscript also never states whether K is chosen per layer or globally, how the 87.5% pruning rate is distributed across layers, and how both input- and output-channel compression are counted in the reported parameter/FLOPs reductions. This is load-bearing for the claim of retaining network topology for hardware acceleration.
  3. [Tables I and II] The SOTA comparison rests on single-run Dice/IoU differences of 0.1–0.5 percentage points (e.g., BUSI U-Net: 0.7841 vs 0.7781 for TPP; FLARE21 MedSAM2: 0.8471 vs 0.8346 for TPP). No standard deviations, number of runs, or significance tests are reported; with a random 80/20 split described in Sec. IV-A, these differences are within plausible seed noise. In addition, Table III states a 87.5% pruning rate but reports parameter reduction of 98.43% for MIS-HCC, which is arithmetically inconsistent and suggests the compression accounting is different from what is described. Please clarify the pruning-rate definition and add repeated-run statistics.
  4. [Section III-C] The core premise that averaging near-duplicate channels 'preserves salient features' is asserted rather than tested. The paper does not examine the distribution of pairwise Wasserstein distances before clustering, the sensitivity of the final accuracy to the chosen linkage or fusion rule, or whether clusters in early vs late layers behave differently. Given that the empirical evidence is single-run, this assumption deserves direct stress-testing (e.g., compare arithmetic averaging with channel selection and convex combinations, and report layer-wise accuracy sensitivity).
minor comments (7)
  1. [Section II-B] There is a dangling citation '[?]' in the sentence about coupling pruning with knowledge distillation.
  2. [Table I] The PFEC entry on ISIC reads '09098' instead of '0.9098'.
  3. [Table III] The method name is inconsistently written as 'TTP' in the table header and rows; the text and other tables use 'TPP'.
  4. [Table IV] The first row is labeled 'Unet-RP' but contains the U-Net baseline values (0.7806/0.7035); this appears to be a typo and should be labeled 'U-Net'.
  5. [Section IV-B-1] The sentence beginning 'For example, the relative performance changes...' is confusing and grammatically garbled; please rewrite.
  6. [Section III-B] There are formatting issues such as 'K≤N in' and 'index sets{Ic}K c=1'; these should be typeset correctly.
  7. [General] There is no code availability statement. Given the reproducibility concerns above, a release of the channel-similarity construction and clustering code would be important.

Circularity Check

0 steps flagged

No significant circularity: the derivation and evaluation are self-contained against external benchmarks; the underspecified Wasserstein distance is a reproducibility/correctness concern, not a circularity.

full rationale

Walking the claimed derivation chain: the Wasserstein similarity matrix (Sec. III-A, Eq. 1), hierarchical clustering (Sec. III-B, Eq. 3), and channel fusion (Sec. III-C, Eq. 4) all take pretrained network weights as input and produce a compressed network. None of these equations is defined in terms of the target quantity (segmentation Dice/IoU), and K is set by the desired compression ratio or dendrogram cut, not fitted to test-set accuracy. The empirical claim that MIS-HCC retains near-baseline accuracy is evaluated on external public benchmarks (BUSI, DSB, ISIC, FLARE21) against external baselines (RP, PFEC, HRank, CHEX, TPP), so the result is not forced by construction. The load-bearing premise in Sec. III-C that 'If channels within a cluster are near-duplicates, replacing them by a single prototype preserves salient features' is an unproven assumption about weight-space redundancy, but it is not circular: the paper does not define 'near-duplicates' in terms of the final segmentation metric, nor does it define its prediction in terms of that premise. The skeptical concern that Eq. (1) leaves the measures mu_p, mu_q and the ground metric d unspecified is a specification/reproducibility and correctness risk, not a circularity pattern: an underspecified distance does not make the output equal to the input by definition. There is also no load-bearing self-citation chain: references [37], [64], [65], [66] are external prior work, and the one author-overlapping citation [27] appears only in a broad list of applications and is not used to justify the compression method. The ablation study (Table IV) compares variants and does not rename a fitted quantity as a prediction. Therefore no enumerated circularity step is present; score 0.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on user-specified per-layer cluster counts, the assumption that weight-space similarity implies functional redundancy, and an under-specified procedure for simultaneous input/output channel compression. The paper introduces no new theoretical entities and no fitted constants beyond the target pruning rate.

free parameters (2)
  • K (number of clusters per layer) = set by pruning rate: 50%, 75%, 87.5%
    The user-specified pruning rate determines the per-layer cluster count. Results depend on this choice, and no automated or data-driven K selection is described.
  • Dendrogram cut threshold = not specified
    Mentioned as an alternative to choosing K, but no threshold value or selection procedure is given in the experiments.
axioms (4)
  • ad hoc to paper Channels with small Wasserstein distance in weight space are functionally redundant and can be replaced by their arithmetic mean without fine-tuning.
    Load-bearing assumption behind Eq. (4); the paper states it in Sec. III-C but provides no theoretical or empirical proof.
  • domain assumption Flattened channel vectors v_p induce measures µ_p for which the Wasserstein distance in Eq. (1) is well-defined and computable.
    Needed for Eq. (1), but the mapping v -> µ and the ground metric d are never specified.
  • ad hoc to paper Compressing input-channel and output-channel dimensions independently across all layers leaves the network structurally valid.
    Sec. III-C says the same compression is applied to output channels, but the paper does not describe how fused output groups align with the next layer's input-channel groups.
  • domain assumption Average-linkage hierarchical clustering yields stable partitions that preserve representational diversity better than other grouping schemes.
    Stated in Sec. III-B with no comparison to other linkage rules beyond the k-means ablation in Table IV.

pith-pipeline@v1.3.0-alltime-deepseek · 11415 in / 11708 out tokens · 110485 ms · 2026-08-01T18:17:07.321326+00:00 · methodology

0 comments
read the original abstract

Medical image segmentation models require both high accuracy and lightweight design to accommodate real-world medical applications. The deployment of these models on resource-limited medical platforms remains a significant challenge due to their high computational and parameter requirements. Existing pruning methods for model compression mostly overlook the intrinsic connections and similarity between the internal structures of complex deep neural networks. As a result, compressed models may not effectively retain the basic features of the pretrained network. To solve this problem, we propose a hierarchical clustering compression method for medical image segmentation models (MIS-HCC). This approach employs hierarchical clustering to partition channels and fuse their parameters efficiently. Specifically, it leverages the Wasserstein distance to represent similarity of channels within layers of pre-trained network, forming a similarity matrix that guides the clustering process. Channels within each cluster are then fused to produce a compressed network. Experimental results on three medical image datasets application demonstrate that MIS-HCC outperforms the state-of-the-art methods in both accuracy and compression efficiency, offering an effective solution for deploying medical image segmentation models on resource-limited medical platforms.

Figures

Figures reproduced from arXiv: 2607.17329 by Bo Zhao, Chang Liu, Haoran Yu, Lifei Liu, Szu-Yu Chen, Yining Liu, Zequn Xie, Zongcheng Chu.

Figure 1
Figure 1. Figure 1: The overview framework of our proposed method. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The visual comparison of our proposed MIS-HCC method with other methods on 2D and 3D segmentation, respectively. [PITH_FULL_IMAGE:figures/full_fig_p005_2.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

66 extracted references · 11 linked inside Pith

  1. [1]

    Knowledge regularized negative feature tuning of vision-language models for out- of-distribution detection,

    W. Zhu, Y . Zhang, X. Jin, W. Zeng, and L. Zhang, “Knowledge regularized negative feature tuning of vision-language models for out- of-distribution detection,” inMM, 2025

    2025

  2. [2]

    Lifelong scene graph generation,

    T. He, X. Hu, T. Wu, D. Zhang, M. Li, Y .-F. Li, and F. R. Yu, “Lifelong scene graph generation,”Pattern Recognition, 2026

    2026

  3. [3]

    Boosting learning efficiency in few-shot tasks with layer-adaptive pid control,

    P. Zhang, X. Li, L. Yu, Z. Zhang, F. Dunkin, H. Liu, and Z. Li, “Boosting learning efficiency in few-shot tasks with layer-adaptive pid control,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  4. [4]

    Adaptive event stream slicing for open-vocabulary event-based object detection via vision-language knowledge distillation,

    J. Zhang, Z. Li, J. Lin, and G. Lu, “Adaptive event stream slicing for open-vocabulary event-based object detection via vision-language knowledge distillation,”arXiv preprint arXiv:2510.00681, 2025

    arXiv 2025

  5. [5]

    Em- bodiment: Self-supervised depth estimation based on camera models,

    J. Zhang, P. K. Reddy, X.-I. Wong, Y . Aloimonos, and G. Lu, “Em- bodiment: Self-supervised depth estimation based on camera models,” inIROS, IEEE, 2024

    2024

  6. [6]

    Vision-language embodiment for monocular depth estimation,

    J. Zhang and G. Lu, “Vision-language embodiment for monocular depth estimation,” inCVPR, 2025

    2025

  7. [7]

    Depth estimation based on 3d gaussian splatting siamese defocus,

    J. Zhang, N. Xu, H. Zhang, and G. Lu, “Depth estimation based on 3d gaussian splatting siamese defocus,” inICRA, IEEE, 2025

    2025

  8. [8]

    Decoding with structured awareness: integrating directional, frequency-spatial, and structural attention for medical image segmentation,

    F. Zhang, Z. Gu, and H. Wang, “Decoding with structured awareness: integrating directional, frequency-spatial, and structural attention for medical image segmentation,” inAAAI, 2026

    2026

  9. [9]

    Computing nodes for plane data points by constructing cubic polynomial with constraints,

    H. Wang and F. Zhang, “Computing nodes for plane data points by constructing cubic polynomial with constraints,”Computer Aided Geometric Design, 2024

    2024

  10. [10]

    Reversible primitive–composition alignment for continual vision–language learn- ing,

    C. Xiao, T. Xu, S. Ma, Y . Jiang, H. Gao, and Y . Wu, “Reversible primitive–composition alignment for continual vision–language learn- ing,” inICLR, 2026

    2026

  11. [11]

    Boosting deep detector efficiency and robustness through detection discriminant reorganization and compression,

    J. Im Choi, Q. Lan, and Q. Tian, “Boosting deep detector efficiency and robustness through detection discriminant reorganization and compression,”Neural Networks, p. 109271, 2026

    2026

  12. [12]

    Visual detector compression via location-aware discriminant analysis,

    Q. Lan, J. I. Choi, and Q. Tian, “Visual detector compression via location-aware discriminant analysis,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 3546–3555, 2026

    2026

  13. [13]

    Map: Parameter-efficient tuning for referring expression comprehen- sion via multi-modal adaptive positional encoding,

    R. Yao, Y . Rong, T. Zou, B. Zhang, J. Li, S. Xiong, and S. Xiong, “Map: Parameter-efficient tuning for referring expression comprehen- sion via multi-modal adaptive positional encoding,” inProceedings of the 33rd ACM International Conference on Multimedia, pp. 2264– 2273, 2025

    2025

  14. [14]

    Frequency-aligned knowledge distillation for lightweight spatiotem- poral forecasting,

    Y . Li, C. Yang, H. Zeng, Z. Dong, Z. An, Y . Xu, Y . Tian, and H. Wu, “Frequency-aligned knowledge distillation for lightweight spatiotem- poral forecasting,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 7262–7272, 2025

    2025

  15. [15]

    A comprehensive survey of interaction techniques in 3d scene generation,

    Y . Li, S. Meng, C. Yang, W. Feng, J. Liu, Z. An, Y . Wang, and Y . Tian, “A comprehensive survey of interaction techniques in 3d scene generation,”Authorea Preprints, 2026

    2026

  16. [16]

    Sepprune: Structured pruning for efficient deep speech separation,

    Y . Li, K. Li, X. Yin, Z. Yang, Z. Dong, Z. Yao, H. Xu, Y . Tian, and Y . Lu, “Sepprune: Structured pruning for efficient deep speech separation,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, pp. 31861–31869, 2026

    2026

  17. [17]

    Detail consistent stage-wise distillation for efficient 3d mri segmentation,

    M. Fan, B. Geng, X. Xiao, T. Wang, S. Mei, P. Che, X. Jiang, and Q. Lan, “Detail consistent stage-wise distillation for efficient 3d mri segmentation,”arXiv preprint arXiv:2605.26382, 2026

    Pith/arXiv arXiv 2026

  18. [18]

    Displacement preserving relational distillation for robust medical segmentation,

    Z. Ding, X. Chu, J. I. Choi, Q. Tian, T. Shi, X. Jiang, L. Zhu, and Q. Lan, “Displacement preserving relational distillation for robust medical segmentation,”arXiv preprint arXiv:2607.04599, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    Acam-kd: adaptive and cooperative attention masking for knowledge distillation,

    Q. Lan and Q. Tian, “Acam-kd: adaptive and cooperative attention masking for knowledge distillation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 3957–3966, 2025

    2025

  20. [20]

    Clockdistill: Consistent location and con- text aware knowledge distillation for detrs,

    Q. Lan and Q. Tian, “Clockdistill: Consistent location and con- text aware knowledge distillation for detrs,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, pp. 7188–7197, 2026

    2026

  21. [21]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMICCAI, Springer, 2015

    2015

  22. [22]

    Transunet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y . Lu,et al., “Transunet: Transformers make strong encoders for medical image segmentation,”arXiv preprint arXiv:2102.04306, 2021

    Pith/arXiv arXiv 2021

  23. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, and e. a. Zhai, Xiaohua, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  24. [24]

    Fusing structure and appearance features in facial expression recognition transformer,

    S. Meng and W. Shi, “Fusing structure and appearance features in facial expression recognition transformer,” inICASSP, 2024

    2024

  25. [25]

    Fakeshield: Explainable image forgery detection and localization via multi-modal large language models,

    Z. Xu, X. Zhang, R. Li, Z. Tang, Q. Huang, and J. Zhang, “Fakeshield: Explainable image forgery detection and localization via multi-modal large language models,” inICLR, 2025

    2025

  26. [26]

    Avatarshield: Visual reinforcement learning for human-centric synthetic video de- tection,

    Z. Xu, X. Zhang, Q. Huang, X. Zhou, and J. Zhang, “Avatarshield: Visual reinforcement learning for human-centric synthetic video de- tection,”arXiv preprint arXiv:2505.15173, 2025

    arXiv 2025

  27. [27]

    Symmetry-aware causal inference for robust neural pde solvers,

    Y . Xie, Y . Xiang, H. You, N. Liu, F. Liu, B. Zhao, Z. Kang, Y . Li, and Y . Li, “Symmetry-aware causal inference for robust neural pde solvers,” inICMR, 2026

    2026

  28. [28]

    Mpq-dmv2: Flexible residual mixed precision quantization for low-bit diffusion models with temporal distillation,

    W. Feng, C. Yang,et al., “Mpq-dmv2: Flexible residual mixed precision quantization for low-bit diffusion models with temporal distillation,”IEEE TPAMI, 2026

    2026

  29. [29]

    Roboalign-r1: Distilled multimodal reward alignment for robot video world models,

    H. Wu, Y . Li, Y . Gao, F. Xu, F. Zhang, K. Wang, P. Zhao, Q. Wang, Y . Zhao, W. Wang,et al., “Roboalign-r1: Distilled multimodal reward alignment for robot video world models,”arXiv preprint arXiv:2605.03821, 2026

    Pith/arXiv arXiv 2026

  30. [30]

    Prediction of weaning from mechanical ventilation using convolutional neural networks,

    Y . Jia, C. Kaul, T. Lawton, R. Murray-Smith, and I. Habli, “Prediction of weaning from mechanical ventilation using convolutional neural networks,”Artificial intelligence in medicine, 2021

    2021

  31. [31]

    Filter pruning via ge- ometric median for deep convolutional neural networks acceleration,

    Y . He, P. Liu, Z. Wang, Z. Hu, and Y . Yang, “Filter pruning via ge- ometric median for deep convolutional neural networks acceleration,” inCVPR, 2019

    2019

  32. [32]

    Learning efficient convolutional networks through network slimming,

    Z. Liu, J. Li, Z. Shen, G. Huang, S. Yan, and C. Zhang, “Learning efficient convolutional networks through network slimming,” inCVPR, 2017

    2017

  33. [33]

    Hrank: Filter pruning using high-rank feature map,

    M. Lin, R. Ji, Y . Wang, Y . Zhang, B. Zhang, Y . Tian, and L. Shao, “Hrank: Filter pruning using high-rank feature map,” inCVPR, 2020

    2020

  34. [34]

    Chex: Channel exploration for cnn model compression,

    Z. Hou, M. Qin, F. Sun, X. Ma, K. Yuan, Y . Xu, Y .-K. Chen, R. Jin, Y . Xie, and S.-Y . Kung, “Chex: Channel exploration for cnn model compression,” inCVPR, 2022

    2022

  35. [35]

    Sauron u-net: Simple automated redundancy elimination in medical image segmentation via filter pruning,

    J. M. Valverde, A. Shatillo, and J. Tohka, “Sauron u-net: Simple automated redundancy elimination in medical image segmentation via filter pruning,”Neurocomputing, vol. 594, p. 127817, 2024

    2024

  36. [36]

    Automatic channel pruning of u-net applied to retinal blood vessels segmentation,

    X. Mei, Z. Zhu, M. Sun, and Y . Ma, “Automatic channel pruning of u-net applied to retinal blood vessels segmentation,” inCCC, IEEE, 2024

    2024

  37. [37]

    Similarity of neural network representations revisited,

    S. Kornblith, M. Norouzi, H. Lee, and G. Hinton, “Similarity of neural network representations revisited,” inICML, PMLR

  38. [38]

    Training-free cross-modal alignment via anchor profiles with statistical significance testing,

    K. Yang, J. Lu, and Y . Fu, “Training-free cross-modal alignment via anchor profiles with statistical significance testing,” inCVPR, 2026

    2026

  39. [39]

    Curvature-adaptive consistency flow matching: Autonomous trajectory optimization via reinforcement learning,

    S. Tian, G. Chen, B. Li, J. Ma, and Z. Yu, “Curvature-adaptive consistency flow matching: Autonomous trajectory optimization via reinforcement learning,” 2026

    2026

  40. [40]

    Divergence of empirical neural tangent kernel in classification problems,

    Z. Yu, S. Tian, and G. Chen, “Divergence of empirical neural tangent kernel in classification problems,” 2025

    2025

  41. [41]

    Un- supervised hyperspectral image super-resolution via self-supervised modality decoupling,

    S. Du, Y . Zou, Z. Wang, X. Li, Y . Li, C. Shang, and Q. Shen, “Un- supervised hyperspectral image super-resolution via self-supervised modality decoupling,”International Journal of Computer Vision, 2026

    2026

  42. [42]

    Pansharpening for thin-cloud contaminated remote sensing images: a unified framework and benchmark dataset,

    S. Du, Y . Zou, J. Li, M. Liu, Y . Li, C. Shang, and Q. Shen, “Pansharpening for thin-cloud contaminated remote sensing images: a unified framework and benchmark dataset,” inAAAI, 2026

    2026

  43. [43]

    Frequency-decoupled learning for joint thin-cloud removal and pansharpening,

    S. Du, Y . Bai, J. Ma, M. Liu, and Y . Li, “Frequency-decoupled learning for joint thin-cloud removal and pansharpening,” inICASSP, IEEE, 2026

    2026

  44. [44]

    Ips: In-prompt process supervision for short video content moderation,

    M. Liu, Y . Sun, R. Sun, X. Dong, X. Shen, H. Wang, H. Xiong, and Y . Song, “Ips: In-prompt process supervision for short video content moderation,” 2026

    2026

  45. [45]

    Reasoning-enhanced domain-adaptive pre- training of multimodal large language models for short video content governance,

    Z. Wang, Y . Sun, H. Wang, B. Jing, X. Shen, X. Dong, Z. Hao, H. Xiong, and Y . Song, “Reasoning-enhanced domain-adaptive pre- training of multimodal large language models for short video content governance,” inEMNLP: Industry Track, ACL, 2025

    2025

  46. [46]

    Audio-enhanced vision-language modeling with latent space broadening for high quality data expansion,

    Y . Sun, Y . Li, R. Sun, C. Liu, F. Zhou, Z. Jin, L. Wang, X. Shen, Z. Hao, and H. Xiong, “Audio-enhanced vision-language modeling with latent space broadening for high quality data expansion,” KDD ’25, Association for Computing Machinery, 2025

    2025

  47. [47]

    Render-in-the-loop: Vector graphics generation via visual self- feedback,

    G. Liang, Z. Wang, J. Hu, H. Zhou, Z. Xue, J. Zhang, D. Xu, and Q. Yu, “Render-in-the-loop: Vector graphics generation via visual self- feedback,”arXiv preprint arXiv:2604.20730, 2026

    Pith/arXiv arXiv 2026

  48. [48]

    Vanim: Rendering-aware sparse state mod- eling for structure-preserving vector animation,

    G. Liang, Z. Wang, C. Wang, J. Hu, H. Zhou, J. Liu, J. Zhang, D. Xu, and Q. Yu, “Vanim: Rendering-aware sparse state mod- eling for structure-preserving vector animation,”arXiv preprint arXiv:2605.01517, 2026

    Pith/arXiv arXiv 2026

  49. [49]

    Multi-object sketch animation with grouping and motion trajectory priors,

    G. Liang, J. Hu, X. Xing, J. Zhang, and Q. Yu, “Multi-object sketch animation with grouping and motion trajectory priors,” inMM, 2025

    2025

  50. [50]

    Spatiotemporal multi-view continual dictionary learning with graph diffusion,

    S. Wu and J. Zhang, “Spatiotemporal multi-view continual dictionary learning with graph diffusion,”Knowledge-Based Systems, 2025

    2025

  51. [51]

    Multi-resolution context augmentation and dual channel attention for 3d lane detection,

    Q. Ning, J. Zhang,et al., “Multi-resolution context augmentation and dual channel attention for 3d lane detection,”IEEE Internet of Things Journal, 2025

    2025

  52. [52]

    Multi-scale convolution and dynamic task interaction detection head for efficient lightweight plum detec- tion,

    J. Wu, J. Zhang, J. Zhu,et al., “Multi-scale convolution and dynamic task interaction detection head for efficient lightweight plum detec- tion,”Food and Bioproducts Processing, 2025

    2025

  53. [53]

    Mappo: Maximum a pos- teriori preference optimization with prior knowledge,

    G. Lan, S. Zhang, T. Wang, Y . Zhang, D. Zhang, X. Wei, X. Pan, H. Zhang, D.-J. Han, and C. G. Brinton, “Mappo: Maximum a pos- teriori preference optimization with prior knowledge,”arXiv preprint arXiv:2507.21183, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    Ants: Adaptive negative textual space shaping for ood detection via test-time mllm understanding and reasoning,

    W. Zhu, Y . Zhang, X. Jin, W. Zeng, and L. Zhang, “Ants: Adaptive negative textual space shaping for ood detection via test-time mllm understanding and reasoning,” inCVPR, 2026

    2026

  55. [55]

    Unet++: A nested u-net architecture for medical image segmentation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” in4th international workshop, DLMIA 2018, and 8th international workshop, ML-CDS 2018, held in conjunction with MICCAI, 2018

    2018

  56. [56]

    Learning how to use tools, not just when: Pattern-aware tool-integrated reasoning,

    N. Xu, Y . Jiang, S. R. Dipta, and H. Zhang, “Learning how to use tools, not just when: Pattern-aware tool-integrated reasoning,” 2026

    2026

  57. [57]

    Find your optimal teacher: Personalized data synthesis via router-guided multi-teacher distillation,

    H. Zhang, S. Yang,et al., “Find your optimal teacher: Personalized data synthesis via router-guided multi-teacher distillation,” 2026

    2026

  58. [58]

    Drp: Distilled reasoning pruning with skill-aware step decomposition for efficient large reasoning models,

    Y . Jiang, D. Li, and F. Ferraro, “Drp: Distilled reasoning pruning with skill-aware step decomposition for efficient large reasoning models,” 2026

    2026

  59. [59]

    Neural clustering based visual representation learning,

    G. Chen, X. Li, Y . Yang, and W. Wang, “Neural clustering based visual representation learning,” inCVPR, 2024

    2024

  60. [60]

    Dataset of breast ultrasound images,

    W. Al-Dhabyani, M. Gomaa, H. Khaled, and A. Fahmy, “Dataset of breast ultrasound images,”Data in brief, vol. 28, p. 104863, 2020

    2020

  61. [61]

    Nucleus segmentation across imaging experiments: the 2018 data science bowl,

    J. C. Caicedo, A. Goodman, K. W. Karhohs, B. A. Cimini, J. Acker- man, M. Haghighi, C. Heng, T. Becker, M. Doan, C. McQuin,et al., “Nucleus segmentation across imaging experiments: the 2018 data science bowl,”Nature methods, vol. 16, no. 12, pp. 1247–1253, 2019

    2018

  62. [62]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic),

    N. Codella, V . Rotemberg, P. Tschandl, M. E. Celebi, S. Dusza, D. Gutman, B. Helba, A. Kalloo, K. Liopyris, M. Marchetti,et al., “Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic),”arXiv preprint arXiv:1902.03368, 2019

    Pith/arXiv arXiv 2018

  63. [63]

    Fast and low-gpu-memory abdomen ct organ segmentation: the flare challenge,

    J. Ma, Y . Zhang, S. Gu, X. An, Z. Wang, C. Ge, C. Wang, F. Zhang, Y . Wang, Y . Xu,et al., “Fast and low-gpu-memory abdomen ct organ segmentation: the flare challenge,”Medical Image Analysis, vol. 82, p. 102616, 2022

    2022

  64. [64]

    Medsam2: Segment anything in 3d medical images and videos,

    J. Ma, Z. Yang, S. Kim, B. Chen, M. Baharoon, A. Fallahpour, R. Asakereh, H. Lyu, and B. Wang, “Medsam2: Segment anything in 3d medical images and videos,”arXiv preprint arXiv:2504.03600, 2025

    Pith/arXiv arXiv 2025

  65. [65]

    Pruning filters for efficient convnets,

    H. Li, A. Kadav, I. Durdanovic, H. Samet, and H. P. Graf, “Pruning filters for efficient convnets,”arXiv preprint arXiv:1608.08710, 2016

    Pith/arXiv arXiv 2016

  66. [66]

    Trainability preserving neural pruning,

    H. Wang and Y . Fu, “Trainability preserving neural pruning,” inICLR, 2023

    2023