Pith. sign in

REVIEW 4 major objections 6 minor 45 references

SAGE: Shape-Adapting Gated Experts for Adaptive Histopathology Image Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read SAGE turns fixed CNN-Transformer hybrids into per-input expert routers and reports top Dice scores on three histopathology benchmarks.

desk verdict A plausible dynamic-routing architecture undermined by evaluation protocol: the SOTA numbers are validation scores compared against test-set baselines. read the letter →

arxiv 2511.18493 v4 pith:LK2XUCUA submitted 2025-11-23 eess.IV cs.AIcs.CV

Gia Huy Thai , Hoang-Nguyen Vu , Anh-Minh Phan , Quang-Thinh Ly , Thi-Ngoc-Truc Nguyen , Nhat Ho This is my paper · ORCID
classification eess.IVcs.AIcs.CV
keywords histopathologysegmentationmixtureofexpertsdynamicexpertroutingCNN-TransformerhybridU-Netwhole-slideimagesglandcolorectalcancer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that the standard design of hybrid CNN-Transformer segmentation networks—a fixed, input-agnostic computation graph—is an unnecessary limitation, and that replacing it with per-input expert routing improves accuracy and domain-shift stability. It introduces SAGE, a dual-path layer that keeps the original backbone's transformation as a main path while a hierarchical gate selects among shared and fine-grained experts, with a Shape-Adapting Hub translating feature tensors between convolutional and transformer formats. Instantiated as SAGE-UNet (ConvNeXt plus ViT encoder on a U-Net decoder), the method reports Dice scores of 95.57% on EBHI, 95.16% on DigestPath, and 94.17% on GlaS Test A+B, claiming the best results among compared methods. A sympathetic reader would care because dynamic, input-dependent computation is a plausible way to handle the wide variance in cell size, shape, and tissue appearance in gigapixel histopathology.

What carries the argument

The load-bearing machinery is the SAGE block: a dual-path layer whose main path runs the pretrained backbone transformation, and whose expert path executes the top-K selected experts; hierarchical gating, which first uses a scalar gate to bias Semantic Affinity Routing logits between shared and fine-grained experts and then picks the top-K; and the Shape-Adapting Hub (SA-Hub), a pair of learnable adapters that transforms feature shapes so a CNN block can invoke a transformer expert and a transformer block can invoke a CNN expert. The dual-path fusion weight alpha is itself learned, giving the model a per-layer dial between stability and adaptation.

What would settle it

Re-run SAGE-UNet and the strongest baselines on the held-out official test sets of EBHI and DigestPath, with all methods at the same input resolution and with checkpoint selection based only on validation; if SAGE's Dice advantage over the best static hybrid falls within noise, the state-of-the-art claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that the static computation graph of hybrid CNN-Transformer networks is a bottleneck: every input is processed through the same fixed sequence of operations regardless of its complexity. SAGE replaces this with a block that routes each feature map through the top-K most semantically relevant experts from a pool of shared and fine-grained specialists, after a high-level gate biases toward either general or specialized computation, and then fuses the expert output with the unchanged main-path output via a learnable scalar. The authors argue that this dynamic restructuring—rather than stacking or concatenating CNN and Transformer modules—explains the measured gains,

Load-bearing premise

The reported state-of-the-art numbers for EBHI and DigestPath rest on the assumption that the checkpoint with the highest validation Dice, evaluated on that same validation split, approximates test-set performance; if validation performance is optimistically biased, the reported margins over prior published results may shrink.

Editorial extensions

If this is right

  • Static hybrid fusion leaves accuracy on the table: a per-input routing between local and global expert computations is the better operating point.
  • The hierarchical gating lets the model lean on shared experts for simple or domain-invariant inputs and reserve specialized experts for complex ones, concentrating compute where it is needed.
  • The SA-Hub demonstrates a general way to mix heterogeneous expert types—CNN, ViT, or others—inside one network, which the authors state is compatible with any CNN or Transformer backbone.
  • Ablation results give two design rules for MoE-style vision networks: enlarging the number of selected experts (top-K) matters more than adding shared experts, and sigmoid gating outperforms softmax gating.
  • The large margin on the domain-shifted GlaS Test B suggests dynamic routing reduces over-segmentation and boundary errors when test images differ from training.

Reading between the lines

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

  • A testable extension: run SAGE-UNet and the strongest static hybrids on the official EBHI and DigestPath test sets at a single input resolution; the reported validation-selected margins may or may not survive this check.
  • The SA-Hub's learnable shape adaptation is a generic cross-architecture adapter and could be used to couple other model families (e.g., state-space models) with CNNs or transformers without redesigning the backbone.
  • The observed layer-wise gating pattern—CNN layers favoring shared experts, deep transformer layers favoring specialized ones—suggests a design heuristic: allocate shared capacity to low-level features and specialization to high-level semantics.
  • Part of SAGE's measured gain may come from the switch to sigmoid gating rather than shape-adaptive routing; applying sigmoid gating to static hybrid baselines would help separate the two contributions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript proposes Shape-Adapting Gated Experts (SAGE), a dynamic-routing framework that augments a static CNN-Transformer U-Net with a dual-path design: a main backbone path and an expert path selected through hierarchical gating (shared vs fine-grained experts, top-K routing), with a Shape-Adapting Hub (SA-Hub) to translate features between CNN and Transformer experts. Evaluations are reported on EBHI, DigestPath, and GlaS. The paper claims state-of-the-art Dice scores of 95.57%, 95.16%, and 94.17%, with additional ablation and routing analyses in the supplement.

Significance. The idea of turning a static hybrid backbone into an input-adaptive expert system is timely and the design is plausible. The paper's strengths include a broad baseline comparison in Table 1, supplementary ablations of the number of selected experts K and shared experts S, and routing visualizations. However, the central SOTA claim is not currently supported: the headline numbers for EBHI and DigestPath are validation-set scores, the SOTA comparison table uses different input resolutions, and the abstract numbers disagree with the full text. If re-evaluated under a sound protocol, the method may still be competitive, but the evidence as presented is insufficient.

major comments (4)
  1. [§4.2 Evaluation and Model Selection; Table 2] For EBHI and DigestPath, the checkpoint with the highest validation Dice is selected and its validation performance is reported. These are validation scores, not test-set scores, and the same validation set is used for both selection and final reporting. This inflates the reported numbers and makes the comparison with published SOTA (which is normally test-set performance) invalid. The abstract's claim that SAGE achieves 95.57% and 95.16% state-of-the-art Dice is not supported without a held-out test evaluation and, ideally, multiple seeds with standard deviations.
  2. [§4.2 Comparison Protocol; Tables 1 and 2] Baseline comparisons use 512×512 inputs, while SOTA comparisons use 224×224 inputs except for DigestPath. The same SAGE-UNet + Sigmoid model appears in both tables with different results: EBHI DSC 95.01 vs 95.57, DigestPath 95.28 vs 95.16, and GlaS A+B 93.85 vs 94.17. The claimed SOTA margin therefore depends on the resolution and table. A direct comparison to literature numbers is only valid if all methods are evaluated at the same resolution and on the same split. The authors should use one consistent protocol for all comparisons.
  3. [§4.1 EBHI Dataset; Table 2] The EBHI evaluation uses only the Adenocarcinoma subset (795 images), but Table 2 gives no indication whether the listed SOTA methods were evaluated on the same subset. If the literature numbers come from the full EBHI dataset or a different class split, the 95.57% figure is not comparable. The authors should clarify this and, if necessary, re-run the SOTA baselines on the same 795-image subset.
  4. [Abstract vs. Table 2] The abstract on the paper's first page reports 95.23% EBHI, 92.78%/91.42% GlaS Test A/B, and 91.26% WSI-level DigestPath, whereas the full-text abstract and Table 2 report 95.57%, 95.16%, and 94.17% (with GlaS A/B of 94.67%/92.67%). These are materially different numbers. The authors must reconcile them and clearly state which numbers are the canonical results.
minor comments (6)
  1. [Section 2, Related Work] The sentence beginning 'The models based on the transformer structure, such as TransUNet ... can understand long State-space models...' is garbled and seems to conflate transformer and state-space model descriptions. Please rewrite for clarity.
  2. [Algorithm 1 vs. Eq. (12)–(13)] Algorithm 1, line 12, omits the output adapter S_out in the expert-path computation, which is inconsistent with Eq. (12)–(13). Please correct the algorithm or clarify the simplification.
  3. [Supplementary Section 8] The ablation text states that sigmoid surpasses softmax with 95.57% vs 95.05% on EBHI, but Table 2 reports the softmax variant at 94.77%. These numbers should be reconciled.
  4. [§4.2 SAGE-UNet Configuration] The internal architectures of the shared experts, fine-grained experts, SA-Hub adapters, and the gating networks are not specified. Add these details or release code, otherwise the method is not reproducible.
  5. [§4.3 Quantitative Results] The paper motivates SAGE by 'extra computation' and 'efficient resource allocation' but reports no FLOPs, parameter counts, or inference-time comparison. If efficiency is claimed, it should be measured.
  6. [Table 2] Each SOTA row lacks a source and a statement of the evaluation split/resolution used to obtain the published number. Add per-row citations and a protocol note.

Circularity Check

1 steps flagged · score 6.0 of 10

EBHI/DigestPath SOTA scores are validation-selected maxima by construction, making two of three headline 'predictions' circular; GlaS and internal baselines remain independent.

  1. fitted input called prediction [Sec. 4.2 'Evaluation and Model Selection'; abstract/Table 2]
    "For EBHI and DigestPath, the model checkpoint with the highest validation Dice score was selected, and its validation performance was reported."

    The abstract/Table 2 SOTA values for EBHI (95.57%) and DigestPath (95.16%) are not independent estimates: the checkpoint was selected by maximizing validation Dice, and the same validation Dice is then reported as the final score. By construction, the reported number is the maximum of the selection metric over checkpoints, so the headline 'superior segmentation' on these two benchmarks is a fit to the validation split rather than a prediction of held-out generalization. Comparing these validation-selected scores in Table 2 to published test-set SOTA numbers forces the comparison in SAGE's favor; the claimed margins are selection artifacts. This is the fitted-input-called-prediction pattern.

full rationale

The method itself is an empirical architecture and does not contain an analytic derivation that could collapse: the dual-path fusion (Eq. 5), hierarchical gating (Eq. 9), and SA-Hub (Eqs. 11-13) are all learned modules evaluated by task loss. The cited sigmoid-gating works [24,38] are co-authored by Nhat Ho, but they are used only as contextual explanation, and the paper's own Figure 10 compares sigmoid vs softmax gating, so they are not load-bearing self-citations; I do not count them as circularity. The one real circularity is evaluative: for EBHI and DigestPath the reported SOTA Dice is the validation Dice used to select the checkpoint, so these two headline numbers reduce by construction to the selection criterion. GlaS is evaluated on the official Test A/B split, so that part is independent; the internal baseline comparison in Table 1 also treats all methods under the same protocol, giving some independent evidence for the architecture. However, the abstract's claimed 'state-of-the-art' on three benchmarks depends on the EBHI/DigestPath validation-selected numbers, and the full-text abstract (95.57/95.16/94.17) disagrees with the supplied abstract (95.23/92.78/91.42/91.26), further weakening confidence. Because two of the three headline SOTA predictions reduce by construction to validation performance, partial circularity is present; the method remains externally testable.

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

The paper is empirical rather than derivational, so the ledger contains hand-selected hyperparameters and evaluation assumptions rather than mathematical axioms. The most important entries are (i) that validation Dice can stand in for test Dice, and (ii) that resolution changes between tables do not distort comparisons. No new physical entities are introduced; SA-Hub is an architectural component evidenced by the reported experiments.

free parameters (5)
  • K (number of activated experts) = 4
    Selected by ablation on EBHI (§8); Table 2 results use K=4. The reported scores depend on this hand-tuned capacity choice.
  • S (number of shared experts) = 4
    Selected jointly with K by ablation on EBHI (§8); S=4 in the reference model.
  • Loss weights lambda_ce, lambda_dice, lambda_lb = 1, 1.5, 1
    Set in §4.2 without sensitivity analysis; central Dice numbers are conditional on this choice.
  • Expert pool size = 20 (4 shared + 16 non-shared)
    Chosen in §4.2; routing diversity and capacity scale depend on it.
  • Input resolutions = 224x224 for SOTA tables, 512x512 for baseline tables, DigestPath 1536x1536 patches
    Different resolutions across comparison tables make cross-table SOTA claims non-comparable; resolution is a free experimental choice.
assumptions (5)
  • ad hoc to paper Validation-set Dice is an estimate of test-set performance.
    Used in §4.2 to report EBHI and DigestPath results from the checkpoint with highest validation Dice without a held-out test set. Statistically unsupported and load-bearing for the SOTA claim.
  • ad hoc to paper Comparison to 'state-of-the-art' methods using 224x224 inputs versus 512x512 inputs does not distort relative ranking.
    §4.2 switches resolution between baseline and SOTA tables; this assumption is not defended.
  • domain assumption Sigmoid gating is more sample-efficient than softmax in this MoE setting [24,38].
    Used in §8 to explain why sigmoid is preferred and to interpret ablation; it is imported from prior work, including co-authored papers.
  • domain assumption Auxiliary load-balancing loss prevents router collapse and preserves generalization.
    Assumed throughout §3.1 and Algorithm 1; no independent validation that this holds at every depth.
  • domain assumption Pretrained ConvNeXt/ViT weights transfer into the dynamic routing scheme.
    The method reuses ImageNet-pretrained backbones (§4.2); successful transfer is assumed to be maintained under dual-path routing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAGE: Shape-Adapting Gated Experts for Adaptive Histopathology Image Segmentation." pith.science (2026). https://pith.science/paper/LK2XUCUA

@misc{pith2026251118493,
  author       = {Pith},
  title        = {Pith review of: SAGE: Shape-Adapting Gated Experts for Adaptive Histopathology Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LK2XUCUA}},
  note         = {Machine review of arXiv:2511.18493}
}
read the original abstract

The significant variability in cell size and shape continues to pose a major obstacle in computer-assisted cancer detection on gigapixel Whole Slide Images (WSIs), due to cellular heterogeneity. Current CNN-Transformer hybrids use static computation graphs with fixed routing. This leads to extra computation and makes it harder to adapt to changes in input. We propose Shape-Adapting Gated Experts (SAGE), an input-adaptive framework that enables dynamic expert routing in heterogeneous visual networks. SAGE reconfigures static backbones into dynamically routed expert architectures via a dual-path design with hierarchical gating and a Shape-Adapting Hub (SA-Hub) that harmonizes feature representations across convolutional and transformer modules. Embodied as SAGE with ConvNeXt and Vision Transformer UNet (SAGE-ConvNeXt+ViT-UNet), our model achieves a Dice score of 95.23% on EBHI, DSC scores of 92.78% and 91.42% on GlaS Test A and Test B, respectively, and 91.26% DSC at the WSI level on DigestPath, while exhibiting robust generalization under distribution shifts by adaptively balancing local refinement and global context. SAGE establishes a scalable foundation for dynamic expert routing in visual networks, thereby facilitating flexible visual reasoning. Project page: https://oxyzgiahuy.github.io/sage/

Figures

Figures reproduced from arXiv: 2511.18493 by the authors.

Figure 1
Figure 1. Explainability visualization of dynamic expert routing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The proposed architecture integrates a ConvNeXt [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The router dynamically divides computing between [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on GlaS test samples. Each column shows (a) the input image with ground-truth annotation, (b) TransUNet, (c) EViT-UNet, and (d) our proposed SAGE-UNet. The top row presents a typical gland structure (GlaS Test A), while the bottom row depicts a c…
Figure 6
Figure 6. Figure 6: Top-K activation map: This binary heatmap shows rout￾ing choice information for K = 4 and over the layers of the model (rows correspond to CNN layers 1 − 4 and Transformer layers 1 − 16) and the 20 experts (columns correspond to E0 through E19). The activation of an ex…
Figure 7
Figure 7. Figure 7: Overview of Semantic Affinity Routing (SAR) Mechanism in SAGE: The query vectors of CNN and Transformer calculate [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The structure of the Shape-Adapting Hub (SA-Hub) over activated experts: The first step of the shape-adaptation process in the [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Shared Gating Scalar Analysis over Training: This figure illustrates the evolution of shared gating scalar [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Ablation of gating strategies and expert capacity. We compare sigmoid and softmax gating while varying Top-K (K) and the number of shared experts (S). Sigmoid gating consis￾tently yields superior performance across all metrics (Accuracy, IoU, DSC). A positive performa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 5 linked inside Pith

  1. [1]

    Understanding of a convolutional neural network

    Saad Albawi, Tareq Abed Mohammed, and Saad Al-Zawi. Understanding of a convolutional neural network. In2017 International Conference on Engineering and Technology (ICET), pages 1–6, 2017. 2

    2017

  2. [2]

    Swin-unet: Unet-like pure transformer for medical image segmentation

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. InEuropean conference on computer vision, pages 205–218. Springer, 2022. 2, 7

    2022

  3. [3]

    Bingzhi Chen, Yishu Liu, Zheng Zhang, Guangming Lu, and Adams Wai Kin Kong. Transattunet: Multi-level attention- guided u-net with transformer for medical image segmenta- tion.IEEE Transactions on Emerging Topics in Computa- tional Intelligence, 8(1):55–68, 2023. 8

    2023

  4. [4]

    Transunet: Transformers make strong encoders for medi- cal image segmentation.arXiv preprint arXiv:2102.04306,

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medi- cal image segmentation.arXiv preprint arXiv:2102.04306,

    arXiv

  5. [5]

    Transunet: Rethinking the u-net architec- ture design for medical image segmentation through the lens of transformers.Medical Image Analysis, 97:103280, 2024

    Jieneng Chen, Jieru Mei, Xianhang Li, Yongyi Lu, Qihang Yu, Qingyue Wei, Xiangde Luo, Yutong Xie, Ehsan Adeli, Yan Wang, et al. Transunet: Rethinking the u-net architec- ture design for medical image segmentation through the lens of transformers.Medical Image Analysis, 97:103280, 2024. 2, 6

    2024

  6. [6]

    Adamv-moe: Adaptive multi-task vision mixture-of- experts

    Tianlong Chen, Xuxi Chen, Xianzhi Du, Abdullah Rashwan, Fan Yang, Huizhong Chen, Zhangyang Wang, and Yeqing Li. Adamv-moe: Adaptive multi-task vision mixture-of- experts. In2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 17300–17311, 2023. 2

    2023

  7. [7]

    Metaxas, Hongsheng Li, Chaofu Wang, and Shaoting Zhang

    Qian Da, Xiaodi Huang, Zhongyu Li, Yanfei Zuo, Chen- bin Zhang, Jingxin Liu, Wen Chen, Jiahui Li, Dou Xu, Zhiqiang Hu, Hongmei Yi, Yan Guo, Zhe Wang, Ling Chen, Li Zhang, Xianying He, Xiaofan Zhang, Ke Mei, Chuang Zhu, Weizeng Lu, Linlin Shen, Jun Shi, Jun Li, Sreehari S, Ganapathy Krishnamurthi, Jiangcheng Yang, Tiancheng Lin, Qingyu Song, Xuechen Liu, S...

    2022

  8. [8]

    Damai Dai, Chengqi Deng, Chenggang Zhao, R. X. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y . Wu, Zhenda Xie, Y . K. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. Deepseek- moe: Towards ultimate expert specialization in mixture-of- experts language models, 2024. 2

    2024

Show all 45 references
  1. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  2. [10]

    Convunext: An efficient convolution neural network for medical im- age segmentation.Knowledge-based systems, 253:109512,

    Zhimeng Han, Muwei Jian, and Gai-Ge Wang. Convunext: An efficient convolution neural network for medical im- age segmentation.Knowledge-based systems, 253:109512,

  3. [11]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 7

    2015

  4. [12]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 2

    2016

  5. [13]

    Universal language model fine-tuning for text classification, 2018

    Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification, 2018. 6

    2018

  6. [14]

    Ebhi: A new enteroscope biopsy histopathological h&e image dataset for image classification evaluation.Physica Medica, 107:102534, 2023

    Weiming Hu, Chen Li, Md Mamunur Rahaman, Haoyuan Chen, Wanli Liu, Yudong Yao, Hongzan Sun, Marcin Grze- gorzek, and Xiaoyan Li. Ebhi: A new enteroscope biopsy histopathological h&e image dataset for image classification evaluation.Physica Medica, 107:102534, 2023. 6

    2023

  7. [15]

    Transformers in vision: A survey.ACM computing surveys (CSUR), 54(10s):1–41, 2022

    Salman Khan, Muzammal Naseer, Munawar Hayat, Syed Waqas Zamir, Fahad Shahbaz Khan, and Mubarak Shah. Transformers in vision: A survey.ACM computing surveys (CSUR), 54(10s):1–41, 2022. 2

    2022

  8. [16]

    Evit-unet: U-net like efficient vision trans- former for medical image segmentation on mobile and edge devices

    Xin Li, Wenhui Zhu, Xuanzhao Dong, Oana M Dumitrascu, and Yalin Wang. Evit-unet: U-net like efficient vision trans- former for medical image segmentation on mobile and edge devices. In2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2025. 8

    2025

  9. [17]

    Ds-transunet: Dual swin transformer u-net for medical image segmentation.IEEE Transactions on Instrumentation and Measurement, 71:1– 15, 2022

    Ailiang Lin, Bingzhi Chen, Jiayu Xu, Zheng Zhang, Guang- ming Lu, and David Zhang. Ds-transunet: Dual swin transformer u-net for medical image segmentation.IEEE Transactions on Instrumentation and Measurement, 71:1– 15, 2022. 8

    2022

  10. [18]

    Swin-umamba: Mamba-based unet with imagenet-based pretraining

    Jiarun Liu, Hao Yang, Hong-Yu Zhou, Yan Xi, Lequan Yu, Cheng Li, Yong Liang, Guangming Shi, Yizhou Yu, Shaot- ing Zhang, et al. Swin-umamba: Mamba-based unet with imagenet-based pretraining. InInternational conference on medical image computing and computer-assisted interven- ...

    2024

  11. [19]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feicht- enhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11966–11976, 2022. 2, 4, 6, 7

    2022

  12. [20]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 6

    2019

  13. [21]

    Segment anything in medical images.Nature Communications, 15(1), 2024

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature Communications, 15(1), 2024. 2

    2024

  14. [22]

    U-mamba: Enhancing long-range dependency for biomedical image segmentation

    Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 2, 7

    2024 arXiv

  15. [23]

    Choos- ing smartly: Adaptive multimodal fusion for object detection in changing environments

    Oier Mees, Andreas Eitel, and Wolfram Burgard. Choos- ing smartly: Adaptive multimodal fusion for object detection in changing environments. In2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 151–156, 2016. 2

    2016

  16. [24]

    Sigmoid gating is more sample efficient than softmax gating in mix- ture of experts, 2024

    Huy Nguyen, Nhat Ho, and Alessandro Rinaldo. Sigmoid gating is more sample efficient than softmax gating in mix- ture of experts, 2024. 2, 3 9

    2024

  17. [25]

    C2gmatch: Leveraging dual-view cross-guidance and co-guidance framework for semi-supervised cell segmentation

    Thi-Ngoc-Truc Nguyen, Xuan-Hong Ong, Hoang-Thien- Nguyen, Van-Hung Bui, Hoang-Nguyen Vu, Thanh Minh Nguyen, Quan Nguyen, and Huu-Hung Nguyen. C2gmatch: Leveraging dual-view cross-guidance and co-guidance framework for semi-supervised cell segmentation. In2025 IEEE 6th Internat...

    2025

  18. [26]

    Atten- tion u-net: Learning where to look for the pancreas.arXiv preprint arXiv:1804.03999, 2018

    Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Atten- tion u-net: Learning where to look for the pancreas.arXiv preprint arXiv:1804.03999, 2018. 8

    2018 arXiv

  19. [27]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Inter- vention – MICCAI 2015, pages 234–241. Springer, Cham,

    2015

  20. [28]

    U-net: Convolutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation,

  21. [29]

    Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer, 2017. 2, 3

    2017

  22. [30]

    Korsuk Sirinukunwattana, Josien P. W. Pluim, Hao Chen, Xi- aojuan Qi, Pheng-Ann Heng, Yun Bo Guo, Li Yang Wang, Bogdan J. Matuszewski, Elia Bruni, Urko Sanchez, Anton B¨ohm, Olaf Ronneberger, Bassem Ben Cheikh, Daniel Raco- ceanu, Philipp Kainz, Michael Pfeiffer, Martin Urschl...

  23. [31]

    Medical image anal- ysis using improved sam-med2d: segmentation and classifi- cation perspectives.BMC Medical Imaging, 24:241, 2024

    Jiahao Sun, Kai Chen, Zhiwei He, et al. Medical image anal- ysis using improved sam-med2d: segmentation and classifi- cation perspectives.BMC Medical Imaging, 24:241, 2024. 2

    2024

  24. [32]

    Mingxing Tan and Quoc V . Le. Efficientnet: Rethinking model scaling for convolutional neural networks.ArXiv, abs/1905.11946, 2019. 2

    1905 arXiv

  25. [33]

    Mingxing Tan and Quoc V . Le. Efficientnet: Rethinking model scaling for convolutional neural networks, 2020. 7

    2020

  26. [34]

    Rachel S. Y . Teo and Tan M. Nguyen. Molex: Mixture of layer experts for finetuning with sparse upcycling, 2025. 3

    2025

  27. [35]

    Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer

    Haonan Wang, Peng Cao, Jiaqi Wang, and Osmar R Zaiane. Uctransnet: rethinking the skip connections in u-net from a channel-wise perspective with transformer. InProceed- ings of the AAAI conference on artificial intelligence, pages 2441–2449, 2022. 8

    2022

  28. [36]

    Moe-nuseg: Enhancing nuclei segmen- tation in histology images with a two-stage mixture of ex- perts network.Alexandria Engineering Journal, 110:557– 566, 2025

    Xuening Wu, Yiqing Shen, Qing Zhao, Yanlan Kang, and Wenqiang Zhang. Moe-nuseg: Enhancing nuclei segmen- tation in histology images with a two-stage mixture of ex- perts network.Alexandria Engineering Journal, 110:557– 566, 2025. 2, 7

    2025

  29. [37]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021. 2, 7

    2021

  30. [38]

    Sigmoid self-attention is better than softmax self-attention: A mixture-of-experts perspective

    Fanqi Yan, Huy Nguyen, Pedram Akbarian, Nhat Ho, and Alessandro Rinaldo. Sigmoid self-attention is better than softmax self-attention: A mixture-of-experts perspective. arXiv preprintarXiv: 2502.00281, 2025. 2, 3

    2025 arXiv

  31. [39]

    Multi-task dense prediction via mixture of low-rank experts, 2024

    Yuqi Yang, Peng-Tao Jiang, Qibin Hou, Hao Zhang, Jinwei Chen, and Bo Li. Multi-task dense prediction via mixture of low-rank experts, 2024. 2

    2024

  32. [40]

    Unet++: A nested u-net ar- chitecture for medical image segmentation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net ar- chitecture for medical image segmentation. InDeep Learn- ing in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, pages 3–11, Cham, 2018. Spri...

    2018

  33. [41]

    Multi-level colonoscopy malignant tissue detection with adversarial cac-unet.Neurocomputing, 438:165–183, 2021

    Chuang Zhu, Ke Mei, Ting Peng, Yihao Luo, Jun Liu, Ying Wang, and Mulan Jin. Multi-level colonoscopy malignant tissue detection with adversarial cac-unet.Neurocomputing, 438:165–183, 2021. 8

    2021

  34. [42]

    Selfreg- unet: Self-regularized unet for medical image segmenta- tion

    Wenhui Zhu, Xiwen Chen, Peijie Qiu, Mohammad Farazi, Aristeidis Sotiras, Abolfazl Razi, and Yalin Wang. Selfreg- unet: Self-regularized unet for medical image segmenta- tion. InInternational Conference on Medical Image Com- puting and Computer-Assisted Intervention, pages 601–...

    2024

  35. [43]

    Semantic Affinity Routing (SAR) Analysis A quantitative analysis of the Mixture-of-Experts (MoE) model’s expert utilization metrics across a heterogeneous instance space confirms an equitable distribution of rout- ing decisions. The supporting evidence, derived from the activa...

  36. [44]

    Recall the hierarchical rout- ing mechanism from the main paper

    Shared Gating Scalar Analysis To gain mechanistic insight into the behavior of the hierar- chical routing mechanism, we visualize the evolution of the shared gating scalarg s on the GlaS dataset during the sec- ond training stage (Figure 9). Recall the hierarchical rout- ing m...

  37. [45]

    Gating Mechanism.Our initial evaluation focused on comparing the performance of sigmoid versus softmax gat- ing functions for the expert routing layer

    Ablation Studies We investigate the impact of gating strategies and expert capacity on the EBHI dataset, with quantitative results il- lustrated in Figure 10. Gating Mechanism.Our initial evaluation focused on comparing the performance of sigmoid versus softmax gat- ing functi...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.