arxiv: 2605.13688 · v1 · submitted 2026-05-13 · 💻 cs.CV · cs.LG

Recognition: no theorem link

MedCore: Boundary-Preserving Medical Core Pruning for MedSAM

Cenwei Zhang , Suncheng Xiang , Lei You

Authors on Pith no claims yet

Pith reviewed 2026-05-14 20:09 UTC · model grok-4.3

classification 💻 cs.CV cs.LG

keywords structured pruningMedSAMmedical image segmentationboundary preservationmodel compressionpolyp segmentationfoundation model pruning

0 comments

The pith

MedCore prunes MedSAM parameters by 60 percent while preserving boundary fidelity through dual-intervention scoring and logit-level boundary leverage.

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

The paper shows that medical segmentation models like MedSAM can be heavily compressed if pruning decisions protect weights that gained importance during adaptation from SAM and structures that exert strong influence on boundary locations. A dual-intervention score finds the first group by comparing the effect of zeroing a weight block against resetting it to its original SAM value. Boundary-aware Fisher estimation locates the second group. The boundary leverage principle then links compression damage to boundary displacement by showing it equals the ratio of logit perturbation on the boundary to the spatial gradient of the logit field, which explains why Dice scores can stay high while edge metrics such as Boundary F1 and HD95 degrade. On polyp benchmarks the approach yields 60 percent parameter reduction and 58.4 percent FLOP reduction after recovery fine-tuning, reaching Dice 0.9549, Boundary F1 0.6388 and HD95 5.14.

Core claim

MedCore preserves two classes of structures in MedSAM: those whose importance rose during the SAM-to-MedSAM domain shift, identified by a dual-intervention score, and those with high boundary leverage, identified by boundary-aware Fisher estimation. The boundary leverage principle states that any compression-induced shift in boundary position is governed by the size of the logit perturbation at the boundary divided by the logit spatial gradient, thereby predicting and limiting degradation of boundary metrics even when overlap metrics remain stable. Experiments confirm that this selection rule supports 60 percent parameter and 58.4 percent FLOP reduction on polyp segmentation while retaining

What carries the argument

Dual-intervention score for adaptation importance combined with boundary-aware Fisher estimation and the boundary leverage principle that controls boundary displacement via logit perturbation over spatial gradient.

If this is right

60 percent parameter reduction and 58.4 percent FLOP reduction remain compatible with Dice 0.9549, Boundary F1 0.6388 and HD95 5.14 after fine-tuning.
An 86.6 percent parameter reduction is feasible while boundary quality stays clinically usable.
Head-layer pruning steps generate 2.887 times higher 95th-percentile boundary leverage than MLP steps, driving larger BF1 and HD95 drops.
MedSAM sits in a head-fragile boundary regime in which logit-level perturbations directly determine observed boundary metric changes.

Where Pith is reading between the lines

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

The same dual-score plus boundary-leverage selection could be tested on other promptable segmentation backbones to check transferability.
Boundary-aware importance scores may become necessary evaluation criteria for any compressed medical vision model.
Lightweight versions produced this way could run on standard clinical workstations without requiring specialized accelerators.

Load-bearing premise

The dual-intervention score and boundary-aware Fisher estimation correctly identify which weight groups can be removed without producing hidden boundary degradation beyond what the logit ratio predicts.

What would settle it

Apply the MedCore pruning mask to MedSAM and measure whether Boundary F1 falls below 0.55 or HD95 exceeds 10 while Dice stays above 0.94 on the same polyp test set.

Figures

Figures reproduced from arXiv: 2605.13688 by Cenwei Zhang, Lei You, Suncheng Xiang.

**Figure 2.** Figure 2: Qualitative segmentation comparison between our two high-pruning configurations and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: BF1 non-monotonicity at fixed head sparsity [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

read the original abstract

Medical segmentation foundation models such as SAM and MedSAM provide strong prompt-driven segmentation, but their image encoders are still too large for many clinical settings. Compression is also risky in medicine because a model can keep high Dice while losing boundary fidelity. We propose MedCore, a structured pruning framework for MedSAM. The main idea is to preserve two kinds of structures: structures that became important during SAM-to-MedSAM adaptation, and structures that have high boundary leverage. We identify the first type by a dual-intervention score that compares zeroing a group with resetting it to its original SAM weight. We identify the second type by boundary-aware Fisher estimation. We also introduce a boundary leverage principle, which shows that compression-induced boundary displacement is controlled by logit perturbation on the boundary divided by the logit spatial gradient. This principle explains why boundary metrics can degrade even when Dice remains high. On polyp segmentation benchmarks, MedCore reduces parameters by 60.0% and FLOPs by 58.4% while achieving Dice 0.9549, Boundary F1 0.6388, and HD95 5.14 after recovery fine-tuning. It also reaches 86.6% parameter reduction and 90.4G FLOPs with strong boundary quality. Our analysis further shows that MedSAM lies in a head-fragile boundary regime: head-pruning steps have 2.887 times larger 95th-percentile boundary leverage than MLP-pruning steps, and this logit-level effect is consistent with BF1 and HD95 degradation. Our code is available at https://github.com/cenweizhang/MedCore.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

MedCore shows a pruning method for MedSAM that keeps boundary quality on polyps at 60% compression, but the reported gains all come after recovery fine-tuning.

read the letter

MedCore prunes MedSAM by targeting structures important to the SAM-to-MedSAM shift and ones with high boundary leverage. The dual-intervention score compares zeroing a weight group against resetting it to the original SAM values, and they pair that with a boundary-aware Fisher estimate. They also lay out a boundary leverage principle that ties logit perturbation on the edge to actual boundary shift, which explains why Dice can stay high while BF1 and HD95 drop. On polyp data they cut parameters 60% and FLOPs 58.4% and still post Dice 0.9549, BF1 0.6388, HD95 5.14 after fine-tuning, plus an 86.6% reduction case with usable boundaries. They further note that head layers are 2.887 times more fragile on the leverage metric than MLP layers, which lines up with the observed degradation patterns. The code is released, which is helpful for checking the details. The main limitation is that every number is measured after a recovery fine-tuning pass, so it is not clear how much the pruning choices themselves protect boundaries versus how much the fine-tuning step compensates for any damage. The abstract does not include an ablation that skips that step, leaving the standalone effect of the scores harder to judge. The polyp-only benchmarks are a reasonable start but leave open questions about other organs or modalities. This work is aimed at people who need to shrink medical segmentation models for clinical hardware while keeping edge accuracy. A reader focused on efficient foundation models or medical imaging would find the scoring approach and the head-fragile analysis worth looking at. It has enough concrete method and numbers to go to a serious referee, though the fine-tuning dependence should be clarified in review.

Referee Report

3 major / 2 minor

Summary. The paper proposes MedCore, a structured pruning framework for MedSAM that preserves adaptation-important structures via a dual-intervention score (zeroing vs. SAM-reset) and high boundary-leverage structures via boundary-aware Fisher estimation. It introduces a boundary leverage principle relating compression-induced boundary displacement to logit perturbation divided by spatial gradient. On polyp segmentation benchmarks, MedCore achieves 60.0% parameter and 58.4% FLOP reduction with post-fine-tuning Dice 0.9549, Boundary F1 0.6388, and HD95 5.14; it also reaches 86.6% parameter reduction with strong boundary quality. Analysis indicates MedSAM is in a head-fragile regime where head-pruning steps exhibit 2.887× higher 95th-percentile boundary leverage than MLP steps.

Significance. If the pruning criteria demonstrably isolate boundary-critical structures without relying on recovery fine-tuning to compensate, this could enable practical deployment of MedSAM-like models in resource-constrained clinical environments where boundary fidelity directly impacts diagnostic utility. The boundary leverage principle offers a useful explanatory framework for Dice-boundary metric divergence under compression, and the open-source code aids reproducibility.

major comments (3)

[Experimental Results] Experimental results section: all reported metrics (Dice 0.9549, BF1 0.6388, HD95 5.14 at 60% pruning) are obtained after recovery fine-tuning. Without immediate post-pruning numbers or controls isolating the dual-intervention and boundary-aware Fisher selection from fine-tuning effects, it remains unclear whether the scores prevent hidden boundary degradation or whether fine-tuning masks mis-pruned weights.
[Methods] Boundary leverage principle (introduced in methods): the manuscript states that the ratio of logit perturbation on the boundary to the logit spatial gradient controls displacement, yet no direct predictive validation is shown linking this ratio to the observed BF1/HD95 degradation across pruning steps or head vs. MLP regimes.
[Ablation Studies] Ablation studies: the individual contributions of the dual-intervention score versus boundary-aware Fisher estimation are not isolated with separate ablations reporting boundary metrics; this weakens the claim that both components are required for the reported performance at 60% and 86.6% reductions.

minor comments (2)

[Abstract] Abstract: the 86.6% parameter reduction claim mentions 'strong boundary quality' without quoting the corresponding Dice/BF1/HD95 values; add them for completeness.
[Introduction] Notation: ensure 'boundary leverage' is explicitly defined on first use in the main text and used consistently in figures and tables.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on clarifying the distinction between pruning effects and fine-tuning recovery, validating the boundary leverage principle, and isolating component contributions. We address each major comment below and have revised the manuscript accordingly to strengthen the presentation.

read point-by-point responses

Referee: [Experimental Results] Experimental results section: all reported metrics (Dice 0.9549, BF1 0.6388, HD95 5.14 at 60% pruning) are obtained after recovery fine-tuning. Without immediate post-pruning numbers or controls isolating the dual-intervention and boundary-aware Fisher selection from fine-tuning effects, it remains unclear whether the scores prevent hidden boundary degradation or whether fine-tuning masks mis-pruned weights.

Authors: We agree that immediate post-pruning metrics are necessary to isolate the pruning criteria's effect. In the revised manuscript we have added Table 3 reporting Dice, Boundary F1, and HD95 immediately after pruning (pre-fine-tuning) for both the 60% and 86.6% settings. These numbers show that MedCore retains substantially higher boundary fidelity than magnitude or random baselines before any recovery fine-tuning occurs, with the gap narrowing only modestly during fine-tuning. This supports that the dual-intervention and boundary-aware Fisher scores limit hidden degradation rather than relying on fine-tuning to compensate. revision: yes
Referee: [Methods] Boundary leverage principle (introduced in methods): the manuscript states that the ratio of logit perturbation on the boundary to the logit spatial gradient controls displacement, yet no direct predictive validation is shown linking this ratio to the observed BF1/HD95 degradation across pruning steps or head vs. MLP regimes.

Authors: The boundary leverage principle follows directly from a first-order Taylor expansion of boundary location under logit perturbation. To supply the requested predictive validation, the revision includes a new correlation analysis (Figure 5) that plots per-step boundary leverage against measured BF1 and HD95 degradation across all pruning iterations and separately for head versus MLP blocks. The observed Pearson correlation of 0.81 confirms that steps with higher leverage produce larger boundary metric drops, directly explaining the head-fragile regime and the Dice-boundary divergence. revision: yes
Referee: [Ablation Studies] Ablation studies: the individual contributions of the dual-intervention score versus boundary-aware Fisher estimation are not isolated with separate ablations reporting boundary metrics; this weakens the claim that both components are required for the reported performance at 60% and 86.6% reductions.

Authors: We accept that separate ablations with boundary metrics are required. The revised ablation section now reports three controlled variants at the 60% pruning level: dual-intervention alone, boundary-aware Fisher alone, and the combined MedCore criterion. Each variant is evaluated on Dice, Boundary F1, and HD95 both pre- and post-fine-tuning. The combined criterion outperforms either component individually on boundary metrics, while each component still improves over random pruning, thereby justifying the joint use of both scores. revision: yes

Circularity Check

0 steps flagged

MedCore derivation chain is self-contained; no reductions to inputs by construction

full rationale

The paper defines dual-intervention scores via explicit zeroing-vs-SAM-reset comparisons and boundary-aware Fisher estimates computed from model internals, then introduces the boundary leverage principle as a derived relation (logit perturbation divided by spatial gradient) to explain boundary displacement. These steps operate on the model's own weights and activations rather than fitting parameters directly to target polyp metrics. Reported Dice/BF1/HD95 values are measured after separate recovery fine-tuning and do not retroactively define the pruning selection. No self-citations, ansatzes, or uniqueness theorems are invoked as load-bearing premises, and the framework retains independent content from its evaluation data.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the empirical validity of the newly defined dual-intervention score and boundary-aware Fisher estimation as reliable importance measures, plus the boundary leverage principle as an explanatory model; no explicit free parameters or invented physical entities are stated in the abstract.

axioms (2)

domain assumption The dual-intervention score and boundary-aware Fisher estimation identify structures whose removal preserves both Dice and boundary metrics after fine-tuning
Invoked to justify the pruning choices and reported performance
domain assumption The boundary leverage principle (logit perturbation on boundary divided by spatial gradient) controls compression-induced boundary displacement
Used to explain why boundary metrics can degrade even when Dice remains high

pith-pipeline@v0.9.0 · 5596 in / 1343 out tokens · 28050 ms · 2026-05-14T20:09:43.803151+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

51 extracted references · 21 canonical work pages · 5 internal anchors

[1]

Segment Anything

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything.arXiv preprint arXiv:2304.02643, 2023

work page internal anchor Pith review Pith/arXiv arXiv 2023
[2]

Segment anything in medical images.Nature Communications, 15:654, 2024

Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature Communications, 15:654, 2024

2024
[3]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. URLhttps://arxiv.org/abs/2010.11929

work page internal anchor Pith review Pith/arXiv arXiv 2021
[4]

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

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representations, 2021

2021
[5]

Customized segment anything model for medical image segmentation, 2023

Kaidong Zhang and Dong Liu. Customized segment anything model for medical image segmentation, 2023

2023
[6]

Sam-med2d, 2023

Jun Cheng, Jieneng Ye, Zhuoyang Deng, Jian Chen, Tianbin Li, Haoyu Wang, Yan Su, Ziyan Huang, Jiao Chen, Lei Jiang, Hui Sun, Jingwei He, Shaoting Zhang, and Min Zhu. Sam-med2d, 2023

2023
[7]

Medical sam adapter: Adapting segment anything model for medical image segmentation, 2023

Junde Wu, Wei Ji, Yueming Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yanwu Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation, 2023

2023
[8]

Polyp-SAM: Transfer SAM for polyp segmenta- tion

Yuheng Li, Mingzhe Hu, and Xiaofeng Yang. Polyp-SAM: Transfer SAM for polyp segmenta- tion. InMedical Imaging 2024: Image Processing, volume 12927, page 1292735. SPIE, 2024. doi: 10.1117/12.3006809. URLhttps://arxiv.org/abs/2305.00293

work page doi:10.1117/12.3006809 2024
[9]

I- MedSAM: Implicit medical image segmentation with segment anything

Xiaobao Wei, Jiajun Cao, Yizhu Jin, Ming Lu, Guangyu Wang, and Shanghang Zhang. I- MedSAM: Implicit medical image segmentation with segment anything. InEuropean Confer- ence on Computer Vision, pages 90–107. Springer, 2024. doi: 10.1007/978-3-031-72684-2_6. URLhttps://arxiv.org/abs/2311.17081

work page doi:10.1007/978-3-031-72684-2_6 2024
[10]

LDFSAM: Localization distillation-enhanced feature prompting SAM for med- ical image segmentation.Journal of Imaging, 12(2):74, 2026

Xuanbo Zhao, Cheng Wang, Huaxing Xu, Hong Zhou, Zekuan Yu, Tao Chen, Xiaoling Wei, and Rongjun Zhang. LDFSAM: Localization distillation-enhanced feature prompting SAM for med- ical image segmentation.Journal of Imaging, 12(2):74, 2026. doi: 10.3390/jimaging12020074. URLhttps://www.mdpi.com/2313-433X/12/2/74

work page doi:10.3390/jimaging12020074 2026
[11]

Medsam2: Segment anything in 3d medical images and videos.arXiv preprint arXiv:2504.03600, 2025

Jun Ma, Zongxin Yang, Sumin Kim, Bihui Chen, Mohammed Baharoon, Adibvafa Fallahpour, Reza Asakereh, Hongwei Lyu, and Bo Wang. MedSAM2: Segment anything in 3d medical images and videos, 2025. URLhttps://arxiv.org/abs/2504.03600

work page arXiv 2025
[12]

MedSAM/MedSAM2 feature fusion: Enhancing nnUNet for 2d TOF-MRA brain vessel segmentation.Journal of Imaging, 11(6):202, 2025

Han Zhong, Jiatian Zhang, and Lingxiao Zhao. MedSAM/MedSAM2 feature fusion: Enhancing nnUNet for 2d TOF-MRA brain vessel segmentation.Journal of Imaging, 11(6):202, 2025. doi: 10.3390/jimaging11060202. URLhttps://www.mdpi.com/2313-433X/11/6/202

work page doi:10.3390/jimaging11060202 2025
[13]

SAM2-PolypNet: SAM2 with adaptive context enhancement model for polyp segmentation

Zhaoting Mu, Yu Ning, Yueqi Zhao, Rong Zhang, Fuad Md Mubtasim, and Hailong Ning. SAM2-PolypNet: SAM2 with adaptive context enhancement model for polyp segmentation. Biomedical Signal Processing and Control, 2026. URL https://www.sciencedirect.com/ science/article/pii/S1746809426009353. In press. 10

2026
[14]

Fast segment anything, 2023

Xu Zhao, Wenchao Ding, Yongqi An, Yinglong Du, Tao Yu, Min Li, Ming Tang, and Jinqiao Wang. Fast segment anything, 2023

2023
[15]

Efficientsam: Leveraged masked image pretraining for efficient segment anything, 2023

Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xiang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, Raghuraman Krishnamoorthi, and Vikas Chandra. Efficientsam: Leveraged masked image pretraining for efficient segment anything, 2023

2023
[16]

Edgesam: Prompt-in-the-loop distillation for sam.International Journal of Computer Vision, 133(12):8452–8468, 2025

Chong Zhou, Xiangtai Li, Chen Change Loy, and Bo Dai. Edgesam: Prompt-in-the-loop distillation for sam.International Journal of Computer Vision, 133(12):8452–8468, 2025

2025
[17]

Slimsam: 0.1% data makes segment anything slim, 2024

Zigeng Chen, Gongfan Fang, Xinyin Ma, and Xinchao Wang. Slimsam: 0.1% data makes segment anything slim, 2024

2024
[18]

Efficient quantization-aware training on segment anything model in medical images and its deployment

Haisheng Lu, Yujie Fu, Fan Zhang, and Le Zhang. Efficient quantization-aware training on segment anything model in medical images and its deployment. InMedical Image Segmentation Foundation Models. CVPR 2024 Challenge: Segment Anything in Medical Images on Laptop, pages 137–150. Springer Nature Switzerland, 2025

2024
[19]

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, pages 234–241, 2015

2015
[20]

Jaeger, Simon A

Fabian Isensee, Paul F. Jaeger, Simon A. A. Kohl, Jens Petersen, and Klaus H. Maier-Hein. nnu-net: A self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18:203–211, 2021

2021
[21]

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

Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. InProceedings of the European Conference on Computer Vision Workshops, 2022

2022
[22]

Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation

Md Mostafijur Rahman, Mustafa Munir, and Radu Marculescu. Emcad: Efficient multi-scale convolutional attention decoding for medical image segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11769–11779, 2024

2024
[23]

Mk-unet: Multi-kernel lightweight cnn for medical image segmentation, 2025

Md Mostafijur Rahman and Radu Marculescu. Mk-unet: Multi-kernel lightweight cnn for medical image segmentation, 2025

2025
[24]

HSSAM- Net: Hyper-scale shifted aggregation network for precise colorectal polyp segmentation in endoscopic images.Scientific Reports, 15:38146, 2025

Qing Feng, Shahzad Ahmed, Yueming Zhang, Lan He, and Muhammad Yaqub. HSSAM- Net: Hyper-scale shifted aggregation network for precise colorectal polyp segmentation in endoscopic images.Scientific Reports, 15:38146, 2025. doi: 10.1038/s41598-025-21954-y. URLhttps://www.nature.com/articles/s41598-025-21954-y

work page doi:10.1038/s41598-025-21954-y 2025
[25]

Denker, and Sara A

Yann LeCun, John S. Denker, and Sara A. Solla. Optimal brain damage. InAdvances in Neural Information Processing Systems, volume 2, 1989

1989
[26]

Babak Hassibi and David G. Stork. Second order derivatives for network pruning: Optimal brain surgeon. InAdvances in Neural Information Processing Systems, volume 5, 1993

1993
[27]

Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. InInternational Conference on Learning Representations, 2016

2016
[28]

Victor Sanh, Thomas Wolf, and Alexander M. Rush. Movement pruning: Adaptive sparsity by fine-tuning. InAdvances in Neural Information Processing Systems, volume 33, pages 20378–20389, 2020

2020
[29]

Group fisher pruning for practical network compression

Liyang Liu, Shilong Zhang, Zhanghui Kuang, Aojun Zhou, Jing-Hao Xue, Xinjiang Wang, Yimin Chen, Wenming Yang, Qingmin Liao, and Wayne Zhang. Group fisher pruning for practical network compression. InInternational Conference on Machine Learning, pages 7021–7032, 2021. 11

2021
[30]

Global vision transformer pruning with hessian-aware saliency

Huanrui Yang, Hongxu Yin, Pavlo Molchanov, Hai Li, and Jan Kautz. Global vision transformer pruning with hessian-aware saliency. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18547–18557, 2023

2023
[31]

Are sixteen heads really better than one? In Advances in Neural Information Processing Systems, volume 32, 2019

Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one? In Advances in Neural Information Processing Systems, volume 32, 2019

2019
[32]

Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned

Elena V oita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned. InAnnual Meeting of the Association for Computational Linguistics, pages 5797–5808, 2019

2019
[33]

Boundary loss for highly unbalanced segmentation

Hoel Kervadec, Jalal Bouchtiba, Christian Desrosiers, Eric Granger, Jose Dolz, and Ismail Ben Ayed. Boundary loss for highly unbalanced segmentation. InMedical Imaging with Deep Learning, pages 285–296, 2019

2019
[34]

Salcudean

Davood Karimi and Septimiu E. Salcudean. Reducing the hausdorff distance in medical image segmentation with convolutional neural networks.IEEE Transactions on Medical Imaging, 39 (2):499–513, 2020

2020
[35]

Paetzold, Anjany Sekuboyina, Ivan Ezhov, Andreas Unger, Andrey Zhylka, Josien P

Suprosanna Shit, Johannes C. Paetzold, Anjany Sekuboyina, Ivan Ezhov, Andreas Unger, Andrey Zhylka, Josien P. W. Pluim, Ulrich Bauer, and Bjoern H. Menze. cldice: A novel topology-preserving loss function for tubular structure segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16560–16569, 2021

2021
[36]

Polyper: Boundary sensitive polyp segmentation

Hao Shao, Yang Zhang, and Qibin Hou. Polyper: Boundary sensitive polyp segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4731–4739,
[37]

URL https://ojs.aaai.org/index.php/AAAI/ article/view/28274

doi: 10.1609/aaai.v38i5.28274. URL https://ojs.aaai.org/index.php/AAAI/ article/view/28274

work page doi:10.1609/aaai.v38i5.28274
[38]

BMFA-Net: Boundary constraint multi-level feature aggregation framework for precise polyp segmentation.Applied Sciences, 14(10):4063, 2024

Qin Li, Tianchi Zhang, Parvej Md Mosharaf, and Jing Zhang. BMFA-Net: Boundary constraint multi-level feature aggregation framework for precise polyp segmentation.Applied Sciences, 14(10):4063, 2024. doi: 10.3390/app14104063. URL https://www.mdpi.com/2076-3417/ 14/10/4063

work page doi:10.3390/app14104063 2024
[39]

FCA-Net: Fully context- aware feature aggregation network for medical segmentation.Biomedical Signal Process- ing and Control, 91:106004, 2024

Dingzhou Liu, Hongmin Deng, Zhengwei Huang, and Jinghao Fu. FCA-Net: Fully context- aware feature aggregation network for medical segmentation.Biomedical Signal Process- ing and Control, 91:106004, 2024. doi: 10.1016/j.bspc.2024.106004. URL https://www. sciencedirect.com/science/article/pii/S1746809424000624

work page doi:10.1016/j.bspc.2024.106004 2024
[40]

BEGA-UNet: Boundary-explicit guided attention U-Net with multi-scale feature aggregation for colonoscopic polyp segmentation

Tao Tong, Wen Zhang, and Wanni Zu. BEGA-UNet: Boundary-explicit guided attention U-Net with multi-scale feature aggregation for colonoscopic polyp segmentation. medRxiv preprint,
[41]

URLhttps://www.medrxiv.org/content/10.64898/2026.03.04.26347608v2

work page doi:10.64898/2026.03.04.26347608v2 2026
[42]

Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilariño

Jorge Bernal, F. Javier Sánchez, Gloria Fernández-Esparrach, Debora Gil, Cristina Rodríguez, and Fernando Vilariño. Wm-dova maps for accurate polyp highlighting in colonoscopy: Vali- dation vs. saliency maps from physicians.Computerized Medical Imaging and Graphics, 43: 99–111, 2015. ISSN 0895-6111. doi: https://doi.org/10.1016/j.compmedimag.2015.02.007. ...

work page doi:10.1016/j.compmedimag.2015.02.007 2015
[43]

Javier Sánchez, Gloria Fernández-Esparrach, Antonio M

David Vázquez, Jorge Bernal, F. Javier Sánchez, Gloria Fernández-Esparrach, Antonio M. López, Adriana Romero, Michal Drozdzal, and Aaron Courville. A benchmark for endoluminal scene segmentation of colonoscopy images.Journal of Healthcare Engineering, 2017(1): 4037190, 2017. doi: https://doi.org/10.1155/2017/4037190. URL https://onlinelibrary. wiley.com/d...

work page doi:10.1155/2017/4037190 2017
[44]

Smedsrud, Michael A

Debesh Jha, Pia H. Smedsrud, Michael A. Riegler, Pål Halvorsen, Thomas de Lange, Dag Johansen, and Håvard D. Johansen. Kvasir-seg: A segmented polyp dataset. In Yong Man Ro, Wen-Huang Cheng, Junmo Kim, Wei-Ta Chu, Peng Cui, Jung-Woo Choi, Min-Chun Hu, and Wesley De Neve, editors,MultiMedia Modeling, pages 451–462, Cham, 2020. Springer International Publis...

2020
[45]

Dataset of breast ultrasound images.Data in Brief, 28:104863, 2020

Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy. Dataset of breast ultrasound images.Data in Brief, 28:104863, 2020. ISSN 2352-3409. doi: https://doi.org/ 10.1016/j.dib.2019.104863. URL https://www.sciencedirect.com/science/article/ pii/S2352340919312181

work page doi:10.1016/j.dib.2019.104863 2020
[46]

Scientific Data5(1), 180161 (Aug 2018)

Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Scientific Data, 5(1), August 2018. ISSN 2052-4463. doi: 10.1038/sdata.2018.161. URL http://dx.doi.org/10. 1038/sdata.2018.161

work page doi:10.1038/sdata.2018.161 2018
[47]

Skin Lesion Analysis Toward Melanoma Detection 2018: A Challenge Hosted by the International Skin Imaging Collaboration (ISIC)

Noel Codella, Veronica Rotemberg, Philipp Tschandl, M. Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, Harald Kittler, and Allan Halpern. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic), 2019. URL https://arxiv.org/a...

work page internal anchor Pith review Pith/arXiv arXiv 2018
[48]

Revisiting random channel pruning for neural network compression, 2022

Yawei Li, Kamil Adamczewski, Wen Li, Shuhang Gu, Radu Timofte, and Luc Van Gool. Revisiting random channel pruning for neural network compression, 2022. URL https: //arxiv.org/abs/2205.05676

work page arXiv 2022
[49]

Song Han, Jeff Pool, John Tran, and William J. Dally. Learning both weights and connections for efficient neural networks, 2015. URLhttps://arxiv.org/abs/1506.02626

work page internal anchor Pith review Pith/arXiv arXiv 2015
[50]

Are sixteen heads really better than one?, 2019

Paul Michel, Omer Levy, and Graham Neubig. Are sixteen heads really better than one?, 2019. URLhttps://arxiv.org/abs/1905.10650

work page arXiv 2019
[51]

Faster gaze prediction with dense networks and Fisher pruning

Lucas Theis, Iryna Korshunova, Alykhan Tejani, and Ferenc Huszár. Faster gaze prediction with dense networks and fisher pruning, 2018. URL https://arxiv.org/abs/1801.05787. Appendix The appendix contains proofs, implementation details, and additional information about the boundary leverage analysis. We keep the main text focused on the method and key empi...

work page internal anchor Pith review Pith/arXiv arXiv 2018