arxiv: 2604.28095 · v1 · submitted 2026-04-30 · 💻 cs.CV

Recognition: unknown

UHR-Net: An Uncertainty-Aware Hypergraph Refinement Network for Medical Image Segmentation

Shuokun Cheng , Jinghao Shi , Kun Sun

Authors on Pith no claims yet

Pith reviewed 2026-05-07 04:53 UTC · model grok-4.3

classification 💻 cs.CV

keywords medical image segmentationhypergraph refinementuncertainty estimationlesion boundaryinstance contrastive pretrainingambiguous region handlingsmall lesion detection

0 comments

The pith

UHR-Net refines lesion boundaries by splitting hyperedge prototypes according to an entropy uncertainty map.

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

The paper presents UHR-Net to improve segmentation of lesions that blend into surrounding tissue or are too small for standard multi-scale features to capture reliably. It begins with Uncertainty-Oriented Instance Contrastive pretraining that combines geometry-aware copy-paste augmentation and hard-negative mining of lesion-like background patches. The core addition is the Uncertainty-Guided Hypergraph Refinement block, which builds an entropy map from a coarse prediction and uses it to separate hyperedge prototypes into distinct foreground and background groups. This separation prevents mixed interactions during higher-order refinement and focuses attention on ambiguous transition zones. If the approach holds, automated outlines become more consistent in exactly the regions clinicians find hardest to judge.

Core claim

The paper claims that deriving an entropy-based uncertainty map from an initial coarse probability map and then splitting hyperedge prototypes into separate foreground and background groups inside the Uncertainty-Guided Hypergraph Refinement block decouples higher-order relations, allowing targeted refinement of ill-defined boundaries and small-lesion cues that would otherwise be diluted.

What carries the argument

The Uncertainty-Guided Hypergraph Refinement block, which constructs an entropy uncertainty map to partition hyperedge prototypes into foreground and background groups so that refinement operates on decoupled higher-order interactions.

If this is right

Segmentation of small lesions and boundary zones becomes measurably more accurate without increasing model size.
Predictions in transition regions between lesion and healthy tissue gain stability across multiple imaging modalities.
The pretraining stage alone reduces under-segmentation of visually similar background patches.
Consistent gains on five public lesion datasets indicate the pattern can transfer to new clinical tasks.
Clinical tools gain a built-in focus on the most uncertain pixels rather than uniform post-processing.

Where Pith is reading between the lines

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

The same uncertainty-driven prototype split could be inserted into other graph-based refinement modules for non-medical images with fuzzy edges.
Pairing the entropy map with additional uncertainty signals such as model ensembles might further reduce error propagation.
Deploying the network in real-time diagnostic pipelines would test whether the refined boundaries change downstream clinical decisions.
The pretraining strategy might serve as a drop-in module for any segmentation backbone that struggles with instance discrimination.

Load-bearing premise

The entropy map computed from the coarse probability map reliably marks the precise regions that need refinement and that feeding this map into the hypergraph does not carry forward mistakes made in the coarse stage.

What would settle it

If a controlled ablation that removes the entropy-guided splitting of hyperedge prototypes into foreground and background groups produces equal or better Dice and boundary scores on the same five benchmarks, the central mechanism would be shown to add no value.

Figures

Figures reproduced from arXiv: 2604.28095 by Jinghao Shi, Kun Sun, Shuokun Cheng.

**Figure 1.** Figure 1: Overall framework of the proposed UHR-Net. The upper left part is the UO-IC pretraining stage. The lower left part is the end-to-end training view at source ↗

**Figure 2.** Figure 2: Qualitative comparisons on ISIC-2016 and Kvasir-SEG. Columns show Image, U-Net++, ESPNet, CMUNeXt, ConDSeg, Ours, and Ground Truth view at source ↗

**Figure 3.** Figure 3: t-SNE visualization of representations learned with and without view at source ↗

**Figure 4.** Figure 4: Visualization of entropy-based uncertainty maps derived from the view at source ↗

read the original abstract

Accurate lesion segmentation is crucial for clinical diagnosis and treatment planning. However, lesions often resemble surrounding tissues and exhibit ill-defined boundaries, leading to unstable predictions in boundary/transition regions. Moreover, small-lesion cues can be diluted by multi-scale feature extraction, causing under- or over-segmentation. To address these challenges, we propose an Uncertainty-Aware Hypergraph Refinement Network (UHR-Net). First, we introduce an Uncertainty-Oriented Instance Contrastive (UO-IC) pretraining strategy that couples geometry-aware copy-paste augmentation with hard-negative mining of lesion-like background regions to improve instance-level discrimination for small and visually ambiguous lesions. Second, we design an Uncertainty-Guided Hypergraph Refinement (UGHR) block, which derives an entropy-based uncertainty map from a coarse probability map to guide hypergraph refinement. By splitting hyperedge prototypes into foreground and background groups, UGHR decouples higher-order interactions and improves refinement in ambiguous regions. Experiments on five public benchmarks demonstrate consistent gains over strong baselines. Code is available at: https://github.com/CUGfreshman/UHR-Net.

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

UHR-Net adds a contrastive pretraining step and an entropy-guided hypergraph split for lesion boundaries, but the refinement still depends on the coarse stage not being too wrong.

read the letter

The paper's main moves are the UO-IC pretraining, which does geometry-aware copy-paste of lesions plus hard-negative mining against lesion-like background, and the UGHR block that turns a coarse probability map into an entropy uncertainty map, then splits hyperedge prototypes into foreground and background groups to refine ambiguous regions. Both pieces are new relative to the cited prior work on contrastive learning and hypergraph segmentation. The target problem—fuzzy boundaries and diluted small-lesion signals—is real and common in medical imaging, so the framing is on point. The abstract's claim of consistent gains across five public benchmarks is the kind of empirical check that matters here. The full paper presumably supplies the numbers, ablations, and code that the abstract omits. The soft spot is the one the stress-test note flags: the uncertainty signal is derived directly from the coarse output, so any under- or over-segmentation in transition zones or small lesions can misdirect the hypergraph grouping and make the refinement step reinforce rather than correct the initial errors. Without seeing explicit tests (for example, swapping the entropy map for ground-truth uncertainty or random splits), it is hard to know how much the pretraining actually stabilizes the coarse stage. The rest of the architecture looks like standard medical segmentation practice with no obvious circularity or parameter-fitting tricks. This is useful reading for people already working on graph or uncertainty refinements for medical images; the specific design choices and benchmark numbers could be worth borrowing or extending. I would send it to peer review—the ideas are clear, the problem is practical, and the claimed gains are testable even if the uncertainty guidance needs tighter validation.

Referee Report

3 major / 3 minor

Summary. The paper proposes UHR-Net for medical image segmentation to handle ill-defined lesion boundaries and small lesions. It introduces an Uncertainty-Oriented Instance Contrastive (UO-IC) pretraining strategy using geometry-aware copy-paste augmentation and hard-negative mining of lesion-like regions, plus an Uncertainty-Guided Hypergraph Refinement (UGHR) block that computes an entropy-based uncertainty map from a coarse probability output to split hyperedge prototypes into foreground/background groups, decoupling higher-order interactions for refinement in ambiguous areas. The authors report consistent gains over strong baselines on five public benchmarks, with code released.

Significance. If the empirical improvements hold under rigorous validation, the work could advance uncertainty-aware refinement techniques in medical segmentation by combining contrastive pretraining for small-lesion discrimination with hypergraph modeling of higher-order relations. The public code release supports reproducibility and is a clear strength.

major comments (3)

[Section 3.2] UGHR block (Section 3.2): The entropy-based uncertainty map is derived directly from the coarse probability map and used to split hyperedge prototypes into foreground and background groups. This guidance assumes the coarse map reliably flags true ambiguities without inheriting its own boundary or small-lesion errors; if the initial prediction under-segments transition zones, the hypergraph refinement risks amplifying mistakes rather than correcting them. The manuscript should add ablations (e.g., UGHR with vs. without uncertainty splitting, or with deliberately perturbed coarse maps) and qualitative comparisons of uncertainty maps against ground-truth error regions to substantiate that the mechanism improves rather than propagates errors.
[Experiments section] Results (Tables 1–5 and associated text): The claim of 'consistent gains' on five benchmarks is central but currently lacks reported standard deviations across runs, statistical significance tests (e.g., paired t-tests or Wilcoxon), and detailed per-dataset metrics with error bars. Without these, it is impossible to determine whether improvements exceed baseline variability or arise from post-hoc choices; the paper must include these to support the performance claims.
[Section 3.1] UO-IC pretraining (Section 3.1): The hard-negative mining thresholds and hypergraph construction parameters are free parameters. The manuscript should specify their selection procedure across the five datasets and report sensitivity analysis, as these choices directly affect the instance-level discrimination for small lesions and could influence the downstream UGHR performance.

minor comments (3)

[Abstract / Introduction] The abstract and introduction should explicitly name the five public benchmarks early for clarity.
[Method] Notation for hyperedge prototypes, uncertainty map, and splitting operation would benefit from explicit equations or a clear algorithmic listing.
[Figure 1] Network architecture diagrams could include annotations for the UO-IC and UGHR components to improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough review and valuable suggestions. We have addressed all major comments by providing clarifications, additional analyses, and revisions to the manuscript. Below we respond point-by-point to the major comments.

read point-by-point responses

Referee: [Section 3.2] UGHR block (Section 3.2): The entropy-based uncertainty map is derived directly from the coarse probability map and used to split hyperedge prototypes into foreground and background groups. This guidance assumes the coarse map reliably flags true ambiguities without inheriting its own boundary or small-lesion errors; if the initial prediction under-segments transition zones, the hypergraph refinement risks amplifying mistakes rather than correcting them. The manuscript should add ablations (e.g., UGHR with vs. without uncertainty splitting, or with deliberately perturbed coarse maps) and qualitative comparisons of uncertainty maps against ground-truth error regions to substantiate that the mechanism improves rather than propagates errors.

Authors: We agree that validating the uncertainty guidance is crucial to ensure it does not propagate errors. The design of UGHR uses the uncertainty map to identify ambiguous regions and decouple the hypergraph interactions accordingly, which is intended to focus refinement efforts where the coarse prediction is least reliable. To substantiate this, we have added new ablation studies in the revised manuscript comparing the full UGHR with variants that disable the uncertainty-based splitting. We also include qualitative comparisons showing the uncertainty maps overlaid with ground-truth error regions, demonstrating that high-uncertainty areas correspond to boundary and small-lesion errors, and that refinement corrects them effectively. These additions confirm the mechanism's benefit. revision: yes
Referee: [Experiments section] Results (Tables 1–5 and associated text): The claim of 'consistent gains' on five benchmarks is central but currently lacks reported standard deviations across runs, statistical significance tests (e.g., paired t-tests or Wilcoxon), and detailed per-dataset metrics with error bars. Without these, it is impossible to determine whether improvements exceed baseline variability or arise from post-hoc choices; the paper must include these to support the performance claims.

Authors: We acknowledge the importance of statistical rigor in reporting results. In the revised version, we have rerun the experiments with multiple random seeds to compute standard deviations for all metrics across the five datasets. We have added these to Tables 1-5 along with error bars in the associated figures. Furthermore, we performed paired t-tests between UHR-Net and the baselines, reporting p-values to demonstrate statistical significance of the improvements. This strengthens the evidence for consistent gains. revision: yes
Referee: [Section 3.1] UO-IC pretraining (Section 3.1): The hard-negative mining thresholds and hypergraph construction parameters are free parameters. The manuscript should specify their selection procedure across the five datasets and report sensitivity analysis, as these choices directly affect the instance-level discrimination for small lesions and could influence the downstream UGHR performance.

Authors: The parameters for hard-negative mining (e.g., similarity threshold) and hypergraph construction (e.g., number of hyperedges) were determined through grid search on the validation sets of each dataset to optimize the pretraining loss. We have expanded Section 3.1 to detail this cross-validation-based selection procedure for all five benchmarks. Additionally, we have included a sensitivity analysis in the supplementary material, varying each parameter over a range and showing the impact on downstream segmentation performance, which indicates robustness to moderate variations. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical architecture proposal with independent validation

full rationale

The paper introduces UHR-Net as a novel segmentation architecture featuring UO-IC pretraining and the UGHR block. The UGHR block derives an entropy-based uncertainty map from a coarse probability map to split hyperedge prototypes, but this is an explicit design choice for guiding refinement rather than a derivation that reduces to its own inputs by construction. No equations are presented that equate a 'prediction' to a fitted parameter, no load-bearing self-citations justify uniqueness theorems, and no ansatz is smuggled via prior work. The central claims rest on experimental gains across five benchmarks, which are externally falsifiable and do not collapse to self-referential quantities. The derivation chain is self-contained as a proposed method with stated assumptions about uncertainty guidance.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 2 invented entities

The central claim rests on the empirical effectiveness of two newly introduced mechanisms whose internal hyperparameters and the reliability of the entropy-derived uncertainty signal are not independently verified outside the reported benchmarks.

free parameters (2)

hypergraph construction parameters
Number of hyperedges, prototype splitting thresholds, and entropy weighting factors are chosen during design or tuning and directly affect the refinement behavior.
contrastive mining thresholds
Hard-negative selection criteria and copy-paste geometry parameters are set to improve discrimination for small lesions.

axioms (2)

domain assumption A coarse probability map yields an entropy map that accurately flags boundary and transition regions without systematic bias.
Invoked to justify using the uncertainty map to guide hypergraph refinement.
domain assumption Higher-order pixel interactions can be usefully decoupled by splitting hyperedge prototypes into foreground and background groups.
Core design choice of the UGHR block.

invented entities (2)

Uncertainty-Guided Hypergraph Refinement (UGHR) block no independent evidence
purpose: To refine segmentation predictions in ambiguous regions by using an entropy uncertainty map to split and process hyperedges separately for foreground and background.
New architectural module introduced by the paper; no independent evidence outside the reported experiments.
Uncertainty-Oriented Instance Contrastive (UO-IC) pretraining strategy no independent evidence
purpose: To improve instance-level discrimination for small and visually ambiguous lesions via geometry-aware copy-paste augmentation and hard-negative mining.
New pretraining procedure introduced by the paper; no independent evidence outside the reported experiments.

pith-pipeline@v0.9.0 · 5494 in / 1691 out tokens · 97436 ms · 2026-05-07T04:53:08.470279+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

30 extracted references · 1 canonical work pages

[1]

Hardnet-cps: colorectal polyp segmen- tation based on harmonic densely united network,

Tong Yu and Qingxiang Wu, “Hardnet-cps: colorectal polyp segmen- tation based on harmonic densely united network,”Biomedical Signal Processing and Control, vol. 85, pp. 104953, 2023

2023
[2]

Lesion-aware dynamic kernel for polyp segmentation,

Ruifei Zhang, Peiwen Lai, Xiang Wan, De-Jun Fan, Feng Gao, Xiao-Jian Wu, and Guanbin Li, “Lesion-aware dynamic kernel for polyp segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2022, pp. 99–109

2022
[3]

H2former: An efficient hierarchical hybrid transformer for medical image segmentation,

Along He, Kai Wang, Tao Li, Chengkun Du, Shuang Xia, and Huazhu Fu, “H2former: An efficient hierarchical hybrid transformer for medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 42, no. 9, pp. 2763–2775, 2023

2023
[4]

Vig- unet: vision graph neural networks for medical image segmentation,

Juntao Jiang, Xiyu Chen, Guanzhong Tian, and Yong Liu, “Vig- unet: vision graph neural networks for medical image segmentation,” in2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). IEEE, 2023, pp. 1–5

2023
[5]

What uncertainties do we need in bayesian deep learning for computer vision?,

Alex Kendall and Yarin Gal, “What uncertainties do we need in bayesian deep learning for computer vision?,”Advances in neural information processing systems, vol. 30, 2017

2017
[6]

Ex- ploring uncertainty measures in deep networks for multiple sclerosis lesion detection and segmentation,

Tanya Nair, Doina Precup, Douglas L Arnold, and Tal Arbel, “Ex- ploring uncertainty measures in deep networks for multiple sclerosis lesion detection and segmentation,”Medical image analysis, vol. 59, pp. 101557, 2020

2020
[7]

Unified medical image segmentation by learning from uncertainty in an end-to-end manner,

Pin Tang, Pinli Yang, Dong Nie, Xi Wu, Jiliu Zhou, and Yan Wang, “Unified medical image segmentation by learning from uncertainty in an end-to-end manner,”Knowledge-Based Systems, vol. 241, pp. 108215, 2022

2022
[8]

Uacanet: Uncertainty augmented context attention for polyp segmentation,

Taehun Kim, Hyemin Lee, and Daijin Kim, “Uacanet: Uncertainty augmented context attention for polyp segmentation,” inProceedings of the 29th ACM international conference on multimedia, 2021, pp. 2167–2175

2021
[9]

Hypergraph neural networks,

Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao, “Hypergraph neural networks,” inProceedings of the AAAI conference on artificial intelligence, 2019, vol. 33, pp. 3558–3565

2019
[10]

Hgnn+: General hypergraph neural networks,

Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji, “Hgnn+: General hypergraph neural networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3181–3199, 2022

2022
[11]

Vision hgnn: An image is more than a graph of nodes,

Yan Han, Peihao Wang, Souvik Kundu, Ying Ding, and Zhangyang Wang, “Vision hgnn: An image is more than a graph of nodes,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19878–19888

2023
[12]

U-net: Convolutional networks for biomedical image segmentation,

Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer- assisted intervention. Springer, 2015, pp. 234–241

2015
[13]

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

Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang, “Unet++: A nested u-net architecture for medical image segmentation,” inInternational workshop on deep learning in medical image analysis. Springer, 2018, pp. 3–11

2018
[14]

Pranet: Parallel reverse attention network for polyp segmentation,

Deng-Ping Fan, Ge-Peng Ji, Tao Zhou, Geng Chen, Huazhu Fu, Jian- bing Shen, and Ling Shao, “Pranet: Parallel reverse attention network for polyp segmentation,” inInternational conference on medical image computing and computer-assisted intervention. Springer, 2020, pp. 263–273

2020
[15]

Tganet: Text-guided attention for improved polyp segmentation,

Nikhil Kumar Tomar, Debesh Jha, Ulas Bagci, and Sharib Ali, “Tganet: Text-guided attention for improved polyp segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2022, pp. 151–160

2022
[16]

Dcsau-net: A deeper and more compact split-attention u-net for medical image segmentation,

Qing Xu, Zhicheng Ma, Wenting Duan, et al., “Dcsau-net: A deeper and more compact split-attention u-net for medical image segmentation,”Computers in Biology and Medicine, vol. 154, pp. 106626, 2023

2023
[17]

Casf-net: Cross-attention and cross-scale fusion network for medical image segmentation,

Jianwei Zheng, Hao Liu, Yuchao Feng, Jinshan Xu, and Liang Zhao, “Casf-net: Cross-attention and cross-scale fusion network for medical image segmentation,”Computer Methods and Programs in Biomedicine, vol. 229, pp. 107307, 2023

2023
[18]

Dtan: Diffusion- based text attention network for medical image segmentation,

Yiyang Zhao, Jinjiang Li, Lu Ren, and Zheng Chen, “Dtan: Diffusion- based text attention network for medical image segmentation,”Com- puters in Biology and Medicine, vol. 168, pp. 107728, 2024

2024
[19]

Cmunext: An efficient medical image segmentation network based on large kernel and skip fusion,

Fenghe Tang, Jianrui Ding, Quan Quan, Lingtao Wang, Chunping Ning, and S Kevin Zhou, “Cmunext: An efficient medical image segmentation network based on large kernel and skip fusion,” in2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–5

2024
[20]

Espnet: Edge-aware feature shrinkage pyramid for polyp segmentation,

Raneem Toman, Venkataraman Subramanian, and Sharib Ali, “Espnet: Edge-aware feature shrinkage pyramid for polyp segmentation,” inIn- ternational Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2025, pp. 157–167

2025
[21]

Condseg: A general medical image segmentation framework via contrast-driven feature enhancement,

Mengqi Lei, Haochen Wu, Xinhua Lv, and Xin Wang, “Condseg: A general medical image segmentation framework via contrast-driven feature enhancement,” inProceedings of the AAAI Conference on Artificial Intelligence, 2025, vol. 39, pp. 4571–4579

2025
[22]

Ce- net: Context encoder network for 2d medical image segmentation,

Zaiwang Gu, Jun Cheng, Huazhu Fu, Kang Zhou, Huaying Hao, Yitian Zhao, Tianyang Zhang, Shenghua Gao, and Jiang Liu, “Ce- net: Context encoder network for 2d medical image segmentation,” IEEE transactions on medical imaging, vol. 38, no. 10, pp. 2281– 2292, 2019

2019
[23]

Fat-net: Feature adaptive transformers for automated skin lesion segmentation,

Huisi Wu, Shihuai Chen, Guilian Chen, Wei Wang, Baiying Lei, and Zhenkun Wen, “Fat-net: Feature adaptive transformers for automated skin lesion segmentation,”Medical image analysis, vol. 76, pp. 102327, 2022

2022
[24]

Bgdiffseg: A fast diffusion model for skin lesion segmentation via boundary enhancement and global recognition guidance,

Yilin Guo and Qingling Cai, “Bgdiffseg: A fast diffusion model for skin lesion segmentation via boundary enhancement and global recognition guidance,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 150–159

2024
[25]

Mscb- unet: Elevating skin lesion segmentation performance with multi-scale spatial-channel bridging network,

Yi Wang, Hanwen Zhang, Jingxin Fu, and Hong Tian, “Mscb- unet: Elevating skin lesion segmentation performance with multi-scale spatial-channel bridging network,”Biomedical Signal Processing and Control, vol. 110, pp. 107986, 2025

2025
[26]

Kvasir- seg: A segmented polyp dataset,

Debesh Jha, Pia H Smedsrud, Michael A Riegler, P ˚al Halvorsen, Thomas De Lange, Dag Johansen, and H ˚avard D Johansen, “Kvasir- seg: A segmented polyp dataset,” inInternational conference on multimedia modeling. Springer, 2019, pp. 451–462

2019
[27]

A comprehensive study on colorectal polyp segmentation with resunet++, conditional random field and test-time augmentation,

Debesh Jha, Pia H Smedsrud, Dag Johansen, Thomas De Lange, H˚avard D Johansen, P ˚al Halvorsen, and Michael A Riegler, “A comprehensive study on colorectal polyp segmentation with resunet++, conditional random field and test-time augmentation,”IEEE journal of biomedical and health informatics, vol. 25, no. 6, pp. 2029–2040, 2021

2029
[28]

Gland segmentation in colon histology images: The glas challenge contest,

Korsuk Sirinukunwattana, Josien PW Pluim, Hao Chen, Xiaojuan Qi, Pheng-Ann Heng, Yun Bo Guo, Li Yang Wang, Bogdan J Ma- tuszewski, Elia Bruni, Urko Sanchez, et al., “Gland segmentation in colon histology images: The glas challenge contest,”Medical image analysis, vol. 35, pp. 489–502, 2017

2017
[29]

David Gutman, Noel CF Codella, Emre Celebi, Brian Helba, Michael Marchetti, Nabin Mishra, and Allan Halpern, “Skin lesion analysis toward melanoma detection: A challenge at the international sympo- sium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic),”arXiv preprint arXiv:1605.01397, 2016

work page arXiv 2016
[30]

IEEE, 2018, pp

Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al., “Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isi...

2017