Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

SafeClick: Error-Tolerant Interactive Segmentation of Any Medical Volumes via Hierarchical Expert Consensus

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A plug-in module lets medical segmentation survive imperfect clicks, lifting Dice by up to 21 points.

desk verdict SafeClick's prompt-robustness gains are real but rest on a single synthetic noise model; with broader evaluation it could be a solid contribution. read the letter →

arxiv 2506.18404 v1 pith:FOF6JN65 submitted 2025-06-23 eess.IV

Yifan Gao , Jiaxi Sheng , Wenbin Wu , Haoyue Li , Yaoxian Dong , Chaoyang Ge , Feng Yuan , Xin Gao This is my paper
classification eess.IV
keywords interactivemedicalimagesegmentationSegmentAnythingModel2error-tolerantpromptingexpertconsensusvolumetricpointandboxpromptsplug-and-playmodule
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

SafeClick claims that the brittleness of interactive medical segmentation under imperfect user prompts can be largely repaired by a small plug-in module, without retraining the base model. On 15 public volumetric datasets, adding SafeClick to SAM 2 and MedSAM 2 improves average Dice by 17.59 and 8.77 percentage points for noisy point prompts and by 10.96 and 6.42 points for noisy bounding boxes, across all tested model sizes. The mechanism is a hierarchical expert consensus: three transformer experts read the image from complementary viewpoints, and a consensus layer cross-references their features before producing the mask. If this holds, radiologists gain reliable segmentation even when clicks are off-target or boxes are loosely drawn.

What carries the argument

The load-bearing mechanism is the pair of CEL and CRL. CEL contains three experts: E1 cross-attends intermediate image features to final image features, E2 self-attends final image features independently of the prompt, and E3 cross-attends prompt features with image features, mirroring the original mask decoder. CRL reshapes expert features into channel-first matrices, computes a contrastive cross-reference attention matrix and a self-attention matrix, blends them with a learnable parameter, and modulates the features before combining them with prompt features through a zero-initialized convolution. The zero-initialized path means training starts from the base model's behavior, preserving its strengths while gradually learning how to down-weight unreliable prompt information.

What would settle it

Run SafeClick and its base model on prompts drawn from real clinical interaction logs rather than synthetic noise, for example clicks recorded from radiologists on the same volumes; if the average Dice advantage over the base model disappears or falls well below the reported 17.59 and 8.77 points under those conditions, the simulated perturbation model was not representative.

Watch

Extended reading notes

Core claim

The paper's central claim is that segmentation quality under imperfect prompts is recoverable by converting a single prompt-conditioned decoder into a consensus of specialized experts. With point displacements of 25% to 100% of object radius and box scales of 50% to 150%, SafeClick consistently beats both zero-shot and fine-tuned base models on all 15 datasets, with the largest gains appearing under the worst prompts. The method introduces a collaborative expert layer (CEL) with three transformer experts and a consensus reasoning layer (CRL) that fuses their outputs through cross-referencing attention. This architecture is presented as transforming segmentation from a prompt-dependent operation into a framework that can weigh image-intrinsic evidence against prompt evidence, preserving accuracy as prompt quality deteriorates.

Load-bearing premise

The simulated prompt errors, random point displacements of 25% to 100% of object radius and box scales of 50% to 150%, are assumed to represent the suboptimal prompts clinicians actually give; if real errors are systematically different, such as clicks outside the target or biased placements, the reported gains may not generalize.

Editorial extensions

If this is right

  • SafeClick works across ViT-T, ViT-S, and ViT-L sizes of both SAM 2 and MedSAM 2, so the robustness gain does not require architecture-specific retraining.
  • Performance gains grow as prompt quality worsens, with the largest improvement (21.21 Dice points) occurring for severely mis-scaled boxes, matching the clinical situation where precise box placement is difficult.
  • Point prompts, which carry minimal spatial guidance, show consistent improvements at every tested displacement level, indicating the module supplements sparse prompt evidence with image features.
  • Because integration starts from a zero-initialized convolution, the module can be inserted into existing pipelines without degrading perfect-prompt performance, while still recovering accuracy on noisy prompts.
  • The roughly 18% increase in inference time is presented as a modest cost for substantial gains in reliability across diverse anatomical regions and imaging modalities.

Reading between the lines

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

  • An implication the paper leaves implicit: the same consensus design should transfer to other prompt modalities such as scribbles, negative clicks, or clicks placed entirely outside the target, since the mechanism does not depend on how the noise was generated; those patterns should be measured directly.
  • A testable extension: because the self-attention expert E2 is prompt-independent, removing the prompt branch entirely should still yield usable masks, predicting that SafeClick can also serve as a weak automatic segmenter on the same 15 datasets.
  • An architectural prediction: the cross-reference attention matrix in CRL can be read as a confidence map of the prompt channel, so visualizing it on a failing prompt should show the method discounting unreliable prompt features, which is a direct, checkable consequence of the design.
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 / 5 minor

Summary. The paper presents SafeClick, a plug-and-play module for SAM 2 and MedSAM 2 that targets error-tolerant interactive segmentation of medical volumes. It consists of a collaborative expert layer (CEL) with three transformer branches and a consensus reasoning layer (CRL) that fuses their outputs with learnable weights. The authors report consistent Dice improvements over zero-shot and fine-tuned baselines across 15 public datasets under both perfect and synthetically perturbed point and box prompts, with average gains of 17.59% and 10.96% for SAM 2 under imperfect point and box prompts, respectively.

Significance. If the reported results are correct, SafeClick offers a practical way to make interactive medical segmentation more robust to imperfect user prompts without modifying the base foundation model, which is clinically relevant. The idea of hierarchical consensus, combining prompt-dependent and prompt-independent features, is reasonable and the code is released. However, the evidence is currently undermined by the reliance on a single synthetic perturbation model, arithmetic inconsistencies in the main table, and the absence of statistical validation, so the reported gains are not yet established.

major comments (4)
  1. [Table 2] The 'Avg.' columns in Table 2 are not consistent with the per-condition values listed in the same row. For example, for zero-shot SAM 2 ViT-T point prompts, the five values (44.08, 43.72, 42.69, 38.57, 33.79) average to 40.57, but the table reports 39.69. Similarly, for SafeClick (SAM 2) ViT-L point prompts, the mean of the five values is 73.48, but the reported average is 72.25. The 'Improve' row for the SAM 2 bounding-box condition also does not match the actual differences between the SafeClick and fine-tuned SAM 2 columns: the listed improvements 1.61, 21.21, 5.97, 4.83, 11.83 differ from the computed differences 1.52, 17.54, 5.61, 3.23, 9.13. Because the headline figures (17.59% and 10.96%) are derived from these averages, the quantitative claims of the paper are not verifiable as printed. Please correct the table and clearly define how 'Avg.' is computed.
  2. [Section 3 (Experiments and Results)] The entire robustness evaluation relies on a single synthetic perturbation generator: point displacements from the center of mass by 25-100% of object radius, and centered bounding-box scaling by 50-150%. The same noise family is used for both training and evaluation, so SafeClick can learn this specific distribution. The paper does not test generalization to other plausible sources of clinical prompt error, such as off-target clicks, biased (non-centered) box shifts, or sequential multi-click corrections. Consequently, the conclusion that SafeClick is 'error-tolerant' in clinical practice is not supported. Please add at least one cross-noise evaluation (e.g., train on the current uniform displacement protocol and test on a different distribution) or a human-prompt study; at a minimum, explicitly acknowledge and discuss this limitation.
  3. [Section 3 (Implementation and Experiment Setting), Table 2] No error bars, confidence intervals, or significance tests are reported anywhere. All results in Table 2 and the ablation in Table 3 are point estimates from a single run, with no statement of random seeds or number of repetitions. Given the scale of the claimed improvements (up to 21 Dice points), it is important to know whether these differences are stable across different data splits and initializations. Please provide standard deviations over at least three runs, or use paired significance tests across the 15 datasets.
  4. [Section 3 (Results), Table 2] The paper compares SafeClick only against zero-shot and fine-tuned versions of SAM 2 and MedSAM 2. It does not compare against any alternative robustification strategy, such as test-time prompt ensembling, multi-prompt averaging, or other adapters already cited (e.g., SAM2-Adapter, Ref. [8]). Without such baselines, it is unclear whether the observed gains come from the specific hierarchical consensus design or simply from additional learned parameters on top of the frozen encoder. Please add at least one simple robustification baseline that uses the same amount of additional computation or parameters.
minor comments (5)
  1. [Table 1] The body-region labels for FLARE22 and MMWHS are incorrect: FLARE22 is an abdominal organ segmentation dataset, and MMWHS is a whole-heart segmentation dataset, not 'Brain' as listed. Please correct these entries so that the dataset description is accurate.
  2. [Table 3] The meaning of the 'Baseline' column in the ablation table is unclear. It should be explicitly defined as the fine-tuned base model without any SafeClick component, or as the full SafeClick model, so that the 'w/o' rows are interpretable. Currently the reader cannot tell what the baseline refers to.
  3. [Section 2, Eq. (7)] The 'contrastive' normalization in Eq. (7), which subtracts the dot-product from the global maximum, is unusual and not intuitively explained. Please provide a brief justification or an ablation of this specific design choice, since it is central to the CRL.
  4. [Section 3, Fig. 3] The x-axis label 'Shift' in Figure 3 is not defined. Please clarify whether it refers to the displacement percentage or the scale percentage, and use consistent labels with Table 2.
  5. [Section 3, Ablation Study] There is a typo in the ablation description: 'When the self-attention transformer layer (E 2 is removed' is missing a closing parenthesis after 'E 2'. Additionally, the sentence about E3 being retained because removing it would change the prompt interaction mode is not an ablation in the usual sense; it would be informative to report what happens when E3 is excluded even if it changes the prompt processing, to quantify its contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; empirical gains on held-out volumes under synthetic prompt noise, with external validity caveats but no derivation that reduces to its inputs.

full rationale

This is an empirical deep-learning study; there is no chain of equations in which an output quantity is defined in terms of, or fitted to, the quantity it is claimed to predict. The reported Dice improvements are computed on held-out test volumes after a 7:1:2 random split ("random partitioning into training, validation, and test sets in a 7:1:2 ratio"), so the Table 2 numbers are not fit-derived. SafeClick's ablations break down component contributions on ASC18, and the claimed compatibility with SAM 2 and MedSAM 2 is tested across 15 public datasets. The main caveat is external validity rather than circularity: imperfect prompts are simulated by one generator ("random displacements from the center of mass ranging from 25% to 100% of the object radius; for bounding boxes, we scale the perfect box by factors ranging from 50% to 150%"), and the paper itself concedes that "extremely poor prompts that completely miss the target region remain challenging." That limits the clinical generalization of the robustness claim but does not make the result equivalent to its inputs. Self-citations in the reference list are not load-bearing; no uniqueness theorem or prior-work ansatz is invoked to force the architecture or the evaluation.

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

The central claim is an empirical performance improvement, so there is no derivation with fitted constants. The free parameters listed are learned or hand-chosen values relevant to the network design. The axioms are domain assumptions about transfer, evaluation, and the ad hoc contrastive attention formula. The paper invents no new physical or conceptual entities.

free parameters (5)
  • learning rate = 1e-4
    AdamW learning rate chosen by hand; no sensitivity analysis provided.
  • batch size = 8
    Chosen by hand; no sensitivity analysis provided.
  • training epochs = 20
    Chosen by hand; no early stopping or alternative epoch counts reported.
  • fusion weight alpha = not reported
    Learnable scalar in Eq. 8 balancing self-attention and cross-attention; final value not given.
  • zero-initialized convolution weights = not reported
    F_conv in Eq. 9 is zero-initialized and learned; final values not reported.
assumptions (5)
  • domain assumption Pre-trained SAM2 and MedSAM2 encoders provide features useful for medical images
    The paper freezes encoder weights and only trains the SafeClick module (Section 3), assuming generic encoder features transfer to medical volumes.
  • domain assumption Random splits of each dataset into 7:1:2 are representative
    Used to justify the evaluation protocol; no repeated splits or cross-validation are reported.
  • ad hoc to paper The contrastive normalization in Eq. 7 (max minus dot product) is a valid attention mechanism
    The formula is introduced without theoretical or empirical justification beyond the reported results; it is designed for this paper.
  • ad hoc to paper Simulated prompt perturbations match real clinical error distributions
    Section 3 defines the simulation; no evidence links these ranges to measured clinical click accuracy.
  • domain assumption Standard deep learning infrastructure delivers reliable results
    Relies on PyTorch, NVIDIA H100 GPUs, and common training practices; no verification of implementation correctness is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SafeClick: Error-Tolerant Interactive Segmentation of Any Medical Volumes via Hierarchical Expert Consensus." pith.science (2026). https://pith.science/paper/FOF6JN65

@misc{pith2026250618404,
  author       = {Pith},
  title        = {Pith review of: SafeClick: Error-Tolerant Interactive Segmentation of Any Medical Volumes via Hierarchical Expert Consensus},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FOF6JN65}},
  note         = {Machine review of arXiv:2506.18404}
}
read the original abstract

Foundation models for volumetric medical image segmentation have emerged as powerful tools in clinical workflows, enabling radiologists to delineate regions of interest through intuitive clicks. While these models demonstrate promising capabilities in segmenting previously unseen anatomical structures, their performance is strongly influenced by prompt quality. In clinical settings, radiologists often provide suboptimal prompts, which affects segmentation reliability and accuracy. To address this limitation, we present SafeClick, an error-tolerant interactive segmentation approach for medical volumes based on hierarchical expert consensus. SafeClick operates as a plug-and-play module compatible with foundation models including SAM 2 and MedSAM 2. The framework consists of two key components: a collaborative expert layer (CEL) that generates diverse feature representations through specialized transformer modules, and a consensus reasoning layer (CRL) that performs cross-referencing and adaptive integration of these features. This architecture transforms the segmentation process from a prompt-dependent operation to a robust framework capable of producing accurate results despite imperfect user inputs. Extensive experiments across 15 public datasets demonstrate that our plug-and-play approach consistently improves the performance of base foundation models, with particularly significant gains when working with imperfect prompts. The source code is available at https://github.com/yifangao112/SafeClick.

Figures

Figures reproduced from arXiv: 2506.18404 by the authors.

Figure 1
Figure 1. Performance comparison between SafeClick and baseline foundation models across different datasets. 2 Methodology [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the proposed SafeClick. The diagram illustrates the two pri￾mary components: the collaborative expert layer (CEL) that generates diverse feature representations through specialized transformer modules, and the consensus reasoning layer (CRL) that performs cross-referencing and adaptive integration of these features. The framework operates as a plug-and-play module compatible with foundation models fo… view at source ↗
Figure 3
Figure 3. Detailed results per dataset with box prompt (left) and point prompt (right) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results in the CHAOS dataset under box prompts. The red star indicates the location of the point prompt. GT: ground truth. Results [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LRMR: LLM-Driven Relational Multi-node Ranking for Lymph Node Metastasis Assessment in Rectal Cancer

    cs.LG 2025-07 reject novelty 5.0 of 10

    A two-stage LLM pipeline that converts lymph node MRI patches into structured reports and ranks patients by pairwise text comparisons reached AUC 0.79 on 36 test patients, marginally above ResNet50.

Reference graph

Works this paper leans on

28 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [8]

    Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more.arXiv preprint arXiv:2408.04579, 2024

    Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more.arXiv preprint arXiv:2408.04579, 2024

    arXiv 2024

  2. [1]

    An anatomy-aware framework for automatic segmentation of parotid tumor from multimodal mri

    Yifan Gao, Yin Dai, Fayu Liu, Weibing Chen, and Lifu Shi. An anatomy-aware framework for automatic segmentation of parotid tumor from multimodal mri. Computers in Biology and Medicine, 161:107000, 2023

    2023

  3. [2]

    Mba-net: Sam-driven bidirec- tional aggregation network for ovarian tumor segmentation

    Yifan Gao, Wei Xia, Wenkui Wang, and Xin Gao. Mba-net: Sam-driven bidirec- tional aggregation network for ovarian tumor segmentation. InInternational Con- ference on Medical Image Computing and Computer-Assisted Intervention, pages 437–447. Springer, 2024

    work page 2024

  4. [3]

    Wega: Weakly-supervised global-local affinity learning framework for lymph node metastasis prediction in rectal cancer.arXiv preprint arXiv:2505.10502, 2025

    Yifan Gao, Yaoxian Dong, Wenbin Wu, Chaoyang Ge, Feng Yuan, Jiaxi Sheng, Haoyue Li, and Xin Gao. Wega: Weakly-supervised global-local affinity learning framework for lymph node metastasis prediction in rectal cancer.arXiv preprint arXiv:2505.10502, 2025

    arXiv 2025

  5. [4]

    Interactive medical image segmentation using deep learning with image-specific fine tuning.IEEE transactions on medical imaging, 37(7):1562–1573, 2018

    Guotai Wang, Wenqi Li, Maria A Zuluaga, Rosalind Pratt, Premal A Patel, Michael Aertsen, Tom Doel, Anna L David, Jan Deprest, S´ ebastien Ourselin, et al. Interactive medical image segmentation using deep learning with image-specific fine tuning.IEEE transactions on medical imaging, 37(7):1562–1573, 2018

    work page 2018

  6. [5]

    Mideepseg: Minimally interactive segmentation of unseen objects from medical images using deep learning.Medical image analysis, 72:102102, 2021

    Xiangde Luo, Guotai Wang, Tao Song, Jingyang Zhang, Michael Aertsen, Jan Deprest, Sebastien Ourselin, Tom Vercauteren, and Shaoting Zhang. Mideepseg: Minimally interactive segmentation of unseen objects from medical images using deep learning.Medical image analysis, 72:102102, 2021

    work page 2021

  7. [6]

    Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨ adle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos.arXiv preprint arXiv:2408.00714, 2024

    arXiv 2024

  8. [7]

    Medical sam 2: Segment medical images as video via segment anything model 2.arXiv preprint arXiv:2408.00874, 2024

    Jiayuan Zhu, Yunli Qi, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2.arXiv preprint arXiv:2408.00874, 2024. 10 Y. Gao et al

    arXiv 2024

Show all 28 references
  1. [9]

    Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation.arXiv preprint arXiv:2408.08870, 2024

    Xinyu Xiong, Zihuang Wu, Shuangyi Tan, Wenxue Li, Feilong Tang, Ying Chen, Siying Li, Jie Ma, and Guanbin Li. Sam2-unet: Segment anything 2 makes strong encoder for natural and medical image segmentation.arXiv preprint arXiv:2408.08870, 2024

    2024

  2. [10]

    Unleashing the potential of sam2 for biomedical images and videos: A survey.arXiv preprint arXiv:2408.12889, 2024

    Yichi Zhang and Zhenrong Shen. Unleashing the potential of sam2 for biomedical images and videos: A survey.arXiv preprint arXiv:2408.12889, 2024

    2024 arXiv

  3. [11]

    Adaptive interactive seg- mentation for multimodal medical imaging via selection engine.arXiv preprint arXiv:2411.19447, 2024

    Zhi Li, Kai Zhao, Yaqi Wang, and Shuai Wang. Adaptive interactive seg- mentation for multimodal medical imaging via selection engine.arXiv preprint arXiv:2411.19447, 2024

    2024 arXiv

  4. [12]

    Desam: Decoupled segment anything model for generalizable medical image segmentation

    Yifan Gao, Wei Xia, Dingdu Hu, Wenkui Wang, and Xin Gao. Desam: Decoupled segment anything model for generalizable medical image segmentation. InInter- national Conference on Medical Image Computing and Computer-Assisted Inter- vention, pages 509–519. Springer, 2024

    2024

  5. [13]

    Segment anything model for medical images?Medical Image Analysis, 92:103061, 2024

    Yuhao Huang, Xin Yang, Lian Liu, Han Zhou, Ao Chang, Xinrui Zhou, Rusi Chen, Junxuan Yu, Jiongquan Chen, Chaoyu Chen, et al. Segment anything model for medical images?Medical Image Analysis, 92:103061, 2024

    2024

  6. [14]

    Scrib- bleprompt: Fast and flexible interactive segmentation for any medical image.arXiv preprint arXiv:2312.07381, 2023

    Hallee E Wong, Marianne Rakic, John Guttag, and Adrian V Dalca. Scrib- bleprompt: Fast and flexible interactive segmentation for any medical image.arXiv preprint arXiv:2312.07381, 2023

    2023 arXiv

  7. [15]

    Adaptive mixtures of local experts.Neural computation, 3(1):79–87, 1991

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts.Neural computation, 3(1):79–87, 1991

    1991

  8. [16]

    Sa-med2d-20m dataset: Segment anything in 2d medical imaging with 20 million masks.arXiv preprint arXiv:2311.11969, 2023

    Jin Ye, Junlong Cheng, Jianpin Chen, Zhongying Deng, Tianbin Li, Haoyu Wang, Yanzhou Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sa-med2d-20m dataset: Segment anything in 2d medical imaging with 20 million masks.arXiv preprint arXiv:2311.11969, 2023

    2023 arXiv

  9. [17]

    A global benchmark of algorithms for segmenting the left atrium from late gadolinium- enhanced cardiac magnetic resonance imaging.Medical image analysis, 67:101832, 2021

    Zhaohan Xiong, Qing Xia, Zhiqiang Hu, Ning Huang, Cheng Bian, Yefeng Zheng, Sulaiman Vesal, Nishant Ravikumar, Andreas Maier, Xin Yang, et al. A global benchmark of algorithms for segmenting the left atrium from late gadolinium- enhanced cardiac magnetic resonance imaging.Medi...

    2021

  10. [18]

    Unleashing the strengths of un- labeled data in pan-cancer abdominal organ quantification: the flare22 challenge

    Jun Ma, Yao Zhang, Song Gu, Cheng Ge, Shihao Ma, Adamo Young, Cheng Zhu, Kangkang Meng, Xin Yang, Ziyan Huang, et al. Unleashing the strengths of un- labeled data in pan-cancer abdominal organ quantification: the flare22 challenge. arXiv preprint arXiv:2308.05862, 2023

    2023 arXiv

  11. [19]

    The medical segmentation decathlon.Nature communications, 13(1):4128, 2022

    Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ron- neberger, Ronald M Summers, et al. The medical segmentation decathlon.Nature communications, 13(1):4128, 2022

    2022

  12. [20]

    Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge

    Bennett Landman, Zhoubing Xu, Juan Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge. InProc. MICCAI multi-atlas labeling beyond cranial vault—workshop challenge, volume 5, page 12. Munich, Ger...

    2015

  13. [21]

    Benchmark for algorithms segmenting the left atrium from 3d ct and mri datasets.IEEE transactions on medical imaging, 34(7):1460–1473, 2015

    Catalina Tobon-Gomez, Arjan J Geers, Jochen Peters, J¨ urgen Weese, Karen Pinto, Rashed Karim, Mohammed Ammar, Abdelaziz Daoudi, Jan Margeta, Zulma San- doval, et al. Benchmark for algorithms segmenting the left atrium from 3d ct and mri datasets.IEEE transactions on medical i...

    2015

  14. [22]

    Evaluation of prostate segmentation algorithms for mri: the promise12 challenge.Medical image analysis, 18(2):359–373, 2014

    Geert Litjens, Robert Toth, Wendy Van De Ven, Caroline Hoeks, Sjoerd Kerkstra, Bram Van Ginneken, Graham Vincent, Gwenael Guillard, Neil Birbeck, Jindang Zhang, et al. Evaluation of prostate segmentation algorithms for mri: the promise12 challenge.Medical image analysis, 18(2)...

    2014

  15. [23]

    Computer aided detection for pulmonary embolism challenge (cad-pe).arXiv preprint arXiv:2003.13440, 2020

    Germ´ an Gonz´ alez, Daniel Jimenez-Carretero, Sara Rodr ´ ıguez-L´ opez, Carlos Cano- Espinosa, Miguel Cazorla, Tanya Agarwal, Vinit Agarwal, Nima Tajbakhsh, Michael B Gotway, Jianming Liang, et al. Computer aided detection for pulmonary embolism challenge (cad-pe).arXiv prep...

    2003 arXiv

  16. [24]

    Fuping Wu and Xiahai Zhuang. Minimizing estimated risks on unlabeled data: A new formulation for semi-supervised medical image segmentation.IEEE Transac- tions on Pattern Analysis and Machine Intelligence, 45(5):6021–6036, 2022

    2022

  17. [25]

    Structseg2019 gtv segmentation, 2023

    Jun Shi. Structseg2019 gtv segmentation, 2023

    2023

  18. [26]

    Emre Kavur, N

    A. Emre Kavur, N. Sinem Gezer, Mustafa Barı¸ s, and Sinem Aslan. CHAOS Chal- lenge - combined (CT-MR) healthy abdominal organ segmentation.Medical Image Analysis, 69:101950, April 2021

    2021

  19. [27]

    Crossmoda 2021 challenge: Benchmark of cross-modality domain adap- tation techniques for vestibular schwannoma and cochlea segmentation.Medical Image Analysis, 83:102628, 2023

    Reuben Dorent, Aaron Kujawa, Marina Ivory, Spyridon Bakas, Nicola Rieke, Samuel Joutard, Ben Glocker, Jorge Cardoso, Marc Modat, Kayhan Batmanghe- lich, et al. Crossmoda 2021 challenge: Benchmark of cross-modality domain adap- tation techniques for vestibular schwannoma and co...

    2021

  20. [28]

    Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the vessel12 study.Medical image anal- ysis, 18(7):1217–1232, 2014

    Rina D Rudyanto, Sjoerd Kerkstra, Eva M Van Rikxoort, Catalin Fetita, Pierre- Yves Brillet, Christophe Lefevre, Wenzhe Xue, Xiangjun Zhu, Jianming Liang, Ilkay ¨Oks¨ uz, et al. Comparing algorithms for automated vessel segmentation in computed tomography scans of the lung: the...

    2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.