Pith. sign in

REVIEW 4 major objections 4 minor 29 references

FunduSAM: A Specialized Deep Learning Model for Enhanced Optic Disc and Cup Segmentation in Fundus Images

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

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

Pith's one-line read FunduSAM, a SAM-based model with adapters, CBAM, and polar transformation, reports the highest optic disc and cup segmentation scores among six methods on the REFUGE dataset.

desk verdict A plausible SAM-adaptation configuration for optic disc/cup segmentation, undermined by an under-documented and partially mislabeled baseline comparison. read the letter →

arxiv 2502.06220 v1 pith:WL3TDF4C submitted 2025-02-10 cs.CV cs.IR

Jinchen Yu , Yongwei Nie , Fei Qi , Wenxiong Liao , Hongmin Cai This is my paper · ORCID
classification cs.CVcs.IR
keywords opticdiscsegmentationcupfundusimagesSegmentAnythingModelparameter-efficientfine-tuningCBAMpolartransformationREFUGEdataset
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

FunduSAM is a deep network built on the Segment Anything Model (SAM) and aimed specifically at segmenting the optic disc (OD) and optic cup (OC) in eye-fundus images, the two structures whose ratio is used to screen for glaucoma. The paper claims that by freezing SAM's image encoder, inserting adapter layers for parameter-efficient fine-tuning, adding a convolutional block attention module, converting images to polar coordinates, and using a loss that enforces the cup to lie inside the disc, FunduSAM reaches the best segmentation scores among six compared methods on the REFUGE dataset. On a random 960/240 split of the 1,200-image dataset, it reports optic-disc Dice of 0.961 and intersection-over-union (IoU) of 0.882, and optic-cup Dice of 0.867 and IoU of 0.789, above five baselines. The motivation is clinical: automated, accurate OD and OC masks make cup-to-disc ratio measurement faster and less subjective.

What carries the argument

The machinery is a set of four modifications to SAM's image encoder and training objective. (1) Parameter-efficient adapters: each of the sixteen transformer blocks gets two adapter layers, one after multi-head attention and one after the feed-forward layer, so only a small fraction of parameters is trained while the pretrained encoder weights stay frozen. (2) Disassembled CBAM: a spatial attention module is placed before the image encoder and a channel attention module after it, to sharpen responses to low-contrast, blurred boundaries. (3) Polar transformation: the cropped optic-disc region is converted from Cartesian to polar coordinates before input, and the predicted mask is converted back afterward, which balances the small cup region and straightens the elliptical boundaries. (4) A joint loss $L = \omega_1 L_{\mathrm{disk}} + \omega_2 L_{\mathrm{cup}} + \omega_3 L_{\mathrm{contain}}$ with a containment term $L_{\mathrm{contain}} = \sum_i y_i(1-x_i)$ that penalizes predicted cup pixels falling outside the predicted disc, encoding the anatomical prior that the cup lies inside the disc.

What would settle it

Re-run ResUNet, nnUNet, TransUNet, Swin-UNetr, and MedSAM under the paper's exact protocol, including the same 960/240 split, optic-disc crop and polar transform, randomly sampled point prompt, and 150-epoch schedule, then compare Dice and IoU with Table I. If any baseline reaches or exceeds FunduSAM's OD Dice 0.961 or OC Dice 0.867, the central superiority claim is refuted; if the published baseline numbers come from the official REFUGE leaderboard rather than from this protocol, the comparison is not controlled.

Watch

Extended reading notes

Core claim

On the paper's terms, the central discovery is that a largely frozen SAM can be turned into a strong specialized fundus segmenter without full fine-tuning. The authors report that FunduSAM achieves the highest Dice and IoU of all six methods for both the optic disc and the optic cup on the REFUGE dataset, with the largest gains coming from the adapter layers, followed by the polar transformation and then the CBAM attention module. The method also keeps SAM's prompt encoder and mask decoder unchanged, so the interactive-prompt behaviour of SAM is preserved while the image encoder is adapted to the fundus domain. In their ablation, removing the adapters lowers optic-disc Dice from 0.961 to 0.893, removing the polar transform lowers it to 0.906, and removing CBAM lowers it to 0.957; each removal also hurts optic-cup scores.

Load-bearing premise

The load-bearing assumption is that the five compared methods were trained and evaluated under the same random 960/240 split, polar-cropping preprocessing, and simulated prompt protocol as FunduSAM, so the reported Dice and IoU gaps reflect the method itself rather than differing experimental conditions.

Editorial extensions

If this is right

  • On the REFUGE dataset under the paper's 960/240 split, FunduSAM reports the top Dice and IoU among the six methods for both structures, with OD Dice 0.961 and OC Dice 0.867.
  • The adapter layers are the most important component in the reported ablation: removing them costs 0.068 in optic-disc Dice and 0.054 in optic-cup Dice, more than removing the polar transform or CBAM.
  • The combination of the three modules lifts plain SAM from OD Dice 0.829 and OC Dice 0.721 to 0.961 and 0.867, respectively.
  • Keeping SAM's prompt encoder and mask decoder fixed means the interactive prompting behavior of SAM remains available while the network specializes, which could ease clinical annotation workflows.

Reading between the lines

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

  • Because the paper reports a custom random 960/240 split rather than the official REFUGE train/validation/test partition, the stated scores are direct evidence only for that split; reproducing the official protocol would be needed to compare with published leaderboard results.
  • The recipe of freezing a vision foundation model, inserting adapters, transforming to polar coordinates, and enforcing a containment loss is likely transferable to other nested anatomical structures that are roughly circular in cross-section, such as the macula or the optic nerve head.
  • The ablation's small CBAM gain suggests a simpler model without CBAM might capture most of the benefit at lower complexity; the paper does not test that configuration.
  • A controlled rerun that trains all baselines with the same simulated prompts and preprocessing would separate the value of the architecture from the value of the experimental protocol.
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 / 4 minor

Summary. The paper introduces FunduSAM, a SAM-based network for optic disc (OD) and optic cup (OC) segmentation in fundus images. The proposed method adds adapter layers to each ViT block of SAM's image encoder for parameter-efficient fine-tuning, inserts a disassembled CBAM module at the encoder input/output, uses polar transformation as preprocessing, and trains with a joint loss combining OD cross-entropy, OC cross-entropy, and a structural containment term. Experiments on the REFUGE dataset (random 960/240 split) report OD Dice 0.961, OD IoU 0.882, OC Dice 0.867, OC IoU 0.789, which the paper claims to be higher than five compared methods (ResUNet, nnUNet, TransUNet, Swin-UNetr, and 'MedSAM'). An ablation study assesses the contributions of CBAM, adapter, and polar transformation.

Significance. If the claims are substantiated, FunduSAM would be a credible application of parameter-efficient SAM adaptation to fundus imaging, with a plausible combination of established components (adapters, CBAM, polar transform) and a structural prior between OD and OC. The paper's strengths include a clear task motivation, a straightforward ablation design, and the use of a widely benchmarked dataset. However, the reported superiority is not yet established because the baseline evaluation protocol is underspecified, the citation of the 'MedSAM' baseline is inconsistent, and the structural loss term is written in a way that makes it independent of the network output. The paper also provides no error bars, no statistical tests, and no code, which limits reproducibility and the strength of the comparative claims.

major comments (4)
  1. [III.D, Eq. (6)] The structural containment loss Lcontain = Σ_i (y_i · (1 - x_i)) is defined using only the ground-truth binary labels x_i and y_i, with no dependence on the predictions ˆx_i, ˆy_i. As written, this term is constant for a given training set and cannot influence the network parameters or 'introduce a structural constraint between the OD and OC.' This is a load-bearing error: the joint loss described in Eq. (3) and the claimed benefit of structural priors in the ablation study are not supported by the given formula. Please clarify the actual loss used, e.g., by replacing one of the labels with the predicted probability, and re-run the experiments with the corrected loss.
  2. [IV.A and Table I] The superiority claim in Table I is only meaningful if all five baselines were trained and evaluated under exactly the same protocol as FunduSAM: the same 960/240 random split, the same polar-transform preprocessing, and the same prompt simulation. Section IV.B reports implementation details only for FunduSAM; no training hyperparameters, preprocessing, prompt rule, or evaluation settings are given for ResUNet, nnUNet, TransUNet, Swin-UNetr, or the method labeled MedSAM. Since the reported margins are small (e.g., 0.84% in OD Dice and 0.34% in OD IoU over Swin-UNetr), even modest protocol differences could change the ranking. The paper must specify the baseline training/evaluation setup, provide per-run variability (e.g., standard deviations over multiple seeds or runs), and ideally use the official REFUGE split for comparability with published numbers.
  3. [Table I and References] The baseline labeled 'MedSAM' is cited as reference [29], which is Yu et al., 'Robust optic disc and cup segmentation with deep learning for glaucoma detection' (Computerized Medical Imaging and Graphics, 2019), not the MedSAM model of Ma et al. (Nature Communications, 2024, reference [8] in the paper). This citation mismatch strongly suggests that the reported MedSAM row was taken from an external publication rather than re-run under this paper's protocol. If so, the comparison is not like-for-like and the 'highest results for both Dice and Iou' claim is not supported. The authors should correct the citation, re-run the true MedSAM under their protocol, or otherwise justify the provenance of the baseline numbers.
  4. [III.D and IV.D] The loss weights ω1, ω2, ω3 in Eq. (3) are not reported. The paper only states that they 'update within the range of [0,1]', which is ambiguous: are they learned during training or manually set? Since the joint loss is central to the method and the ablation study attributes performance gains to the structural constraint, the exact values (or selection procedure) are necessary for reproducibility and for interpreting the ablation results. Please report the values and how they were chosen.
minor comments (4)
  1. [III.C, Eqs. (1)-(2)] The text says Eq. (1) is the conversion from Cartesian to polar coordinates, but Eq. (1) (x = r cos θ, y = r sin θ) is the polar-to-Cartesian mapping; Eq. (2) is the Cartesian-to-polar mapping. Please fix the description to avoid confusion.
  2. [IV.B] The sentence 'The loss function is the joint loss of OD and OC cross entropy and structural constraints a priori' contains a grammatical error and does not specify the loss weights or the exact form of the structural constraint; please revise for clarity.
  3. [Table I] It would be helpful to report the number of independent runs and standard deviations for each metric, as the current table gives no indication of result stability.
  4. [IV.E] The visualization section claims superiority for low-contrast images based on qualitative inspection; consider providing a quantitative low-contrast subset analysis to support the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FunduSAM's contribution is an empirical segmentation method compared against external baselines, with no derivation step that reduces to its own inputs.

full rationale

The paper does not contain a theoretical derivation chain whose outputs could be equivalent to its inputs by construction; it is an empirical deep-learning paper. The central claim, that FunduSAM achieves the highest Dice and IoU on the REFUGE dataset, is supported by Table I, which compares FunduSAM against five external methods, and by an ablation study in Table II that varies the proposed components. No parameter is fitted to the test set and then renamed as a prediction; the loss weights are stated only loosely but are not claimed as fitted predictions. There are no load-bearing self-citations, and no uniqueness theorem or prior-work-derived ansatz is invoked to force the method. The reviewer-visible concerns are real but are not circularity: the manuscript does not document the training protocol for the baseline methods, and the baseline labeled 'MedSAM' cites reference [29], which is Yu et al. 2019 on optic disc and cup segmentation rather than the MedSAM paper (Ma et al., Nature Communications 2024). These are reproducibility and comparison-fairness issues, not cases where a result is equivalent to its inputs by definition. The evaluation is self-contained in the sense that the proposed model's performance is independently measured against external published methods, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper does not derive the reported performance; it is an empirical engineering claim. It depends on standard polar-coordinate math, anatomical priors, and several unreported design choices (loss weights, adapter dimensions, crop parameters). No new physical or conceptual entities are introduced.

free parameters (3)
  • Loss weights omega_1, omega_2, omega_3 = Not reported
    In Eq. (3), the joint loss is a weighted sum; the weights are said to update within [0,1] but their values or learning scheme are not given, and they directly affect the reported Dice/IoU.
  • Adapter bottleneck dimension = Not reported
    The adapter layers in each ViT block (down projection, activation, up projection) require a hidden dimension; it is not specified and controls the PEFT capacity and final segmentation quality.
  • Polar transformation crop parameters = Not reported
    The OD crop center and radius used before polar mapping are taken from 'the method from Wang et al.' but not specified; segmentation results depend on this preprocessing.
assumptions (5)
  • standard math Polar coordinate transform (Eqs. 1-2) is invertible and does not distort the OD/OC structure relevant to segmentation.
    Invoked in Section III.C; standard but assumes a known disc center and an ROI crop that preserves all relevant anatomy.
  • domain assumption The optic cup is always contained within the optic disc in every annotation, so L_contain = sum_i y_i(1-x_i) is a valid prior.
    Used in Eq. (6), Section III.D; if any ground-truth mask violates containment, this loss penalizes a correct segmentation.
  • domain assumption A single random foreground point is an adequate prompt for SAM's mask decoder at training and test time.
    Section IV.B states 'we randomly sample a point in the foreground region of the label to simulate the prompt process'; no prompt sensitivity analysis is given.
  • domain assumption The random 960/240 split of all 1,200 images is a representative evaluation protocol for comparing segmentation methods.
    Section IV.A; this assumes the official cross-camera/domain-shift test is not needed and that single-run results transfer.
  • domain assumption SAM's frozen image encoder, with adapters and CBAM, can capture fundus-specific features such as blurred boundaries and low contrast.
    Core design premise in Sections III.A-III.B; supported only by the single REFUGE experiment, not by independent benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FunduSAM: A Specialized Deep Learning Model for Enhanced Optic Disc and Cup Segmentation in Fundus Images." pith.science (2026). https://pith.science/paper/WL3TDF4C

@misc{pith2026250206220,
  author       = {Pith},
  title        = {Pith review of: FunduSAM: A Specialized Deep Learning Model for Enhanced Optic Disc and Cup Segmentation in Fundus Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WL3TDF4C}},
  note         = {Machine review of arXiv:2502.06220}
}
read the original abstract

The Segment Anything Model (SAM) has gained popularity as a versatile image segmentation method, thanks to its strong generalization capabilities across various domains. However, when applied to optic disc (OD) and optic cup (OC) segmentation tasks, SAM encounters challenges due to the complex structures, low contrast, and blurred boundaries typical of fundus images, leading to suboptimal performance. To overcome these challenges, we introduce a novel model, FunduSAM, which incorporates several Adapters into SAM to create a deep network specifically designed for OD and OC segmentation. The FunduSAM utilizes Adapter into each transformer block after encoder for parameter fine-tuning (PEFT). It enhances SAM's feature extraction capabilities by designing a Convolutional Block Attention Module (CBAM), addressing issues related to blurred boundaries and low contrast. Given the unique requirements of OD and OC segmentation, polar transformation is used to convert the original fundus OD images into a format better suited for training and evaluating FunduSAM. A joint loss is used to achieve structure preservation between the OD and OC, while accurate segmentation. Extensive experiments on the REFUGE dataset, comprising 1,200 fundus images, demonstrate the superior performance of FunduSAM compared to five mainstream approaches.

Figures

Figures reproduced from arXiv: 2502.06220 by the authors.

Figure 1
Figure 1. The pipeline of our method for optic disc and cup segmentation, the input fundus image is first transformed in polar coordinates, followed by learning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed FunduSAM consists of Adapter layers, Spatial Attention Module and Channel Attention Module in CBAM [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Visualization of polar transformation, which simplifies the boundary [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization of the results of different methods for fundus optic disc and cup segmentation [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 21 canonical work pages

  1. [29]

    Robust optic disc and cup segmentation with deep learning for glaucoma detection

    Shuang Yu, Di Xiao, Shaun Frost, and Yogesan Kanagasingam. Robust optic disc and cup segmentation with deep learning for glaucoma detection. Computerized Medical Imaging and Graphics , 74:61–71, 2019

    work page 2019

  2. [8]

    Seg- ment anything in medical images

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Seg- ment anything in medical images. Nature Communications , 15(1):654, 2024

    work page 2024

  3. [1]

    Ranking of optic disc variables for detection of glaucomatous optic nerve damage

    Jost B Jonas, Antonio Bergua, Paul Schmitz-Valckenberg, Konstantinos I Papastathopoulos, and Wido M Budde. Ranking of optic disc variables for detection of glaucomatous optic nerve damage. Investigative Oph- thalmology & Visual Science , 41(7):1764–1773, 2000

    work page 2000

  4. [2]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 3431– 3440, 2015

    work page 2015

  5. [3]

    U-net: Con- volutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Con- volutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 , pages 234–241. Springer, 2015

    2015

  6. [4]

    Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Mur- phy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence , 40(4):834–848, 2017

    work page 2017

  7. [5]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017

    2017

  8. [6]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4015– 4026, 2023

    work page 2023

Show all 29 references
  1. [7]

    Customized segment anything model for medical image segmentation, 2023

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

    2023

  2. [9]

    How to efficiently adapt large segmentation model (sam) to medical images

    Xinrong Hu, Xiaowei Xu, and Yiyu Shi. How to efficiently adapt large segmentation model (sam) to medical images. arXiv preprint arXiv:2306.13731, 2023

    2023 arXiv

  3. [10]

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

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

    2023

  4. [11]

    Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation

    Xian Lin, Yangyang Xiang, Li Zhang, Xin Yang, Zengqiang Yan, and Li Yu. Samus: Adapting segment anything model for clinically-friendly and generalizable ultrasound image segmentation. arXiv preprint arXiv:2309.06824, 2023

    2023 arXiv

  5. [12]

    Parameter-efficient transfer learning for nlp

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Syl- vain Gelly. Parameter-efficient transfer learning for nlp. In International conference on machine learning , pages 2790–2799. PMLR, 2019

    2019

  6. [13]

    Parameter-efficient transfer learning with diff pruning, 2020

    Demi Guo, Alexander M Rush, and Yoon Kim. Parameter-efficient transfer learning with diff pruning, 2020

    2020

  7. [14]

    The power of scale for parameter-efficient prompt tuning

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. 2021

    2021

  8. [15]

    Prefix-tuning: Optimizing continuous prompts for generation

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. 2021

    2021

  9. [16]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685 , 2021

    2021 arXiv

  10. [17]

    Parameter-efficient model adaptation for vision transformers

    Xuehai He, Chunyuan Li, Pengchuan Zhang, Jianwei Yang, and Xin Eric Wang. Parameter-efficient model adaptation for vision transformers. 37(1):817–825, 2023

    2023

  11. [18]

    Adaptformer: Adapting vision transformers for scalable visual recognition, 2022

    Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition, 2022

    2022

  12. [19]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV) , pages 3–19, 2018

    2018

  13. [20]

    A cbam based multiscale transformer fusion approach for remote sensing image change detection

    Wei Wang, Xinai Tan, Peng Zhang, and Xin Wang. A cbam based multiscale transformer fusion approach for remote sensing image change detection. IEEE Journal of Selected Topics in Applied Earth Observa- tions and Remote Sensing , 15:6817–6825, 2022

    2022

  14. [21]

    Joint optic disc and cup segmentation based on multi-label deep network and polar transformation

    Huazhu Fu, Jun Cheng, Yanwu Xu, Damon Wing Kee Wong, Jiang Liu, and Xiaochun Cao. Joint optic disc and cup segmentation based on multi-label deep network and polar transformation. IEEE transactions on medical imaging , 37(7):1597–1605, 2018

    2018

  15. [22]

    Efficient optic cup detection from intra-image learning with retinal structure priors

    Yanwu Xu, Jiang Liu, Stephen Lin, Dong Xu, Carol Y Cheung, Tin Aung, and Tien Yin Wong. Efficient optic cup detection from intra-image learning with retinal structure priors. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2012: 15th International Conferen...

    2012

  16. [23]

    Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening, 2022

    Huihui Fang, Fei Li, Junde Wu, Huazhu Fu, Xu Sun, Jaemin Son, Shuang Yu, Menglu Zhang, Chenglang Yuan, Cheng Bian, et al. Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening, 2022

    2022

  17. [24]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation, 2016

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation, 2016

    2016

  18. [25]

    Road extraction by deep residual u-net

    Zhengxin Zhang, Qingjie Liu, and Yunhong Wang. Road extraction by deep residual u-net. IEEE Geoscience and Remote Sensing Letters , 15(5):749–753, 2018

    2018

  19. [26]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation

    Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods , 18(2):203–211, 2021

    2021

  20. [27]

    Transunet: Transformers make strong encoders for medical image segmentation, 2021

    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 medical image segmentation, 2021

    2021

  21. [28]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images, 2021

    Ali Hatamizadeh, Vishwesh Nath, Yucheng Tang, Dong Yang, Holger R Roth, and Daguang Xu. Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images, 2021

    2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.