Pith. sign in

REVIEW 4 major objections 5 minor 48 references

Adversarial Vessel-Unveiling Semi-Supervised Segmentation for Retinopathy of Prematurity Diagnosis

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read A semi-supervised teacher–student method segments retinal vessels in retinopathy of prematurity (ROP) without manual ROP annotations, and its vessel masks improve four-stage ROP classification when fused with fundus images.

desk verdict A credible SSL pipeline for ROP vessel segmentation with solid public-dataset results, but a key loss equation and softmax axis are under-specified and the ROP test set is thin. read the letter →

arxiv 2411.09140 v1 pith:N7746CZT submitted 2024-11-14 cs.CV

classification cs.CV
keywords semi-supervisedsegmentationretinalvesselretinopathyofprematurityteacher-studentlearningdomainadversarialuncertainty-weightedconsistencymulti-stageROPclassification
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

Retinopathy of prematurity (ROP) is a leading cause of blindness in preterm infants, and its diagnosis depends on reading retinal vessel structure; but preterm vessels are thin and faint, and manual annotation is prohibitively difficult. This paper argues that a semi-supervised teacher–student segmentation model can extract those vessels without any manually annotated ROP image, using only 18 labeled adult/child retinal images from public datasets plus 90 unlabeled ROP images. The two technical proposals are an uncertainty-weighted vessel-unveiling module, which amplifies faint or hidden vessels in the teacher's predictions, and a domain-adversarial discriminator that aligns feature distributions across the public and ROP domains. If the argument holds, clinicians could get vessel maps for ROP staging at no annotation cost, and the vessel maps themselves improve multi-stage classification accuracy when fused with original fundus images. The central evidence is a Dice score of 45.24% on the ROP test set, over 6 percentage points above the second-best baseline.

What carries the argument

The load-bearing mechanism is the uncertainty-weighted vessel-unveiling module inside the teacher branch. For each unlabeled ROP image, the teacher performs $K=8$ stochastic forward passes with dropout and soft augmentation, producing probability maps $p_k$; a per-pixel vessel entropy $I_{\mathrm{vessel}}$ is derived across these passes, and the softmax of $(1 - I_{\mathrm{vessel}})$ is element-wise multiplied with the averaged prediction to form the vessel-unveiling target $y_w$. The consistency loss combines an unweighted term comparing the student to the teacher's averaged prediction and a distance to $y_w$ weighted by $(1 - I_{\mathrm{vessel}})$, so the student is pushed toward faint vessels while being allowed to stay uncertain where the teacher is uncertain. A PatchGAN-style discriminator takes encoder features from labeled and unlabeled images and adversarially aligns the public and ROP domains, and the student uses three regularized decoders (main, feature-noise, feature-dropout) with supervised loss on the labeled public images.

What would settle it

Compare the vessel-unveiling target $y_w$ with the 12 manually annotated ROP test images: if $y_w$ has a lower Dice against ground truth than the teacher's plain averaged prediction, then the entropy weighting is selecting noise rather than vessels, and the consistency loss is training the student toward a worse target.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that combining adversarial feature alignment with an uncertainty-weighted consistency target enables a teacher–student network to segment retinal vessels in unlabeled ROP fundus images, despite a large domain gap from the public datasets it trains on. The method is the first to attempt ROP vessel segmentation in a semi-supervised setting without any ROP ground-truth vessel annotations during training; the only ROP annotations are 12 manually traced test images used for evaluation. The reported Dice of 45.24% on the ROP test set, exceeding the second-best baseline by more than 6%, is taken as evidence that the vessel-unveiling module and domain alignment extract vessels that other semi-supervised methods miss. The paper further shows that feeding the segmented vessel masks into a fusion classifier with the original fundus images raises four-stage ROP classification accuracy from 74.76% to 76.38%.

Load-bearing premise

The teacher's uncertainty-weighted vessel-unveiling prediction $y_w$ is adopted as a reliable learning target for unlabeled ROP images, but it is never checked against any annotated ROP vessel map before training starts; if the uncertainty weighting emphasizes false positives, the consistency loss will amplify them and the reported ROP gains will not transfer.

Editorial extensions

If this is right

  • ROP vessel segmentation can be performed without any manually annotated ROP images, removing the main annotation bottleneck for this disease.
  • The model retains source-domain performance on CHASEDB and STARE while improving ROP test Dice, showing the adversarial alignment does not sacrifice labeled-domain accuracy.
  • Vessel masks from the segmentation model carry complementary signal: fusing them with fundus images raises four-stage ROP classification accuracy from 74.76% to 76.38% and precision from 71.61% to 75.79%.
  • The same semi-supervised recipe can be applied to other retinal vessel segmentation tasks where target-domain labels are unavailable.

Reading between the lines

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

  • Editorial inference: the paper never measures overlap between the entropy-weighted target $y_w$ and true ROP vessels; computing that overlap on the 12 annotated test images would settle whether the unveiling module selects vessels or noise.
  • Editorial inference: the reported ROP gain might come from the teacher's averaged prediction rather than the entropy weighting; a control with a plain averaged teacher target would isolate the module's contribution.
  • Editorial inference: if the domain-adversarial alignment transfers public-dataset knowledge to ROP images, the same recipe can be tested on other pediatric imaging domains with scarce infant annotations.
  • Editorial inference: the 1.62-point accuracy gain from fusing vessel masks could reflect extra model capacity; a control with a second image channel carrying no vessel information would test that.
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 proposes a semi-supervised retinal vessel segmentation framework for retinopathy of prematurity (ROP), combining a teacher-student consistency model with an uncertainty-weighted vessel-unveiling module and a feature-level adversarial domain alignment discriminator. The model is trained on publicly labeled datasets (CHASEDB, STARE) together with unlabeled in-house ROP images, and is evaluated on small ROP and public test sets. The authors report consistent improvements over several semi-supervised baselines, and they extend the framework to a downstream ROP multi-stage classification task in which predicted vessel masks are fused with fundus images.

Significance. If the reported results hold, the paper addresses a real annotation bottleneck in ROP vessel segmentation and demonstrates a potentially useful downstream diagnostic application. The work is among the first to attempt vessel segmentation across all four ROP stages without ground-truth ROP vessel annotations, and the combination of adversarial feature alignment with uncertainty-weighted teacher-student learning is a reasonable direction. The paper also reports results on multiple public datasets and a downstream classification task. However, the central vessel-unveiling loss is described inconsistently in the equations, the target construction is under-specified and potentially degenerate, and the evaluation is based on very small test sets without statistical significance assessment. These issues currently prevent verification of the central claims.

major comments (4)
  1. [Section III-B, Eqs. (3)-(5)] Equation (4) is labeled MSE(ŷ_j, ŷ'_j), but its right-hand side is (1/(H×W)) Σ (ŷ_j − yw)², using yw instead of ŷ'_j. With Equation (5), both consistency terms in Equation (3) compare the student output ŷ_j to the vessel-unveiling target yw, so the teacher output ŷ'_j never actually enters the consistency loss as written. If the implementation follows Eq. (4) literally, the intended teacher-student MSE term is absent and the loss reduces to two weighted variants of the same student-to-yw comparison. The authors must either correct the equation or clarify the implementation and confirm that the reported results are produced by the described loss.
  2. [Section III-B and Fig. 3] The construction of the vessel-unveiling target yw is under-specified and potentially degenerate. The operation softmax(1 − vessel_entropy) is written without specifying the axis of the softmax. If softmax is applied over the full H×W spatial map, the weights sum to 1 and, when multiplied elementwise by the averaged prediction, yield values on the order of 1/(H×W); for the 400×400 patches used in training this makes yw nearly zero everywhere, which would collapse the consistency loss toward predicting all-background. If softmax is applied over the channel dimension, the single-channel softmax is identically 1 and the uncertainty weighting has no effect. The authors must specify the intended normalization and provide evidence, e.g., histograms of yw on unlabeled ROP images, that the target is non-degenerate.
  3. [Section IV-B, IV-E, Tables I-III] The evaluation is severely underpowered for the central claims. The ROP test set consists of only 12 manually annotated images, and the public test sets consist of 5 images each. No standard deviations, confidence intervals, or significance tests are reported for any table. The headline result that the proposed model beats the second-best method by more than 6% Dice on the ROP test set (45.24% versus 39.00%) relies on a single split of 12 images. Please report repeated-seed experiments with variance measures and, if possible, a larger annotated ROP test set or a significance test (e.g., paired bootstrap over images).
  4. [Section V-A and V-C] The downstream classification experiment uses 3,873 fundus images from 217 patients, with approximately 5 images per eye, but the paper does not state how the train/test split was performed. If images from the same patient or the same eye appear in both training and test partitions, the reported fusion improvement (76.38% versus 74.76% accuracy) will be inflated by intra-patient correlation. The authors should specify a patient-level or eye-level split and report per-patient or per-eye metrics, together with confidence intervals, before claiming clinical utility for the downstream task.
minor comments (5)
  1. [Section I, Contributions] The sentence 'In our work, we make four three contributions' contains an editing error; it should list either four or three contributions consistently.
  2. [Fig. 3] The vessel-entropy formula in Figure 3 is written as '−1.0×K ∑ipilog(pi)', which is ambiguous. It should clarify whether the average is over the K dropout passes and whether the entropy is computed over the class dimension of the probabilities.
  3. [Eq. 5] The variable Ivessel is called 'vessel entropy' but it is not defined whether higher Ivessel corresponds to higher uncertainty or higher vessel probability. Since Eq. (5) multiplies the distance by (1−Ivessel), the meaning of Ivessel is load-bearing for interpreting the weighting.
  4. [Section VI, Conclusion] The conclusion refers to a 'vessel-veiling module' where the rest of the paper uses 'vessel-unveiling module'; this typo should be corrected.
  5. [Tables I-III] The captions state 'UpperBound' is fully supervised with only the labeled public images. This is clear, but the tables would benefit from a note that no ROP ground-truth vessels are used in training, to avoid confusion with the 12 held-out ROP test annotations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported segmentation and classification results are measured against external held-out labels, not against the paper's own training targets.

full rationale

The paper's central claims are empirical: a Dice score of 45.24% on a 12-image ROP test set with manual annotations (Section IV-B, IV-E), and a downstream classification gain from 74.76% to 76.38% accuracy on a separate 3,873-image cohort (Section V). These test labels are external to the semi-supervised training signal. The vessel-unveiling target yw is generated from the teacher's MC-dropout predictions and entropy and is used only as a consistency regularizer; it is not the quantity reported as a result, so the evaluation does not reduce to the training target by construction. The adversarial domain alignment and the multi-decoder student are standard architectural components, and the paper's citations (MC dropout [16], PatchGAN [23], U-Net [19], Mean Teacher [22]) are independent external sources rather than load-bearing self-citations. No uniqueness theorem or author-derived premise is invoked to force the method choice. The main caveats are correctness and reproducibility issues rather than circularity: Eq. 4 labels a term MSE(ˆyj, ˆy'_j) but writes it with yw instead of ˆy'_j, contradicting Eq. 3; and Fig. 3 does not specify the axis for softmax, so the vessel-unveiling weights are underspecified. These internal inconsistencies should be checked against the code, but they do not make the reported external test-set numbers equivalent to the model's inputs by construction. The paper also does not state how hyperparameters were selected, which is a reporting gap, not evidence that the central results were fitted to the test labels.

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

The free parameters are the hyperparameters of the training procedure that are not derived from data: the consistency-loss weight alpha, noise scale sigma, attention-dropout threshold t, MC dropout count K, and EMA schedule. The key axioms are that transferable vessel morphology exists between public adult/adolescent fundus images and preterm ROP images, that the teacher's uncertainty-weighted predictions are reliable training targets, and that the 12 held-out manual annotations are representative. No new entities are invented.

free parameters (5)
  • consistency loss weight alpha (Eq. 3) = not reported
    The total loss L = Lsup + alpha*Lcons + Ladv, but no value or tuning procedure for alpha is given in Section IV-A.
  • feature noise scale sigma = not reported
    Used for z_noise = z + epsilon with epsilon ~ U(-sigma, sigma) in Section III-A; the scale is never specified.
  • attention dropout threshold t = not reported
    Used to build the mask M(z)=I(A(z)<t) in Section III-A; no procedure for choosing t is described.
  • MC dropout passes K = 8
    K=8 is set in Section IV-A, but no sensitivity analysis is provided.
  • EMA decay schedule = min(1 - 1/(epoch+1), 0.95)
    The schedule is fixed in Section IV-A; no analysis of its effect on the results is given.
assumptions (3)
  • domain assumption Public labeled datasets (CHASEDB, STARE) and unlabeled ROP images share vessel morphology and feature structure sufficient for adversarial alignment to transfer segmentation knowledge.
    Invoked in Section III-C and Section I; if false, the domain-adversarial alignment cannot improve ROP segmentation.
  • domain assumption The teacher's uncertainty-weighted prediction (vessel-unveiling target) approximates true vessel structure well enough to serve as a training target for unlabeled images.
    Core of Section III-B; no validation of the unveiling target against annotations before training is reported.
  • domain assumption The 12 manually annotated ROP images (held out for evaluation) are representative of the unlabeled ROP population and were not used for hyperparameter selection.
    Section IV-B states these annotations are used only for quantitative evaluation, but no split or selection protocol is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adversarial Vessel-Unveiling Semi-Supervised Segmentation for Retinopathy of Prematurity Diagnosis." pith.science (2026). https://pith.science/paper/N7746CZT

@misc{pith2026241109140,
  author       = {Pith},
  title        = {Pith review of: Adversarial Vessel-Unveiling Semi-Supervised Segmentation for Retinopathy of Prematurity Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7746CZT}},
  note         = {Machine review of arXiv:2411.09140}
}
read the original abstract

Accurate segmentation of retinal images plays a crucial role in aiding ophthalmologists in diagnosing retinopathy of prematurity (ROP) and assessing its severity. However, due to their underdeveloped, thinner vessels, manual annotation in infant fundus images is very complex, and this presents challenges for fully supervised learning. To address the scarcity of annotations, we propose a semi supervised segmentation framework designed to advance ROP studies without the need for extensive manual vessel annotation. Unlike previous methods that rely solely on limited labeled data, our approach leverages teacher student learning by integrating two powerful components: an uncertainty weighted vessel unveiling module and domain adversarial learning. The vessel unveiling module helps the model effectively reveal obscured and hard to detect vessel structures, while adversarial training aligns feature representations across different domains, ensuring robust and generalizable vessel segmentations. We validate our approach on public datasets (CHASEDB, STARE) and an in-house ROP dataset, demonstrating its superior performance across multiple evaluation metrics. Additionally, we extend the model's utility to a downstream task of ROP multi-stage classification, where vessel masks extracted by our segmentation model improve diagnostic accuracy. The promising results in classification underscore the model's potential for clinical application, particularly in early-stage ROP diagnosis and intervention. Overall, our work offers a scalable solution for leveraging unlabeled data in pediatric ophthalmology, opening new avenues for biomarker discovery and clinical research.

Figures

Figures reproduced from arXiv: 2411.09140 by the authors.

Figure 1
Figure 1. The domain difference among publicly available datasets (CHASEDB [3] & STARE [24]) and our ROP dataset. CHASEDB primar￾ily consists of images from adolescents and healthy populations, while STARE comprises images from adults and emphasizes heterogeneous pathology cases. In contrast, our ROP dataset is curated for infants’ retinopathy of prematurity detection. We introduce a semi-supervised learning (SSL) approach to… view at source ↗
Figure 2
Figure 2. The framework of our proposed semi-supervised learning method for vessel segmentation. The student network learns from the supervised and consistency loss. The vessel-unveiling prediction extracted from the uncertainty-aware vessel-unveiling module reveals the hidden vessels to ensure boosted student network’s performance on unlabeled inputs. The discriminator takes latent representation of both labeled and unlabele… view at source ↗
Figure 3
Figure 3. Uncertainty-aware vessel-unveiling Module. The vessel￾unveiling reveals hidden vessels due to ROP image domain internal challenges discussed in section I. In the heatmap scale, the red color represents a higher value. The final consistency loss is defined as: Lcons = α × MSE(ˆyj , yˆ ′ j ) + DIST(ˆyj , yw) (3) where α is a weighting factor used to balance the two components of the loss. In this equation, the student… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization result of all baselines and our model on ROP test set with 18 CHASEDB labeled data and 90 ROP unlabeled images during training. The ROP vessels by their underdeveloped structure and background tissue colors, are harder to recognize than public datasets. O…
Figure 5
Figure 5. Figure 5: Visualization result of different methods on CHASEDB test set with 5 labeled data during training. To illustrate the discrepancy between other baselines, we set bounding boxes in the color red to share false positive regions. TABLE III QUANTITATIVE COMPARISON OF EXPERI…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 47 canonical work pages

  1. [1]

    V . M. Yildiz, P. Tian, I. Yildiz, J. M. Brown, J. Kalpathy-Cramer, J. Dy, S. Ioannidis, D. Erdogmus, S. Ostmo, S. J. Kim, and R. P. Chan, ”Plus disease in retinopathy of prematurity: convolutional neural network performance using a combined neural network and feature extraction approach,” in Translational Vision Science & Technology, vol. 9, no. 2, pp. 1...

  2. [2]

    J. P. Campbell, E. Ataer-Cansizoglu, V . Bolon-Canedo, A. Bozkurt, D. Erdogmus, J. Kalpathy-Cramer, S. N. Patel, J. D. Reynolds, J. Horowitz, K. Hutcheson, and M. Shapiro, ”Expert diagnosis of plus disease in retinopathy of prematurity from computer-based image analysis,” in JAMA Ophthalmology, vol. 134, no. 6, pp. 651-657, 2016

  3. [3]

    M. M. Fraz, P. Remagnino, A. Hoppe, B. Uyyanonvara, A. R. Rudnicka, C. G. Owen, and S. A. Barman, ”An ensemble classification-based approach applied to retinal blood vessel segmentation,” in IEEE TBME, vol. 59, no. 9, pp. 2538-2548, 2012

  4. [4]

    Q. Hu, M. D. Abr `amoff, and M. K. Garvin, ”Automated separation of binary overlapping trees in low-contrast color retinal images,” in MICCAI, vol. 16, pp. 436-443, Springer Berlin Heidelberg, 2013

  5. [5]

    K. Jin, X. Huang, J. Zhou, Y . Li, Y . Yan, Y . Sun, Q. Zhang, Y . Wang, and J. Ye, ”Fives: A fundus image dataset for artificial Intelligence-based vessel segmentation,” in Scientific Data, vol. 9, no. 1, p. 475, 2022

  6. [6]

    X. Lyu, L. Cheng, and S. Zhang, ”The reta benchmark for retinal vascular tree analysis,” in Scientific Data, vol. 9, no. 1, p. 397, 2022

  7. [7]

    M. U. Akram, S. Akbar, T. Hassan, S. G. Khawaja, U. Yasin, and I. Basit, ”Data on fundus images for vessels segmentation, detection of hypertensive retinopathy, diabetic retinopathy and papilledema,” in Data in Brief, vol. 29, p. 105282, 2020

  8. [8]

    Staal, M

    J. Staal, M. D. Abr `amoff, M. Niemeijer, M. A. Viergever, and B. Van Ginneken, ”Ridge-based vessel segmentation in color images of the retina,” in IEEE TMI, vol. 23, no. 4, pp. 501-509, 2004

Show all 48 references
  1. [9]

    J. C. Wigdahl, C. Agurto, S. C. Nemeth, V . S. Joshi, W. Bauman, P. Soliz, and E. S. Barriga, ”Detection of plus disease in retinopathy of prematurity using automatic vessel tortuosity measurements,” in Investigative Ophthalmology & Visual Science , vol. 58, no. 8, pp. 654- 654, 2017

  2. [10]

    X. Chen, Y . Yuan, G. Zeng, and J. Wang, ”Semi-supervised semantic segmentation with cross pseudo supervision,” in CVPR, pp. 2613-2622, 2021

  3. [11]

    J. Hou, X. Ding, and J. D. Deng, ”Semi-supervised semantic segmenta- tion of vessel images using leaking perturbations,” in WACV, pp. 2625- 2634, 2022

  4. [12]

    X. Luo, J. Chen, T. Song, and G. Wang, ”Semi-supervised medical image segmentation through dual-task consistency,” in AAAI, vol. 35, no. 10, pp. 8801-8809, May 2021

  5. [13]

    Lahiri, V

    A. Lahiri, V . Jain, A. Mondal, and P. K. Biswas, ”Retinal vessel seg- mentation under extreme low annotation: A GAN based semi-supervised approach,” in ICIP, pp. 418-422, October 2020

  6. [14]

    Y . Wu, Z. Ge, D. Zhang, M. Xu, L. Zhang, Y . Xia, and J. Cai, ”Mutual consistency learning for semi-supervised medical image segmentation,” in Medical Image Analysis , vol. 81, p. 102530, 2022

  7. [15]

    Zhang, L

    Y . Zhang, L. Yang, J. Chen, M. Fredericksen, D. P. Hughes, and D. Z. Chen, ”Deep adversarial networks for biomedical image segmentation utilizing unannotated images,” in MICCAI, Part III, vol. 20, pp. 408-416, Springer International Publishing, 2017

  8. [16]

    Gal and Z

    Y . Gal and Z. Ghahramani, ”Dropout as a Bayesian approximation: Representing model uncertainty in deep learning,” in ICML, pp. 1050- 1059, PMLR, June 2016

  9. [17]

    D. Zhai, B. Hu, X. Gong, H. Zou, and J. Luo, ”ASS-GAN: Asymmetric semi-supervised GAN for breast ultrasound image segmentation,” in Neurocomputing, vol. 493, pp. 204-216, 2022

  10. [18]

    L. Yu, S. Wang, X. Li, C. W. Fu, and P. A. Heng, ”Uncertainty-aware self-ensembling model for semi-supervised 3D left atrium segmenta- tion,” in MICCAI, Part II, vol. 22, pp. 605-613, Springer International Publishing, 2019

  11. [19]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox, ”U-net: Convolutional networks for biomedical image segmentation,” in MICCAI, Part III, vol. 18, pp. 234-241, Springer International Publishing, 2015

  12. [20]

    C. Guo, M. Szemenyei, Y . Pei, Y . Yi, and W. Zhou, ”SD-UNet: A structured dropout U-Net for retinal vessel segmentation,” inIEEE BIBE, pp. 439-444, October 2019

  13. [21]

    Berthelot, N

    D. Berthelot, N. Carlini, I. Goodfellow, N. Papernot, A. Oliver, and C. A. Raffel, ”Mixmatch: A holistic approach to semi-supervised learning,” in NeurIPS, vol. 32, 2019

  14. [22]

    Tarvainen and H

    A. Tarvainen and H. Valpola, ”Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,” in NeurIPS, vol. 30, 2017

  15. [23]

    Isola, J

    P. Isola, J. Y . Zhu, T. Zhou, and A. A. Efros, ”Image-to-image translation with conditional adversarial networks,” in CVPR, pp. 1125-1134, 2017

  16. [24]

    A. D. Hoover, V . Kouznetsova, and M. Goldbaum, ”Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response,” in IEEE TMI, vol. 19, no. 3, pp. 203-210, 2000

  17. [25]

    S. Qiao, W. Shen, Z. Zhang, B. Wang, and A. Yuille, ”Deep co-training for semi-supervised image recognition,” in ECCV, pp. 135-152, 2018

  18. [26]

    X. Luo, M. Hu, T. Song, G. Wang, and S. Zhang, ”Semi-supervised medical image segmentation via cross-teaching between CNN and transformer,” inInternational Conference on Medical Imaging with Deep Learning, pp. 820-833, PMLR, December 2022

  19. [27]

    M. R. Amini, V . Feofanov, L. Pauletto, E. Devijver, and Y . Maximov, ”Self-training: A survey,” arXiv arXiv:2202.12040, 2022

  20. [28]

    L. Yang, W. Zhuo, L. Qi, Y . Shi, and Y . Gao, ”St++: Make self-training work better for semi-supervised semantic segmentation,” in CVPR, pp. 4268-4277, 2022

  21. [29]

    L. Wang, Y . Sun, and Z. Wang, ”CCS-GAN: A semi-supervised gen- erative adversarial network for image classification,” in The Visual Computer, pp. 1-13, 2022

  22. [30]

    R. He, Z. Tian, and M. J. Zuo, ”A semi-supervised GAN method for RUL prediction using failure and suspension histories,” in Mechanical Systems and Signal Processing , vol. 168, p. 108657, 2022

  23. [31]

    Ouali, C

    Y . Ouali, C. Hudelot, and M. Tami, ”Semi-supervised semantic seg- mentation with cross-consistency training,” in CVPR, pp. 12674-12684, 2020

  24. [32]

    J. Fan, B. Gao, H. Jin, and L. Jiang, ”Ucc: Uncertainty guided cross- head co-training for semi-supervised semantic segmentation,” in CVPR, pp. 9947-9956, 2022

  25. [33]

    Tan and Q

    M. Tan and Q. Le, ”Efficientnet: Rethinking model scaling for convo- lutional neural networks,” in ICML, pp. 6105-6114, PMLR, May 2019

  26. [34]

    Dammann, M

    O. Dammann, M. E. Hartnett, and A. Stahl, ”Retinopathy of prematu- rity,” in Dev. Med. Child Neurol. , vol. 65, no. 5, pp. 625–631, 2023

  27. [35]

    T. K. Redd, J. P. Campbell, J. M. Brown, S. J. Kim, S. Ostmo, R. V . P. Chan, J. Dy, D. Erdogmus, S. Ioannidis, J. Kalpathy-Cramer, M. F. Chiang, ”Evaluation of a deep learning image assessment system for detecting severe retinopathy of prematurity,” The British Journal of Oph...

  28. [36]

    Madhu, S

    K. Madhu, S. M. John, A. Joseph, and B. Abraham, ”A study on Retinopathy of Prematurity (ROP) Screening Using Deep Learning Approaches and a Blood Vessel Segmentation Framework for ROP Affected Eyes,” in IEEE ICETITE, pp. 1-10, February 2024

  29. [37]

    F. Lv, T. Liang, X. Chen, and G. Lin, ”Cross-domain semantic segmen- tation via domain-invariant interactive relation transfer,” in CVPR, pp. 4334-4343, 2020

  30. [38]

    Minsu, S

    K. Minsu, S. Joung, S. Kim, J. Park, I. Kim, and K. Sohn. ”Cross-domain grouping and alignment for domain adaptive semantic segmentation.” in AAAI, vol. 35, no. 3, pp. 1799-1807. 2021

  31. [39]

    Cheng, Q

    C. Cheng, Q. Dou, H. Chen, J. Qin, and P. Heng. ”Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation.” in AAAI, vol. 33, no. 01, pp. 865-872. 2019

  32. [40]

    Xiaohui, S

    L. Xiaohui, S. Niu, X. Gao, X. Zhou, J. Dong, and H. Zhao. ”Self- training adversarial learning for cross-domain retinal OCT fluid seg- mentation.” in Computers in Biology and Medicine , 155, 2023

  33. [41]

    Shujun, L

    W. Shujun, L. Yu, K. Li, X. Yang, C.W. Fu, and P.A. Heng. ”Boundary and entropy-driven adversarial learning for fundus image segmentation.” in MICCAI, Part I 22, pp. 102-110. Springer International Publishing, October 2019

  34. [42]

    Yaroslav, E

    G. Yaroslav, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. March, and V . Lempitsky. ”Domain-adversarial training of neural networks.” Journal of machine learning research , vol. 17, no. 59, pp: 1-35, 2016

  35. [43]

    F., Wei, L. Wang, L. Ju, X. Zhao, X. Wang, X. Shi, and Z. Ge. ”Un- supervised domain adaptive fundus image segmentation with category- level regularization.” in MICCAI, pp. 497-506. Cham: Springer Nature Switzerland, 2022

  36. [44]

    Hao, and M

    G. Hao, and M. Liu. ”Domain adaptation for medical image analysis: a survey.” in IEEE TBME, 69, no. 3, pp: 1173-1185, 2021

  37. [45]

    Meila, ”Comparing clusterings by the variation of information”, in Learning Theory and Kernel Machines , pp, 173-187, Springer Berlin Heidelberg, August 2003

    M. Meila, ”Comparing clusterings by the variation of information”, in Learning Theory and Kernel Machines , pp, 173-187, Springer Berlin Heidelberg, August 2003

  38. [46]

    W. M. Rand, ”Objective criteria for the evaluation of clustering meth- ods”, in Journal of the American Statistical association , no. 336, pp. 846-850, 1971

  39. [47]

    K. Li, S. Wang, L. Yu, and P. A. Heng, ”Dual-teacher++: Exploiting intra-domain and inter-domain knowledge with reliable transfer for cardiac segmentation,” IEEE TMI, vol. 40, no. 10, pp. 2771-2782, 2020

  40. [48]

    and Kulkarni, S., ”HVDROPDB datasets for research in retinopathy of prematurity,” Data in Brief , vol

    Agrawal, R., Walambe, R., Kotecha, K., Gaikwad, A., Deshpande, C.M. and Kulkarni, S., ”HVDROPDB datasets for research in retinopathy of prematurity,” Data in Brief , vol. 52, p.109839, 2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.