Pith. sign in

REVIEW 4 major objections 4 minor 30 references

PC-Seg claims that a five-stage curriculum of cross-view consistency and ensemble pseudo-labeling lifts sparse 2D OCT annotations to 3D segmentation accuracy on par with full supervision using only 0.7% of the training labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:11 UTC pith:OZVZPWO3

load-bearing objection PC-Seg is a sensible five-stage curriculum for 3D OCT segmentation from sparse 2D labels, but the 0.7% claim isn't fully backed by single-run experiments and unmeasured pseudo-labels. the 4 major comments →

arxiv 2607.17718 v2 pith:OZVZPWO3 submitted 2026-07-20 cs.CV

PC-Seg: Progressive Cross-View Consistency for 3D OCT Segmentation from Sparse 2D Annotations

Tsubasa Konno , Takahiro Ninomiya , Yukun Zhou , Koichi Ito , Siegfried K. Wagner , Yiqun Lin , Pearse A. Keane , Toru Nakazawa
show 1 more author
Takafumi Aoki
This is my paper
classification cs.CV
keywords OCT segmentationretinal layer segmentationsemi-supervised learningcurriculum learningcross-view consistencypseudo-labeling3D medical imagingsparse annotations
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

PC-Seg is a training strategy for 3D OCT segmentation that starts from just a handful of annotated 2D B-scans and still ends with accuracy comparable to fully supervised training. The paper proposes a five-stage curriculum: train a single 2D model on the annotated B-scans, adapt it to orthogonal slices using its own pseudo-labels, cross-teach the two planes, distill the resulting pseudo-labels into a 3D model, and finally co-train the 2D and 3D models with ensemble pseudo-labeling. On the MSHC dataset, the ensemble reaches an F-score of 0.9174 with 60 labeled patches (about 0.7% of training data), against 0.9193–0.9212 for fully supervised baselines; on Duke DME it reaches a mean F-score of 0.848 including fluid detection at 0.742. The central claim is that progressive cross-view consistency prevents error accumulation and lets unlabeled orthogonal slices and volumes substitute for dense manual voxel labels.

Core claim

The paper claims that the bottleneck in semi-supervised OCT segmentation is not the lack of labels but the failure to exploit 3D spatial context, and that this context can be injected through a curriculum rather than through dense annotations. Its discovery is that a single 2D model, trained first on standard B-scans and then forced to agree with its own predictions on orthogonal slices, generates volumetric pseudo-labels accurate enough to teach a 3D model; once the 3D model exists, the two models refine each other in a co-training loop. The authors report that this pipeline closes the gap to fully supervised performance in both a layer-segmentation dataset (MSHC) and a pathological dataset

What carries the argument

The load-bearing mechanism is the five-stage progressive curriculum with ensemble pseudo-labeling. Stage 1 trains a 2D model on sparse B-scans with a weak-to-strong consistency regularization baseline; Stage 2 adds orthogonal slices under the 2D model's own pseudo-labels; Stage 3 lets the two planes cross-teach each other; Stage 4 distills ensembled two-plane predictions into a 3D model; Stage 5 alternately trains both models against dynamically updated ensemble pseudo-labels. The key design choice is that the curriculum expands dimensions and views gradually rather than training a 3D model from scratch, and that the pseudo-labeling confidence threshold is lowered from 0.95 to 0.5 for lesion

Load-bearing premise

The five-stage curriculum works only if the pseudo-labels produced at each stage are accurate enough to supervise the next stage, and in Stages 4 and 5 this includes accepting the model's own predictions at confidence threshold 0.5 for lesion classes; if the initial 2D model makes systematic errors on orthogonal slices, those errors propagate into the 3D model and can be reinforced by co-training.

What would settle it

Take a few training volumes from MSHC or Duke DME with complete manual annotations (available in the same datasets), and at the end of Stage 2 and Stage 3 compute the Dice score of the generated pseudo-labels for orthogonal B-scans against these ground-truth slices. If the Stage-2 pseudo-label Dice is no higher than the Stage-1 B-scan model's Dice on B-scans, or if the Stage-3 pseudo-labels show no gain over Stage-2, the progressive-refinement mechanism is not doing the work the paper attributes to it. A stronger test: randomize the stage order (e.g., train the 3D model first, then adapt to 2D

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • With 60 labeled B-scan patches, PC-Seg reaches an F-score of 0.9174 on MSHC, essentially matching fully supervised baselines (0.9193–0.9212) while using 0.7% of the labels.
  • PC-Seg's 3D model alone (F=0.9020) outperforms a standard 3D model trained with the same semi-supervised baseline (0.7843) under sparse labels, showing that the cross-view pseudo-labels teach the 3D model more than direct semi-supervised training does.
  • On Duke DME, the ensemble achieves a Fluid F-score of 0.742 versus 0.3–0.6 for conventional methods, suggesting that 3D context is decisive for complex pathological structures.
  • The 'Only stage 3' ablation (F=0.8713) versus the full pipeline (F=0.9047) shows that skipping the warm-up and orthogonal-adaptation stages causes pseudo-label error accumulation; the progressive order is necessary for the claim.

Where Pith is reading between the lines

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

  • The same progressive cross-view curriculum could transfer to other highly anisotropic 3D imaging modalities (e.g., ultrasound volumes or histology stacks) where annotations are naturally 2D, though the paper does not test this.
  • Because the method uses a single 2D model rather than separate multi-view networks, its parameter and memory footprint is lower than co-training two networks from scratch; the authors cite this as motivation but do not provide a direct compute comparison, so it remains open whether the multi-stage training time offsets that saving.
  • The paper does not measure pseudo-label accuracy against ground truth at each stage; an independent audit of Stage-2/Stage-3 pseudo-label Dice would reveal whether the curriculum's benefit comes from label-quality improvement or from the 3D distillation step alone.
  • A testable extension: if the confidence threshold for lesion classes is perturbed around 0.5 in Stages 4–5, the reported Fluid F-score should be sensitive; mapping this sensitivity would quantify how much the result depends on the threshold choice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes PC-Seg, a five-stage curriculum for semi-supervised 3D OCT segmentation trained from sparse 2D annotations. Stage 1 trains a 2D B-scan model with a weak-to-strong consistency SSL framework; Stage 2 adapts it to orthogonal slices using pseudo-labels; Stage 3 performs cross-teaching between the two orthogonal planes; Stage 4 distills ensemble pseudo-labels into a 3D model; Stage 5 co-trains 2D and 3D models with ensemble pseudo-labeling. On MSHC, with 60 labeled patches (~0.7% of training data), the proposed ensemble reaches F=0.9174, close to fully supervised scores (0.9193–0.9212). On Duke DME, with 66 labeled B-scans, it reaches mean F=0.848 and Fluid F=0.742, outperforming the reported baselines. The central claim is that the progressive curriculum prevents pseudo-label error accumulation and enables fully-supervised-level accuracy from very sparse labels.

Significance. If the reported results are reproducible, the paper makes a practically important contribution: a model-agnostic curriculum that lifts sparse 2D slice annotations to 3D volumetric OCT segmentation, with code publicly released, achieving Dice within ~0.002 of full supervision on MSHC and substantially improving fluid detection on Duke DME. The idea of using a single 2D model to learn cross-view consistency before introducing a 3D model is clean and well motivated, and the ablation study is thoughtfully designed to separate the contributions of orthogonal slices, cross-teaching, distillation, and co-training. The main weaknesses are empirical validation gaps: no uncertainty quantification, no direct measurement of pseudo-label quality, and a partially truncated ablation. These are addressable and do not invalidate the approach, but they currently limit the strength of the headline claims.

major comments (4)
  1. [Sect. 2.2, Table 1] The central premise that Stages 1–2 generate 'reliable' pseudo-labels and that the progressive curriculum 'mitigates error accumulation' is not directly tested. No pseudo-label accuracy against ground truth is reported at any stage. This matters because the threshold reduction to 0.5 for lesion classes in Stages 4–5 (Sect. 3.2) is justified by the assertion that ensemble predictions are 'inherently more robust.' Please add per-stage pseudo-label Dice/volume overlap versus ground truth and a sensitivity analysis of the confidence threshold. Without this, the claimed mechanism of the method is unsupported.
  2. [Table 1, 'Only stage 3' row] The 'Only stage 3' ablation is truncated or confounded as presented. If it is intended as a method that stops after Stage 3, then its 0.8713/0.8748 scores cannot be compared with the final ensemble 0.9047, because the comparison conflates skipping Stages 1–2 with the absence of Stages 4–5. If it is intended to skip Stages 1–2 and still run the full pipeline, the table only reports the Stage-3 checkpoint. In either case, the conclusion that the progressive curriculum is 'essential' for the final 3D model is not established. Please run the ablated scheme through the full pipeline and report the final 2D, 3D, and ensemble scores, or explicitly restrict the claim to Stage-3 accuracy.
  3. [Tables 2 and 3] All results are single-run scores without error bars or statistical significance tests. Several margins are small: the 60-label ensemble F on MSHC is 0.9174 versus 0.9085 for SemiVL and 0.9135 for 2D ResUNet w/ UM; the full-supervision comparison is 0.9212 versus 0.9193/0.9198; on Duke, the mean F is 0.848 versus 0.839 for GD-Net. Without multiple seeds and standard deviations, 'consistently outperforms' and 'significant improvement' are not supported. Please report mean ± std over at least three seeds (and ideally paired tests) for the main comparisons.
  4. [Tables 2 and 3] Several baseline scores are cited from other papers rather than re-run under the same preprocessing, backbone, training schedule, and augmentation protocol. Table 2 notes that results for [10], [17], [7], and [22] are cited from [22]; Table 3 results from Language [26] through GD-Net [3] are cited from [3]. Since the proposed method uses UniMatch, ResUNet, and OCT-specific augmentations, part of the gain may come from these choices rather than from the proposed curriculum. Please re-run the main baselines under the same protocol or explicitly state that the comparison is to reported numbers and restrict 'outperforms' claims to the implemented baselines.
minor comments (4)
  1. [Sect. 3.2] The number of epochs per stage, the pseudo-label update frequency, and the ensemble fusion rule are not specified. Since the curriculum is defined by 'predefined number of epochs,' these details are necessary for reproducibility. The public code helps, but the paper should be self-contained.
  2. [Abstract / Sect. 3.1] The abstract states 'labels for only about 0.7% of the training data,' but the Duke DME experiment uses 66 labeled B-scans out of 366 training B-scans (~18%). Please clarify that the 0.7% figure refers to the MSHC patch-level setting and describe the Duke labeling budget explicitly.
  3. [Sect. 2.1] The paper uses 'F-score' for Dice score. Define this explicitly at first use and state whether background is included; the current text says background is included for MSHC and excluded for Duke, but this is only stated in Sect. 3.1, after the metric is already used.
  4. [Table 2] For the 'Full (8,820)' column, clarify whether all methods are trained with the same augmentations and training schedule, or whether this column mixes reported numbers with re-run baselines. A footnote distinguishing 're-run' from 'cited' would improve interpretability.

Circularity Check

0 steps flagged

No load-bearing circularity: reported F-scores are computed against held-out ground truth; the two self-citations are component-level and non-load-bearing.

full rationale

The paper's derivation chain is externally grounded and shows no load-bearing circularity. The central claim — that PC-Seg matches fully supervised performance with only ~0.7% of labels (60/8,820 patches on MSHC) — is supported by test-set evaluation on held-out subjects (MSHC: 20 test volumes; Duke: 2 test volumes) using ground-truth annotations, so the reported F-scores cannot be equal to any training input by construction. Self-training and cross-teaching are inherent to the method (pseudo-labels generated from model predictions in Stages 1–5 and fed back), but the final metrics are computed against ground truth, not against the pseudo-labels; the resulting confirmation-bias risk is a correctness/robustness concern, not a definitional circularity under the rubric. The two self-citations — [13] FDDA/PRLC augmentations and [14] 1D+2D U-Net — are components and baselines, not the claimed novelty; no 'uniqueness theorem' from the authors' prior work is invoked to force the design, and the adoption of UniMatch [29] and ResUNet rests on external work or an unshown 'preliminary evaluations' claim rather than on a self-citation chain. Two evidence gaps should be flagged, though neither constitutes circularity: (i) Section 3.2 asserts that lowering the pseudo-label threshold to 0.5 for lesion classes at Stages 4–5 is safe because ensemble predictions are 'inherently more robust and stable, effectively preventing the inclusion of erroneous pseudo-labels,' yet pseudo-label accuracy is never measured against ground truth at any stage; (ii) the 'Only stage 3' ablation in Table 1 is truncated at Stage 3 (B-scan F=0.8713) and is not carried through the Stage-4 distillation and Stage-5 co-training, so the claim that the progressive curriculum 'mitigates error accumulation' is supported only by an intermediate-stage gap rather than a full-curriculum counterfactual. These are support gaps affecting the strength of the evidence for the curriculum's necessity, not a reduction of the result to its inputs. Score 2 reflects only the minor, non-load-bearing self-citations present in the paper.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method rests on the reliability of self-generated pseudo-labels and on treating orthogonal slices as useful supervision, both of which are domain assumptions rather than derived facts. No new physical entities are introduced. The main hand-tuned quantities are SSL hyperparameters, especially the confidence threshold whose reduction to 0.5 for lesion classes is a post-hoc adjustment that directly influences the Duke DME fluid score.

free parameters (4)
  • Confidence threshold tau = 0.95 initial; 0.5 for lesion classes in Stages 4-5
    Controls which pseudo-labeled pixels are used for training; the threshold was lowered specifically after observing low confidence on lesion classes, directly affecting the reported Fluid F-score.
  • Stage durations (epochs) = not reported
    The text says each stage proceeds for a predefined number of epochs but never gives the numbers; the curriculum's success depends on these durations.
  • Consistency-loss weight lambda = not reported
    The total SSL loss is a weighted sum of supervised and unsupervised terms, but the weight is not specified; it determines how strongly pseudo-labels drive training.
  • Ensemble fusion rule = not reported
    The final Ensemble scores combine B-scan 2D, ortho 2D, and 3D predictions, but the paper does not state whether this is an equal average or weighted fusion.
axioms (4)
  • domain assumption Self-generated pseudo-labels are accurate enough to supervise the next stage without unbounded confirmation bias.
    Stages 2, 4, and 5 rely on pseudo-labels from the model's own earlier checkpoints as training targets; the only safeguard described is the curriculum order and confidence threshold. The 'Only stage 3' ablation shows this can fail.
  • domain assumption Orthogonal slices resampled from OCT volumes provide meaningful, non-redundant supervision despite strong anisotropy.
    Stage 2 introduces ortho B-scan pseudo-labels as additional unlabeled data; the method assumes these views help the 2D model become spatially consistent rather than injecting slice-direction artifacts.
  • domain assumption Retinal flattening preprocessing preserves the information needed for segmentation.
    All volumes are flattened before training and testing, following prior works. The method's applicability to non-flattened or differently preprocessed OCT is not established.
  • domain assumption The data splits are representative and the reported single-split results generalize to held-out patients and datasets.
    MSHC uses one 12/3/20 subject split and Duke DME uses one 6/2/2 split; no cross-validation or multi-seed results are provided, so the point estimates assume split stability.

pith-pipeline@v1.3.0-alltime-deepseek · 9256 in / 13177 out tokens · 143583 ms · 2026-08-01T17:11:12.698343+00:00 · methodology

0 comments
read the original abstract

Volumetric segmentation of optical coherence tomography (OCT) images is essential for diagnosing ocular diseases but requires labor-intensive voxel-wise annotations. While semi-supervised learning (SSL) can reduce annotation costs, most existing methods process data slice by slice and fail to exploit the inherent 3D spatial context. We propose PC-Seg, a progressive cross-view consistency framework that learns high-accuracy 3D segmentation models from sparse 2D annotations. Unlike conventional multi-view approaches, PC-Seg uses a single 2D model to learn cross-view consistency from standard B-scans and orthogonal slices, thereby generating reliable volumetric pseudo-labels. These pseudo-labels are then distilled into a 3D model, followed by a co-training stage in which the 2D and 3D models mutually refine each other through ensemble pseudo-labeling. Experiments on the MSHC and Duke DME datasets demonstrate that PC-Seg achieves accuracy comparable to fully supervised learning while using labels for only about 0.7% of the training data, outperforming state-of-the-art semi-supervised and retinal layer segmentation methods. Our code is publicly available at https://github.com/gsisaoki/pc-seg-official.

Figures

Figures reproduced from arXiv: 2607.17718 by Koichi Ito, Pearse A. Keane, Siegfried K. Wagner, Takafumi Aoki, Takahiro Ninomiya, Toru Nakazawa, Tsubasa Konno, Yiqun Lin, Yukun Zhou.

Figure 1
Figure 1. Figure 1: Overview of the five-stage curriculum learning in PC-Seg. The 2D model trained in Stage 3 is utilized again in Stage 5 for the final ensemble. The “Train” arrows in each stage denote the optimization using the baseline SSL framework in Sect. 2.1. the baseline SSL framework, we adopt UniMatch [29], a weak-to-strong con￾sistency regularization method. UniMatch first generates a reliable pseudo-label from the… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison of segmentation results on the Duke DME dataset. the ensemble prediction. This result demonstrates that incorporating 3D spatial context is crucial for accurately identifying complex pathological structures [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references

  1. [1]

    Cai, H., Qi, L., Yu, Q., Shi, Y., Gao, Y.: 3D medical image se gmentation with sparse annotation via cross-teaching between 3D and 2D netw orks. Proc. Int’l Conf. Medical Image Computing and Computer Assisted Interventio n 14222, 614–624 (Oct 2023)

    2023

  2. [2]

    Cao, G., Wu, Y., Peng, Z., Zhou, Z., Dai, C.: Self-attentio n CNN for retinal layer segmentation in OCT. Biomed. Opt. Express 15(3), 1605–1617 (Mar 2024)

    2024

  3. [3]

    Biomedical Signal Processing and Control 98, 106702 (Dec 2024)

    Cao, G., Zhou, Z., Wu, Y., Peng, Z., Yan, R., Zhang, Y., Jian g, B.: GCN-enhanced spatial-spectral dual-encoder network for simultaneous s egmentation of retinal lay- ers and fluid in OCT images. Biomedical Signal Processing and Control 98, 106702 (Dec 2024)

    2024

  4. [4]

    Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin-Unet: Unet-like pure transformer for medical image segmentation . Proc. European Conf. Computer Vision Workshops 13803, 205–218 (Oct 2023)

    2023

  5. [5]

    Chiu, S.J., Allingham, M.A., Mettu, P.S., Cousins, S.W., Izatt, J.A., Farsiu, S.: Kernel regression based segmentation of optical coherence tomography images with diabetic macular edema. Biomed. Opt. Express 6(4), 1172–1194 (Apr 2015)

    2015

  6. [6]

    Farshad, A., Yeganeh, Y., Gehlbach, P., Navab, N.: Y-Net: A spatiospectral dual- encoder network for medical image segmentation. Proc. Int’ l Conf. Medical Image Computing and Computer Assisted Intervention pp. 582–592 ( Sep 2022)

    2022

  7. [7]

    Fazekas, B., Aresta, G., Lachinov, D., Riedl, S., Mai, J., Schmidt-Erfurth, U., Bogunović, H.: SD-LayerNet: Semi-supervised retinal laye r segmentation in OCT using disentangled representation with anatomical priors . Proc. Int’l Conf. Medical Image Computing and Computer Assisted Intervention pp. 320 –329 (Sep 2022)

    2022

  8. [8]

    IEEE Trans

    Fazekas, B., Aresta, G., Seeböck, P., Mai, J., Schmidt-Er furth, U., Bogunović, H.: SD-RetinaNet: Topologically constrained semi-supervise d retinal lesion and layer segmentation in OCT. IEEE Trans. Med. Imaging pp. 1–15 (Oct 2 025) 10 T. Konno et al

  9. [9]

    Hatamizadeh, A., Tang, Y., Nath, V., Yang, D., Myronenko, A., Landman, B., Roth, H.R., Xu, D.: UNETR: Transformers for 3D medical image segmentation. Proc. IEEE/CVF Winter Conf. Applications of Computer Visio n pp. 574–584 (Jan 2022)

    2022

  10. [10]

    , Saidha, S., Calabresi, P.A., Prince, J.L.: Structured layer surface segmentation for retina OCT using fully convolutional regression networks

    He, Y., Carass, A., Liu, Y., Jedynak, B.M., Solomon, S.D. , Saidha, S., Calabresi, P.A., Prince, J.L.: Structured layer surface segmentation for retina OCT using fully convolutional regression networks. Medical Image Analysi s 68, 101856 (Feb 2021)

    2021

  11. [11]

    Data Brief 22, 601–604 (Feb 2018)

    He, Y., Carass, A., Solomon, S.D., Saidha, S., Calabresi , P.A., Prince, J.L.: Reti- nal layer parcellation of optical coherence tomography ima ges: Data resource for multiple sclerosis and healthy controls. Data Brief 22, 601–604 (Feb 2018)

    2018

  12. [12]

    Hoyer, L., Tan, D.J., Naeem, M.F., Gool, L.V., Tombari, F .: SemiVL: Semi- supervised semantic segmentation with vision-language gu idance. Proc. European Conf. Computer Vision pp. 257–275 (Oct 2024)

    2024

  13. [13]

    Konno, T., Ninomiya, T., Miura, K., Ito, K., Himori, N., S harma, P., Nakazawa, T., Aoki, T.: Formula-driven data augmentation and partial ret inal layer copying for retinal layer segmentation. Proc. Ophthalmic Medical Imag e Analysis Workshop on MICCAI pp. 136–145 (Oct 2024)

    2024

  14. [14]

    Konno, T., Ninomiya, T., Miura, K., Ito, K., Himori, N., S harma, P., Nakazawa, T., Aoki, T.: Retinal layer segmentation using 1D+2D U-Net f rom OCT images. Proc. IEEE Int’l Symp. Biomedical Imaging pp. 1–5 (May 2024)

    2024

  15. [15]

    Lang, A., Carass, A., Hauser, M., Sotirchos, E.S., Calab resi, P.A., Ying, H.S., Prince, J.L.: Retinal layer segmentation of macular OCT ima ges using boundary classification. Biomed. Opt. Express 4(7), 1133–1152 (Jul 2013)

    2013

  16. [16]

    Medical Image Analysis 91(103019), 1–14 (Jan 2024)

    Liu, H., Wei, D., Lu, D., Tang, X., Wang, L., Zheng, Y.: Sim ultaneous alignment and surface regression using hybrid 2D–3D networks for 3D co herent layer seg- mentation of retinal OCT images with full and sparse annotat ions. Medical Image Analysis 91(103019), 1–14 (Jan 2024)

    2024

  17. [17]

    : Semi-supervised automatic segmentation of layer and fluid region in retinal o ptical coherence to- mography images using adversarial learning

    Liu, X., Cao, J., Fu, T., Pan, Z., Hu, W., Zhang, K., Liu, J. : Semi-supervised automatic segmentation of layer and fluid region in retinal o ptical coherence to- mography images using adversarial learning. IEEE Access 7, 3046–3061 (Dec 2018)

    2018

  18. [18]

    Loshchilov, L., Hutter, F.: Decoupled weight decay regu larization. Proc. Int’l Conf. Learning Representations pp. 1–10 (May 2019)

    2019

  19. [19]

    IEEE Int’l Conf

    Lu, Y., Shen, Y., Xing, X., Meng, M.H.: Multiple consiste ncy supervision based semi-supervised OCT segmentation using very limited annot ations. IEEE Int’l Conf. Robotics and Automation pp. 8483–8489 (May 2022)

    2022

  20. [20]

    Maier, H., Faghihroohi, S., Navab, N.: A line to align: De ep dynamic time warping for retinal OCT segmentation. Proc. Int’l Conf. Medical Ima ge Computing and Computer Assisted Intervention 12901, 709–719 (Sep 2021)

    2021

  21. [21]

    Automatika 62(3-4), 375–385 (Aug 2021)

    Melinščak, M., Radmilović, M., Vatavuk, Z., Lončarić, S .: Annotated retinal optical coherence tomography images (AROI) database for joint reti nal layer and fluid segmentation. Automatika 62(3-4), 375–385 (Aug 2021)

    2021

  22. [22]

    Ong, C.Z.L., Ali, A.A.B., Rajapakse, J.C.: Generative a dversarial learning for semi- supervised retinal layer segmentation in OCT images. Proc. IEEE EMBS Int. Conf. Biomedical and Health Informatics pp. 1–8 (Nov 2024)

    2024

  23. [23]

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolut ional networks for biomedi- cal image segmentation. Proc. Int’l Conf. Medical Image Com puting and Computer Assisted Intervention pp. 234–241 (Oct 2015)

    2015

  24. [24]

    , Wachinger, C., Navab, N.: ReLayNet: Retinal layer and fluid segmentation of macula r optical coherence tomography using fully convolutional networks

    Roy, A., Conjeti, S., Karri, S., Sheet, D., Katouzian, A. , Wachinger, C., Navab, N.: ReLayNet: Retinal layer and fluid segmentation of macula r optical coherence tomography using fully convolutional networks. Biomed. Op t. Express 8(8), 3627– 3642 (Aug 2017) Progressive Cross-View Consistency for 3D OCT Segmentatio n 11

    2017

  25. [25]

    IEEE Trans

    Tan, Y., Shen, W.D., Wu, M.Y., Liu, G.N., Zhao, S.X., Chen , Y., Yang, K.F., Li, Y.J.: Retinal layer segmentation in OCT images with boun dary regression and feature polarization. IEEE Trans. Medical Imaging 43(2), 686–700 (Feb 2024)

    2024

  26. [26]

    Tran, A., Weiss, J., Albarqouni, S., Faghi Roohi, S., Nav ab, N.: Retinal layer segmentation reformulated as OCT language processing. Pro c. Int’l Conf. Medical Image Computing and Computer Assisted Intervention 12265, 694–703 (Oct 2020)

    2020

  27. [27]

    Xia, Y., Liu, F., Yang, D., Cai, J., Yu, L., Zhu, Z., Xu, D., Yuille, A., Roth, H.: 3D semi-supervised learning with uncertainty-aware multi -view co-training. Proc. IEEE/CVF Winter Conf. Applications of Computer Vision pp. 3 646–3655 (Mar 2020)

    2020

  28. [28]

    Xie, E., Wang, W., Yu, Z., Anandkumar, A., Alvarez, J.M., Luo, P.: Segformer: Simple and efficient design for semantic segmentation with tr ansformers. Proc. Advances in Neural Information Processing Systems 34, 12077–12090 (Dec 2021)

    2021

  29. [29]

    Yang, L., Qi, L., Feng, L., Zhang, W., Shi, Y.: Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. Proc. IEEE/CVF Conf. Computer Vi- sion and Pattern Recognition pp. 7236–7246 (Jun 2023)

    2023

  30. [30]

    Zhang, Y., Liu, J., Liao, B., Wang, Q.: Multi-task semi-s upervised 3D medical image segmentation based on CNN and self-attention. Proc. I nt’l Conf. Health Big Data and Intelligent Healthcare pp. 233–237 (Dec 2024)

    2024