Pith. sign in

REVIEW 6 major objections 5 minor 65 references

P3Net: Progressive and Periodic Perturbation for Semi-Supervised Medical Image Segmentation

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A progressive, periodic CutMix schedule plus a boundary-focused loss achieves state-of-the-art semi-supervised medical image segmentation on four benchmark datasets.

desk verdict A plausible and genuinely new perturbation schedule with real empirical signal, but the defining equations are self-contradictory and the data reporting has errors; worth a serious referee, not citable as-is. read the letter →

arxiv 2505.15861 v1 pith:YZDIJNJW submitted 2025-05-21 eess.IV cs.CV

Zhenyan Yao , Miao Zhang , Lanhu Wu , Yongri Piao , Feng Tian , Weibing Sun , Huchuan Lu This is my paper
classification eess.IVcs.CV
keywords semi-supervisedlearningmedicalimagesegmentationinterpolation-basedperturbationCutMixboundary-focusedlossprogressiveperiodicmeanteacherconsistencyregularization
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

This paper argues that interpolation-based perturbation for semi-supervised medical image segmentation should follow a human-like learning schedule: the perturbation strength, defined as the fraction of an unlabeled image pasted into a labeled image via CutMix, should ramp gradually from 0.25 to 0.9 and then reset, rather than staying fixed or being chosen randomly. It also introduces a boundary-focused weighted cross-entropy and Dice loss intended to penalize errors near interpolation boundaries. The paper reports state-of-the-art Dice scores on ACDC, Decathlon Prostate, LA, and Pancreas-NIH at 5-10% labeled data, with the largest gain on Pancreas-NIH (79.69% Dice with 6% labels, 14.25 points above the previous best). It further shows that the schedule and the loss each improve existing interpolation methods such as BCP, ClassMix, and MUM, making both components plug-in additions.

What carries the argument

The core machinery is the progressive and periodic perturbation schedule, a time-varying CutMix ratio $\alpha(\mathrm{iter}) = \exp(\mathrm{iter}/S) + \gamma$ that rises from 0.25 to 0.9 over a period $T=8000$ iterations and then resets, controlling the area of unlabeled content pasted into labeled images in a mean-teacher framework. The second component is a boundary-focused loss consisting of weighted cross-entropy and Dice terms whose per-pixel weight $\mu_{i,j}$ is meant to flag pixels that disagree with their $5\times5$ surroundings, applied through a mask $M_l$ that isolates a band around the interpolation box. The weight equation as printed, however, sums the center pixel's own correctness indicator over the neighborhood, so it does not actually depend on surrounding pixels; the boundary-focusing behavior therefore rests on an unverified correction to Eq. (11).

What would settle it

Construct a map where a center pixel's prediction matches its label but all 24 pixels in its 5x5 neighborhood are wrong; substituting into Eq. (11) gives $\mu = 0$, even though the pixel sits on a strong boundary. The same substitution gives $\mu = 1$ for a wrong pixel surrounded by correct pixels, so the printed formula does not respond to neighborhood context and cannot implement the claimed boundary weighting.

Watch

Extended reading notes

Core claim

The central claim is that dynamically controlling interpolation strength in a progressive and periodic way turns unlabeled data into positive guidance rather than noise, and that a boundary-focused loss corrects the edge errors that mixed-image training creates. In the proposed P3Net, a mean-teacher network is trained on CutMix-style images whose labeled/unlabeled mix ratio follows $\alpha(\mathrm{iter}) = \exp(\mathrm{iter}/S) + \gamma$, rising from 0.25 to 0.9 over a period of 8000 iterations and then resetting. The boundary-focused loss re-weights pixel-level cross-entropy and Dice terms by a factor $\mu$ that is described as large where a pixel disagrees with its $5\times5$ surroundings, and it is applied through a mask that selects a band around the interpolation box. The paper states that these two components achieve state-of-the-art results on two 2D and two 3D datasets and that each component improves existing interpolation-based methods when transplanted into them.

Load-bearing premise

The boundary-focused loss assumes Eq. (11) computes a per-pixel weight that grows where a pixel disagrees with its 5x5 neighborhood; as printed, the sum compares the center pixel's prediction and label only, so the weight reduces to a per-pixel correctness indicator and the boundary-focusing mechanism is not actually specified.

Editorial extensions

If this is right

  • On the four datasets tested, the reported Dice, Jaccard, HD95, and ASD values improve over prior methods at the same labeled-data ratios, and at 5-6% labels the model sometimes outperforms baselines that use 10% labels.
  • The P3M schedule can be grafted onto existing interpolation methods such as BCP and ClassMix, improving their Dice; the paper reports a 28.30% Dice gain for ClassMix on ACDC at 5% labels.
  • The boundary-focused loss can be added to interpolation methods such as BCP, ClassMix, and MUM, improving their Dice; the paper reports a 5.55% Dice gain for MUM on ACDC at 5% labels.
  • The schedule's period and its upper and lower bounds are tunable, and the paper's ablation suggests a moderate period (T=8000) with bounds 0.25 and 0.9 is near-optimal, while extreme bounds degrade performance.
  • The method transfers across 2D (U-Net backbone) and 3D (V-Net backbone) datasets without additional network components.

Reading between the lines

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

  • If the periodic reset is the key ingredient, the exact curve shape may matter less than the bounds and the period; the paper's own ablation shows several curve families performing comparably, which points to a curriculum over perturbation strength rather than the specific exponential formula.
  • A testable extension is to combine P3M with pseudo-label or contrastive consistency losses rather than only interpolation losses, to see whether the schedule generalizes to other perturbation families in semi-supervised learning.
  • The boundary-weighting idea could be transferred to any task with composited or stitched training images, such as domain adaptation or object detection on mosaics, wherever a mask identifies artificial seams whose edges lack context.
  • The large reported gain on Pancreas-NIH at 6% labels is worth stress-testing with different seeds and data splits to confirm that the schedule, rather than dataset idiosyncrasy, drives the improvement.
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

6 major / 5 minor

Summary. The paper proposes P3Net, a semi-supervised medical image segmentation method whose main components are a progressive and periodic perturbation mechanism (P3M) and a boundary-focused weighted cross-entropy/Dice loss. The P3M schedule is intended to ramp the interpolation ratio between labeled and unlabeled data from 0.25 to 0.9 and then reset periodically, while the boundary-focused loss is meant to emphasize pixels near CutMix interpolation boundaries. The authors report state-of-the-art results on ACDC, Decathlon Prostate, LA, and Pancreas-NIH at 5–10% labeled data, and also show plug-in gains when applying P3M or the loss to existing methods such as BCP, ClassMix, and MUM. The central claim is that a gradually increasing, periodically resetting perturbation ratio and a boundary-aware loss improve semi-supervised segmentation accuracy.

Significance. If the method works as described, it is an attractive plug-in contribution: the schedule and loss add no network components, apply to both 2D and 3D inputs, and improve several existing interpolation-based methods. The experimental evaluation is broad (four datasets, multiple labeled ratios, four metrics) and the ablations cover the period, the bounds, and the curve shape. However, the as-printed specification contains several load-bearing ambiguities and errors: the direction of alpha in the mixing equation is contradictory, Eq. (11) does not implement the claimed neighborhood comparison, Eq. (14) applies a mask to scalar losses, and the main table contains a full-supervision copy-paste error. In addition, the schedule hyperparameters are selected on the same evaluation sets used to report final performance. These issues do not necessarily invalidate the underlying idea, but they prevent the reader from reproducing or independently verifying the reported numbers as the paper now stands.

major comments (6)
  1. [Section 3.2, Eqs. (5)-(8)] The quantity alpha is not defined consistently. The sentence preceding Eq. (5) says alpha_1(iter) 'represents the proportion of labeled data in the interpolation at iteration iter, specifically the area occupied by unlabeled data after using CutMix.' Eq. (7) constructs X_mix = X_u * M_c + X_s * (1 - M_c), and the text states that the zero-value region of M_c has size (alpha*W, alpha*H). Since 1 - M_c equals 1 on that region, X_s occupies fraction alpha of the mixed image; under this reading alpha is the labeled fraction. The schedule then decreases the unlabeled fraction from 0.75 to 0.1, contradicting the stated goal of learning more from unlabeled data. Section 4.4 and Table 5 read alpha as the unlabeled fraction ('setting the lower bound to 0 will reduce the amount of unlabeled data the network learns'). The paper therefore does not uniquely specify the schedule that produced Tables 1-5.
  2. [Section 3.3, Eq. (11)] The boundary-weight mu_{i,j} is defined as 1 minus the average of 1(y^s_{i,j} = p^s_{i,j}) over A_{i,j}, where the indicator compares the fixed center indices (i,j). The sum therefore contains 25 identical terms and mu_{i,j} collapses to 0 or 1 depending only on whether the center pixel agrees with its own prediction. This does not measure the difference between the center pixel and its surroundings, so the boundary-focused loss as written cannot target interpolation-boundary pixels. The authors need to provide the correct neighbor comparison, e.g., comparing y^s_{k,l} with p^s_{k,l} over the 5x5 window, and re-run the affected experiments if the implementation differs from the printed formula.
  3. [Section 3.3, Eqs. (9)-(14)] L_wce, L_wdice, L_bf, and L_seg are written as scalar losses, but Eq. (14) applies the mask M_l with element-wise multiplication. Unless these losses are per-pixel maps, Eq. (14) multiplies a scalar by a mask, which does not produce the intended spatial weighting. The authors should define the per-pixel loss maps explicitly and clarify how M_l is applied to them.
  4. [Table 1, full-supervision rows] The 100% labeled-data rows report identical Dice/Jaccard/HD95/ASD values for ACDC and Decathlon Prostate, and identical values for LA and Pancreas-NIH. These datasets differ in modality, resolution, and anatomy, so identical full-supervision scores are implausible and indicate a copy-paste error. The corrected numbers may change the reported gap between the semi-supervised method and the full-supervision upper bound.
  5. [Abstract and Section 1] The abstract and introduction state that the method obtains 79.69% Dice with 6% labeled data on Pancreas-NIH, 'which is 14.25% higher than the previous SOTA.' In Table 1, the best baseline at that ratio is MagicNet at 70.06, an absolute difference of 9.63 Dice points (or 11.37 points over PMT). The 14.25% figure is not traceable to the table; please state whether it is absolute, relative, or computed against a different baseline.
  6. [Section 4.4, Tables 3-5] The period T=8000, the bounds (0.25, 0.9), and the chosen red curve are selected by comparing Dice on the ACDC 5% and LA 5% evaluation sets, and the same sets are then used to report the final state-of-the-art numbers. Selecting hyperparameters on the test data overstates expected performance. The authors should use a held-out validation split or nested evaluation for these selections, or clearly report the selection procedure and re-validate on truly held-out data.
minor comments (5)
  1. [Abstract] The abstract says 'two 2D and 3D datasets'; this should be 'two 2D and two 3D datasets'.
  2. [Section 3.1, Eq. (4)] The unsupervised loss in the pre-warm step appears to be missing a '+' sign between the Dice and cross-entropy terms.
  3. [Figure 4 caption] The caption says 'Visualization of 3D datasets LA with 5% labeled data and Pancreas-NIH with 6% labeled data,' but the figure shows ACDC and Decathlon Prostate, which are 2D datasets.
  4. [Sections 2.2 and 5] The acronym 'SSLMIS' is used where 'SSMIS' is intended; please correct these typos.
  5. [Section 3.2, Eq. (7)] The sentence 'We first generate a mask M_c(t) with all elements set to 1' is confusing when followed by 'The size of the zero-value region is (alpha W x alpha H)'; please state that the mask is initialized to 1 and a box of that size is then set to 0.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: P3M is an empirical schedule/loss proposal whose main claims are tested against fixed-ratio and single-cycle baselines, not reduced by construction to its own inputs.

full rationale

The paper's central claims are empirical rather than derived. The P3M schedule (Eqs. 5-8) is explicitly an adopted curve with chosen endpoints, and the paper itself states that 'a reasonable upper and lower limit is far more important than the function itself.' The claim that periodic perturbation helps is supported by comparing periodic T values against the single-cycle T=75000 in Table 4, and the claim that a progressive ramp beats fixed interpolation is tested against the constant-ratio cyan curve in Table 3. Likewise, the boundary-focused loss is an introduced loss term (Eqs. 9-14) whose contribution is isolated in Table 2 by adding it last. The load-bearing comparisons are internal empirical tests, not quantities forced by construction: no parameter is fitted to a subset of data and then reported as a prediction of that same subset, no cited result by the same authors is used to preclude alternatives, and no known empirical pattern is merely renamed. The manuscript does contain serious technical ambiguities: Eq. (7) conflicts with the surrounding text about whether alpha is the labeled or unlabeled fraction; Eq. (11) as printed makes mu_{i,j} constant over the image because the indicator compares only the fixed center indices; and the final choices T=8000, bounds 0.25/0.9, and the red curve are selected by ablations on the same benchmark datasets. These are reproducibility and validation concerns rather than circularity; correcting them would not change the fact that the schedule-vs-constant and periodic-vs-single-cycle comparisons are honest empirical tests. Under the circularity rubric, the honest finding is no significant circularity.

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

The method rests on standard mean-teacher and CutMix machinery plus two hand-designed components: a periodic interpolation schedule and a boundary weight. The schedule's period, bounds and curve shape are empirically fitted; the boundary weight's stated computation is not implementable as printed. No new physical or conceptual entities are introduced.

free parameters (5)
  • Period T = 8000 iterations
    Set by ablation; Table 4 shows T=8000 best on ACDC and LA.
  • Upper interpolation bound = 0.9
    Chosen from the grid in Table 5 on ACDC 5%; controls maximum unlabeled share in CutMix.
  • Lower interpolation bound = 0.25
    Chosen from the grid in Table 5; controls starting unlabeled share.
  • Boundary ring width epsilon = 13 pixels
    Set in Sec. 4.2; not ablated, controls size of the boundary-focused loss region.
  • Neighborhood size for mu = 5x5
    Chosen in Eq. 11; not ablated.
assumptions (5)
  • domain assumption Teacher EMA provides reliable pseudo-labels for unlabeled data
    Eq. 1 (EMA) and Eq. 4 supervise the student with teacher pseudo-labels; standard in mean-teacher SSL, not proved here.
  • domain assumption CutMix-style mixing of labeled and unlabeled images with pseudo-label supervision is a valid training signal
    Eqs. 7-8 define the mixed input and pseudo-label; inherited from BCP/CutMix literature.
  • ad hoc to paper A progressive and periodic schedule of the interpolation ratio improves learning
    Core hypothesis of P3M; motivated by human learning but not derived, only empirically ablated (Tables 2-5).
  • ad hoc to paper Local agreement between predictions and pseudo-labels in a 5x5 neighborhood identifies boundary error pixels
    Basis of Eq. 11; as written the equation does not implement this because indices are fixed.
  • domain assumption Pseudo-labels are derived by thresholding teacher probability maps at 0.5 (binary) or argmax
    Sec. 3.2; standard practice but known to inject noise in low-data regimes.

how reviews work

0 comments
Cite this review

Pith. "Pith review of P3Net: Progressive and Periodic Perturbation for Semi-Supervised Medical Image Segmentation." pith.science (2026). https://pith.science/paper/YZDIJNJW

@misc{pith2026250515861,
  author       = {Pith},
  title        = {Pith review of: P3Net: Progressive and Periodic Perturbation for Semi-Supervised Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YZDIJNJW}},
  note         = {Machine review of arXiv:2505.15861}
}
read the original abstract

Perturbation with diverse unlabeled data has proven beneficial for semi-supervised medical image segmentation (SSMIS). While many works have successfully used various perturbation techniques, a deeper understanding of learning perturbations is needed. Excessive or inappropriate perturbation can have negative effects, so we aim to address two challenges: how to use perturbation mechanisms to guide the learning of unlabeled data through labeled data, and how to ensure accurate predictions in boundary regions. Inspired by human progressive and periodic learning, we propose a progressive and periodic perturbation mechanism (P3M) and a boundary-focused loss. P3M enables dynamic adjustment of perturbations, allowing the model to gradually learn them. Our boundary-focused loss encourages the model to concentrate on boundary regions, enhancing sensitivity to intricate details and ensuring accurate predictions. Experimental results demonstrate that our method achieves state-of-the-art performance on two 2D and 3D datasets. Moreover, P3M is extendable to other methods, and the proposed loss serves as a universal tool for improving existing methods, highlighting the scalability and applicability of our approach.

Figures

Figures reproduced from arXiv: 2505.15861 by the authors.

Figure 1
Figure 1. We assessed the learning progress of unlabeled data every 200 iterations. The Dice coe [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1-2
Figure 1-2. Based on the above observation, we argue that the perturbation-based methods should be adjusted to form a gradual and intentional learning mechanism for better adapting to SSMIS. From the perspective of education and cognitive psychology [27], 3 [PITH_FULL_IMAGE:figures/full_fig_p003_1-2.png] view at source ↗
Figure 2
Figure 2. The whole pipeline of our proposed P3Net (left) and progressive and periodic perturbation mecha￾nism (right) in the second stage. mechanism. 3. Method 3.1. Overview Following the SSMIS [15, 7], we aim to train the segmentation encoder and decoder on both labeled and unlabeled data simultaneously. Without introducing extra network components, P3M leverages a progressively changing interpolation ratio to enable the ne… view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: Visualization of 3D datasets LA with 5% labeled data and Pancreas-NIH with 6% labeled data and [PITH_FULL_IMAGE:figures/full_fig_p014_3.png]
Figure 4
Figure 4. Figure 4: Visualization of 3D datasets LA with 5% labeled data and Pancreas-NIH with 6% labeled data and [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Demonstrate a period of different P3M curve functions. (Corresponding to each row in Tab. 3.) Period Dice Jaccard HD95 ASD ACDC T = 4000 89.07 80.94 3.09 1.10 T = 8000(our) 89.71 81.88 1.85 0.64 T = 12000 89.54 81.60 1.63 0.55 T = 16000 89.24 81.16 2.36 0.67 T = 75000 …
Figure 6
Figure 6. Figure 6: Application of P3M. Results are evaluated on the ACDC(5%) and LA(5%) dataset. Blue represents without our P3M, and orange represents with our P3M. (best viewed by zoom-in on screen.) the upper bound will increase the proportion of unlabeled data, introduce more noise, …
Figure 7
Figure 7. Figure 7: Application of boundary-focused loss. Results are evaluated on the ACDC(5%) and LA(5%) [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 56 canonical work pages

  1. [1]

    Clinically applicable deep learning for diagnosis and re- ferral in retinal disease.Nature medicine, 24(9):1342–1350, 2018

    Jeffrey De Fauw et al. Clinically applicable deep learning for diagnosis and re- ferral in retinal disease.Nature medicine, 24(9):1342–1350, 2018

    work page 2018

  2. [2]

    Video-based ai for beat-to-beat assessment of cardiac func- tion.Nature, 580(7802):252–256, 2020

    David Ouyang et al. Video-based ai for beat-to-beat assessment of cardiac func- tion.Nature, 580(7802):252–256, 2020

    work page 2020

  3. [3]

    Bidirectional copy-paste for semi-supervised medical image segmentation

    Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. InProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 11514–11524, 2023

    work page 2023

  4. [4]

    Mcf: Mutual correction framework for semi-supervised medical image segmentation

    Yongchao Wang, Bin Xiao, Xiuli Bi, Weisheng Li, and Xinbo Gao. Mcf: Mutual correction framework for semi-supervised medical image segmentation. InPro- ceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 15651–15660, 2023

    work page 2023

  5. [5]

    Pseudo-label guided contrastive learning for semi-supervised medical image segmentation

    Hritam Basak and Zhaozheng Yin. Pseudo-label guided contrastive learning for semi-supervised medical image segmentation. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 19786–19797, 2023. 22

    work page 2023

  6. [6]

    Orthog- onal annotation benefits barely-supervised medical image segmentation

    Heng Cai, Shumeng Li, Lei Qi, Qian Yu, Yinghuan Shi, and Yang Gao. Orthog- onal annotation benefits barely-supervised medical image segmentation. InPro- ceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 3302–3311, 2023

    work page 2023

  7. [7]

    Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium seg- mentation

    Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium seg- mentation. InMedical Image Computing and Computer Assisted Intervention– MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22, pages 605–613. Springer, 2019

    work page 2019

  8. [8]

    Shape-aware semi-supervised 3d semantic segmentation for medical images

    Shuailin Li, Chuyu Zhang, and Xuming He. Shape-aware semi-supervised 3d semantic segmentation for medical images. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 552–561. Springer, 2020

    work page 2020

Show all 65 references
  1. [9]

    Semi-supervised medi- cal image segmentation through dual-task consistency

    Xiangde Luo, Jieneng Chen, Tao Song, and Guotai Wang. Semi-supervised medi- cal image segmentation through dual-task consistency. InProc. AAAI Conf. Artif. Intell. (AAAI), volume 35, pages 8801–8809, 2021

    2021

  2. [10]

    Magicnet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery

    Duowen Chen, Yunhao Bai, Wei Shen, Qingli Li, Lequan Yu, and Yan Wang. Magicnet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 23869–23878, 2023

    2023

  3. [11]

    Local contrastive loss with pseudo-label based self-training for semi-supervised medical image segmentation.Med

    Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Local contrastive loss with pseudo-label based self-training for semi-supervised medical image segmentation.Med. Image Anal., 87:102792, 2023

    2023

  4. [12]

    Reference-guided pseudo-label generation for medical semantic segmentation

    Constantin Marc Seibold, Simon Reiß, Jens Kleesiek, and Rainer Stiefelhagen. Reference-guided pseudo-label generation for medical semantic segmentation. In Proc. AAAI Conf. Artif. Intell. (AAAI), volume 36, pages 2171–2179, 2022. 23

    2022

  5. [13]

    Intra-and inter-pair consistency for semi-supervised gland segmenta- tion.IEEE Trans

    Yutong Xie, Jianpeng Zhang, Zhibin Liao, Johan Verjans, Chunhua Shen, and Yong Xia. Intra-and inter-pair consistency for semi-supervised gland segmenta- tion.IEEE Trans. Image Process., 31:894–905, 2021

    2021

  6. [14]

    Enhancing pseudo label quality for semi-supervised domain-generalized medical image segmentation

    Huifeng Yao, Xiaowei Hu, and Xiaomeng Li. Enhancing pseudo label quality for semi-supervised domain-generalized medical image segmentation. InProc. AAAI Conf. Artif. Intell. (AAAI), volume 36, pages 3099–3107, 2022

    2022

  7. [15]

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight- averaged consistency targets improve semi-supervised deep learning results.Ad- vances in neural information processing systems, 30, 2017

    2017

  8. [16]

    Adaptive hierarchical dual consistency for semi-supervised left atrium segmentation on cross-domain data.IEEE Trans

    Jun Chen et al. Adaptive hierarchical dual consistency for semi-supervised left atrium segmentation on cross-domain data.IEEE Trans. Med. Imag., 41(2):420– 433, 2021

    2021

  9. [17]

    Transformation-consistent self-ensembling model for semisupervised medical image segmentation.IEEE Trans

    Xiaomeng Li, Lequan Yu, Hao Chen, Chi-Wing Fu, Lei Xing, and Pheng- Ann Heng. Transformation-consistent self-ensembling model for semisupervised medical image segmentation.IEEE Trans. Neural Netw. Learn. Syst., 32(2):523– 534, 2020

    2020

  10. [18]

    Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation.IEEE Trans

    Yinghuan Shi, Jian Zhang, Tong Ling, Jiwen Lu, Yefeng Zheng, Qian Yu, Lei Qi, and Yang Gao. Inconsistency-aware uncertainty estimation for semi-supervised medical image segmentation.IEEE Trans. Med. Imag., 41(3):608–620, 2021

    2021

  11. [19]

    Mum: Mix image tiles and unmix feature tiles for semi-supervised object detection

    JongMok Kim, JooYoung Jang, Seunghyeon Seo, Jisoo Jeong, Jongkeun Na, and Nojun Kwak. Mum: Mix image tiles and unmix feature tiles for semi-supervised object detection. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pat- tern Recognit. (CVPR), pages 14512–14521, 2022

    2022

  12. [20]

    Perturbed and strict mean teachers for semi-supervised se- mantic segmentation

    Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised se- mantic segmentation. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 4258–4267, 2022. 24

    2022

  13. [21]

    Instance- specific and model-adaptive supervision for semi-supervised semantic segmenta- tion

    Zhen Zhao, Sifan Long, Jimin Pi, Jingdong Wang, and Luping Zhou. Instance- specific and model-adaptive supervision for semi-supervised semantic segmenta- tion. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 23705–23714, 2023

    2023

  14. [22]

    Revisiting weak-to-strong consistency in semi-supervised semantic segmentation

    Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. InPro- ceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 7236–7246, 2023

    2023

  15. [23]

    Augmentation matters: A simple-yet-effective approach to semi-supervised se- mantic segmentation

    Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised se- mantic segmentation. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 11350–11359, 2023

    2023

  16. [24]

    Dsp: Dual soft-paste for unsupervised domain adaptive semantic segmentation

    Li Gao, Jing Zhang, Lefei Zhang, and Dacheng Tao. Dsp: Dual soft-paste for unsupervised domain adaptive semantic segmentation. InProceedings of the 29th ACM International Conference on Multimedia, pages 2825–2833, 2021

    2021

  17. [25]

    Class- mix: Segmentation-based data augmentation for semi-supervised learning

    Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Class- mix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1369–1378, 2021

    2021

  18. [26]

    Cutmix: Regularization strategy to train strong classifiers with localizable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. InProceedings of the Proc. IEEE Int. Conf. Comput. Vis. (ICCV), pages 6023–6032, 2019

    2019

  19. [27]

    Remind your neural network to prevent catastrophic forgetting

    Tyler L Hayes, Kushal Kafle, Robik Shrestha, Manoj Acharya, and Christopher Kanan. Remind your neural network to prevent catastrophic forgetting. InProc. Eur. Conf. Comput. Vis. (ECCV), pages 466–483. Springer, 2020

    2020

  20. [28]

    Tripled-uncertainty guided mean teacher model for semi-supervised 25 medical image segmentation

    Kaiping Wang, Bo Zhan, Chen Zu, Xi Wu, Jiliu Zhou, Luping Zhou, and Yan Wang. Tripled-uncertainty guided mean teacher model for semi-supervised 25 medical image segmentation. InMedical Image Computing and Computer As- sisted Intervention–MICCAI 2021: 24th International Confere...

    2021

  21. [29]

    Semi-supervised neuron segmentation via reinforced consis- tency learning.IEEE Trans

    Wei Huang, Chang Chen, Zhiwei Xiong, Yueyi Zhang, Xuejin Chen, Xiaoyan Sun, and Feng Wu. Semi-supervised neuron segmentation via reinforced consis- tency learning.IEEE Trans. Med. Imag., 41(11):3016–3028, 2022

    2022

  22. [30]

    Dacs: Domain adaptation via cross-domain mixed sampling

    Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1379– 1389, 2021

    2021

  23. [31]

    mixup: Beyond empirical risk minimization.arXiv:1710.09412, 2017

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization.arXiv:1710.09412, 2017

    2017 arXiv

  24. [32]

    Yolox: Exceeding yolo series in 2021.arXiv:2107.08430, 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021.arXiv:2107.08430, 2021

    2021 arXiv

  25. [33]

    Improved regularization of convolu- tional neural networks with cutout.arXiv:1708.04552, 2017

    Terrance DeVries and Graham W Taylor. Improved regularization of convolu- tional neural networks with cutout.arXiv:1708.04552, 2017

    2017 arXiv

  26. [34]

    Interpolation consistency training for semi- supervised learning.Neural Networks, 145:90–106, 2022

    Vikas Verma, Kenji Kawaguchi, Alex Lamb, Juho Kannala, Arno Solin, Yoshua Bengio, and David Lopez-Paz. Interpolation consistency training for semi- supervised learning.Neural Networks, 145:90–106, 2022

    2022

  27. [35]

    Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems, 32, 2019

    David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning.Advances in neural information processing systems, 32, 2019

    2019

  28. [36]

    Remixmatch: Semi-supervised learning with dis- tribution alignment and augmentation anchoring.arXiv:1911.09785, 2019

    David Berthelot, Nicholas Carlini, Ekin D Cubuk, Alex Kurakin, Kihyuk Sohn, Han Zhang, and Colin Raffel. Remixmatch: Semi-supervised learning with dis- tribution alignment and augmentation anchoring.arXiv:1911.09785, 2019. 26

    1911 arXiv

  29. [37]

    Semi-supervised semantic segmentation needs strong, varied perturba- tions.arXiv:1906.01916, 2019

    GeoffFrench, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Fin- layson. Semi-supervised semantic segmentation needs strong, varied perturba- tions.arXiv:1906.01916, 2019

    1906 arXiv

  30. [38]

    Structured consistency loss for semi-supervised semantic segmentation.arXiv:2001.04647, 2020

    Jongmok Kim, Jooyoung Jang, Hyunwoo Park, and SeongAh Jeong. Structured consistency loss for semi-supervised semantic segmentation.arXiv:2001.04647, 2020

    2001 arXiv

  31. [39]

    Optimization of repetition spacing in the practice of learning.Acta neurobiologiae experimentalis, 54:59–59, 1994

    Piotr A Wozniak and Edward J Gorzelanczyk. Optimization of repetition spacing in the practice of learning.Acta neurobiologiae experimentalis, 54:59–59, 1994

    1994

  32. [40]

    Memory: A contribution to experimental psychology

    Our Knowledge Concerning Memory. Memory: A contribution to experimental psychology

  33. [41]

    A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging.Med

    Zhaohan Xiong et al. A global benchmark of algorithms for segmenting the left atrium from late gadolinium-enhanced cardiac magnetic resonance imaging.Med. Image Anal., 67:101832, 2021

    2021

  34. [42]

    Explor- ing smoothness and class-separation for semi-supervised medical image segmen- tation

    Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Explor- ing smoothness and class-separation for semi-supervised medical image segmen- tation. InInternational Conference on Medical Image Computing and Computer- Assisted Intervention, pages 34–43. Springer, 2022

    2022

  35. [43]

    Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation

    Holger R Roth et al. Deeporgan: Multi-level deep convolutional networks for automated pancreas segmentation. InMedical Image Computing and Computer- Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part I 18, pa...

    2015

  36. [44]

    Deep learning techniques for automatic mri cardiac multi- structures segmentation and diagnosis: is the problem solved?IEEE Trans

    Olivier Bernard et al. Deep learning techniques for automatic mri cardiac multi- structures segmentation and diagnosis: is the problem solved?IEEE Trans. Med. Imag., 37(11):2514–2525, 2018

    2018

  37. [45]

    Cross-patch dense contrastive learning for semi-supervised segmentation of cellular nuclei in 27 histopathologic images

    Huisi Wu, Zhaoze Wang, Youyi Song, Lin Yang, and Jing Qin. Cross-patch dense contrastive learning for semi-supervised segmentation of cellular nuclei in 27 histopathologic images. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 11666–11675, 2022

    2022

  38. [46]

    SSL4MIS.https://github.com/HiLab-git/SSL4MIS, 2020

    Xiangde Luo. SSL4MIS.https://github.com/HiLab-git/SSL4MIS, 2020

    2020

  39. [47]

    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

  40. [48]

    Cat: Constrained adversarial training for anatomically- plausible semi-supervised segmentation.IEEE Trans

    Ping Wang, Jizong Peng, Marco Pedersoli, Yuanfeng Zhou, Caiming Zhang, and Christian Desrosiers. Cat: Constrained adversarial training for anatomically- plausible semi-supervised segmentation.IEEE Trans. Med. Imag., 2023

    2023

  41. [49]

    Temporal ensembling for semi-supervised learning

    Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv:1610.02242, 2016

    2016 arXiv

  42. [50]

    Semi-supervised semantic segmentation with cross pseudo supervision

    Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 2613–2622, 2021

    2021

  43. [51]

    Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation

    Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. InProc. IEEE Int. Conf. Comput. Vis. (ICCV), pages 6930–6940, 2021

    2021

  44. [52]

    Semi-supervised semantic seg- mentation with cross-consistency training

    Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic seg- mentation with cross-consistency training. InProceedings of the Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pages 12674–12684, 2020

    2020

  45. [53]

    Fully convolutional neural networks for volumetric medical image segmentation

    F Milletari, N Navab, SA VN Ahmadi, and V Net. Fully convolutional neural networks for volumetric medical image segmentation. InProceedings of the 2016 Fourth International Conference on 3D Vision (3DV), pages 565–571

    2016

  46. [54]

    U-net: Convolutional net- works for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional net- works for biomedical image segmentation. InMedical Image Computing and 28 Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, P...

    2015

  47. [55]

    Correlation- aware mutual learning for semi-supervised medical image segmentation

    Shengbo Gao, Ziji Zhang, Jiechao Ma, Zihao Li, and Shu Zhang. Correlation- aware mutual learning for semi-supervised medical image segmentation. InIn- ternational Conference on Medical Image Computing and Computer-Assisted In- tervention, pages 98–108. Springer, 2023

    2023

  48. [56]

    Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation

    Wenjing Lu, , et al. Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 662–672. Springer, 2023

    2023

  49. [57]

    Combinatorial cnn-transformer learning with manifold constraints for semi-supervised medical image segmentation

    Huimin Huang, Yawen Huang, Shiao Xie, Lanfen Lin, Ruofeng Tong, Yen-Wei Chen, Yuexiang Li, and Yefeng Zheng. Combinatorial cnn-transformer learning with manifold constraints for semi-supervised medical image segmentation. In Proceedings of the AAAI Conference on Artificial Int...

    2024

  50. [58]

    Consistency-guided differential decoding for enhancing semi- supervised medical image segmentation.IEEE Transactions on Medical Imaging, pages 1–1, 2024

    Qingjie Zeng, Yutong Xie, Zilin Lu, Mengkang Lu, Jingfeng Zhang, Yuyin Zhou, and Yong Xia. Consistency-guided differential decoding for enhancing semi- supervised medical image segmentation.IEEE Transactions on Medical Imaging, pages 1–1, 2024

    2024

  51. [59]

    Pmt: Progressive mean teacher via exploring temporal consistency for semi-supervised medical image segmentation

    Ning Gao, Sanping Zhou, Le Wang, and Nanning Zheng. Pmt: Progressive mean teacher via exploring temporal consistency for semi-supervised medical image segmentation. InEuropean Conference on Computer Vision, pages 144–160. Springer, 2025

    2025

  52. [60]

    Decoupled consis- tency for semi-supervised medical image segmentation

    Faquan Chen, Jingjing Fei, Yaqi Chen, and Chenxi Huang. Decoupled consis- tency for semi-supervised medical image segmentation. InInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention, pages 551–561. Springer, 2023. 29

    2023

  53. [61]

    Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints

    Jinhua Liu, Christian Desrosiers, and Yuanfeng Zhou. Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pa...

    2022

  54. [62]

    Self-aware and cross-sample prototypical learning for semi- supervised medical image segmentation.arXiv:2305.16214, 2023

    Zhenxi Zhang et al. Self-aware and cross-sample prototypical learning for semi- supervised medical image segmentation.arXiv:2305.16214, 2023

    2023 arXiv

  55. [63]

    Adaptive bidirec- tional displacement for semi-supervised medical image segmentation

    Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirec- tional displacement for semi-supervised medical image segmentation. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 4070–4080, 2024

    2024

  56. [64]

    Gradient-aware for class-imbalanced semi- supervised medical image segmentation.pancreas, 98:1–86

    Wenbo Qi1 Jiafei Wu and SC Chan. Gradient-aware for class-imbalanced semi- supervised medical image segmentation.pancreas, 98:1–86

  57. [65]

    Alternate diverse teaching for semi-supervised medical image segmen- tation

    Zhen Zhao, Zicheng Wang, Longyue Wang, Dian Yu, Yixuan Yuan, and Luping Zhou. Alternate diverse teaching for semi-supervised medical image segmen- tation. InEuropean Conference on Computer Vision, pages 227–243. Springer, 2025. 30

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.