Learning Disentangled Stain and Structural Representations for Semi-Supervised Histopathology Segmentation

Ha-Hieu Pham; Huy-Hieu Pham; Min Xu; Nguyen Lan Vi Vu; Thanh-Huy Nguyen; Trung-Nghia Le; Ulas Bagci

REVIEW 3 major objections 6 minor 18 references

Learning Disentangled Stain and Structural Representations for Semi-Supervised Histopathology Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CSDS gives stain and structure separate student networks and claims state-of-the-art low-label gland segmentation.

desk verdict A plausible dual-student SSL method for histopathology whose 'disentanglement' claim is more branding than guarantee; worth reviewing after fixes. read the letter →

arxiv 2507.03923 v2 pith:KZYPPBEC submitted 2025-07-05 cs.CV cs.AI

Ha-Hieu Pham , Nguyen Lan Vi Vu , Thanh-Huy Nguyen , Ulas Bagci , Min Xu , Trung-Nghia Le , Huy-Hieu Pham This is my paper · ORCID

classification cs.CVcs.AI

keywords semi-supervisedsegmentationhistopathologyglandstaindisentanglementmeanteacheruncertaintyestimationH&Estainingrepresentationlearning

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

The paper introduces Color-Structure Dual-Student (CSDS), a semi-supervised segmentation framework for H&E histopathology, and claims it outperforms eight semi-supervised baselines when only 5% or 10% of the data is labeled. The central thesis is that stain appearance and tissue morphology are two distinct, entangled sources of variation, and modeling each with its own student network yields better representations than processing images holistically. On the GlaS and CRAG benchmarks, CSDS reports Dice improvements of up to 1.2% and 0.7% at 5% labels, and 0.7% and 1.4% at 10% labels, over prior methods. If correct, this gives a practical recipe for producing clinically useful gland segmentations with very few annotated slides.

What carries the argument

The load-bearing mechanism is the dual-student decoupling: two student networks of identical architecture receive deliberately different augmentation streams, one aimed at color variability and one at morphological variability, so that each is pushed to become invariant to one nuisance factor. A shared teacher updated by EMA over the averaged student weights supplies pseudo-labels, and per-pixel uncertainty maps—computed from teacher softmax entropy and then amplified in color-variance-rich and edge-rich regions—modulate how strongly each student's unsupervised consistency loss counts. This design attaches different input distributions and different reliability weights to each student, which is what the paper argues produces the low-label gains.

What would settle it

Train CSDS with the augmentations swapped between students, or replace the two students with a single student that sees both augmentation families at the same capacity; if Dice does not drop appreciably, the dual-student disentanglement is not what produces the reported result.

Watch

Extended reading notes

Core claim

The paper's central claim is that a Mean Teacher framework with two specialized students—one trained only on stain-augmented inputs (color jittering and histogram matching) and one trained only on structurally deformed inputs (elastic deformation)—learns cleaner disentangled representations than a single student trained holistically. A shared teacher, updated by exponential moving average of both students' weights, generates pseudo-labels, and the unsupervised loss is weighted by entropy-based uncertainty maps modulated by chromatically diverse and edge-rich regions. With a DeepLabV3+ ResNet-101 backbone, CSDS achieves 82.86 Dice on GlaS and 75.25 Dice on CRAG at 5% labeled data, and 85.06 Dice on GlaS and 79.50 Dice on CRAG at 10%, consistently above all compared baselines.

Load-bearing premise

The framework assumes color augmentation changes stain appearance without touching tissue structure, and elastic deformation changes structure without touching stain; if the two augmentation streams mix these factors, the two students learn redundant views and the reported gains may come from extra parameters and augmentation diversity rather than from disentanglement.

Editorial extensions

If this is right

At 5% labeled data, CSDS reaches 82.86 Dice on GlaS, leaving only a 0.81-point gap to full supervision, so near-fully-supervised quality may be attainable at a fraction of the annotation cost.
Removing either the color or structure student lowers Dice on both label ratios, indicating the two branches carry complementary information rather than one being redundant.
Averaging both students into the EMA teacher outperforms alternating or best-student updates, so aggregating both views is itself part of the improvement.
Because the same DeepLabV3+/ResNet-101 backbone is used throughout, the reported gains come from the training strategy rather than from a stronger segmentation architecture.

Reading between the lines

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

Editorial inference: since the unsupervised loss in Eq. 12 imposes no explicit independence constraint between the two students, the gains could partly reflect an ensemble of two augmentation views; adding a decorrelation penalty would test whether true disentanglement is being learned.
Editorial inference: swapping the augmentation pipelines—color student on elastically deformed images and structure student on color-jittered images—is an untreated control that would reveal whether specialization or augmentation diversity is responsible for the performance.
Editorial inference: training one student on both augmentation families with twice the capacity would tell whether the two-network design is necessary or whether a single model with richer augmentation suffices.
Editorial inference: if the disentanglement generalizes, the same color-structure split could be carried to other stain-variable domains such as immunohistochemistry or frozen sections, and to instance-level tasks like nuclei detection and counting.

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.

Desk Editor's Note

read the letter

Colleague,

Quick take: this is a sensible extension of teacher-student SSL to histopathology, with a dual-student split along color and structure augmentations. The reported gains are real-ish—about a Dice point on two public benchmarks—but the headline claim of 'explicit decoupling' is not enforced by the loss. The method works, or at least appears to, for reasons that could be just extra parameters and augmentation diversity.

What's new: the specific combination of a color student (jittering/histogram matching), a structure student (elastic deformation), and entropy-based uncertainty masks modulated by color variance and edge maps. I don't know of an exact prior version of this. The ablations (Tables 2 and 3) show both branches contribute and the mean EMA strategy helps. Code is promised, which is good.

Soft spots, in order of importance:

1. Disentanglement is asserted, not demonstrated. Eq. 12 gives each student its own augmented view but has no independence term; both students share the same supervised loss and the same teacher's pseudo-labels. Nothing stops the color student from learning structure or the structure student from learning stain. The stress-test note is right: the gains could come from having two networks and stronger augmentation coverage. To support the 'disentangled' wording, you'd need to show representations specialize, e.g., via intervention or representation-similarity analysis.

2. Several numeric inconsistencies. Section 3.4 swaps the 5% and 10% GlaS Dice values (text says 10% is 82.86; table says 85.06). The abstract's '1.4% improvement' on CRAG at 10% doesn't match the table (gap vs FDCL is 3.22 points). The claim of 'narrowing the gap to full supervision to just 0.81%' at 5% is off by a factor of ten (actual gap ~7.98). These undermine trust.

3. Missing details: thresholds tau_color, tau_structure, lambda_C, lambda_S, lambda_unsup, and EMA decay are not reported. That's a reproducibility blocker.

4. No significance tests across the five folds. Many numbers overlap within one standard deviation; 'consistently outperforms' is a strong reading.

5. FDCL, the strongest baseline, shares authors with this paper but this is not disclosed. That's a conflict-of-interest flag that should be stated.

None of this kills the method. The core idea is credible and the experiments are on public data with fixed splits. But the paper overclaims, and the errors need correction.

Who this is for: anyone doing SSL for histopathology or medical imaging with stain variation. It's a useful incremental data point, not a conceptual breakthrough.

Recommendation: accept for peer review only after the authors fix the inconsistencies, report all hyperparameters, disclose the FDCL relationship, and either soften the disentanglement language or provide direct evidence of specialization. I'd give it a borderline accept at a workshop, not a top conference.

Referee Report

3 major / 6 minor

Summary. The paper proposes Color-Structure Dual-Student (CSDS), a semi-supervised segmentation framework for histopathology images. CSDS uses two student networks, one trained with color augmentations (color jittering, histogram matching) and one with elastic deformations, plus a shared EMA-updated teacher that provides pseudo-labels. Uncertainty maps based on entropy, color variance, and edge magnitude modulate the unsupervised losses. Experiments on GlaS and CRAG at 5% and 10% labeled-data ratios report improved Dice and Jaccard scores over eight SSL baselines, with ablation studies on the dual-student design and EMA update strategies.

Significance. If the central claim of explicit stain/structure disentanglement were validated, CSDS would be a practically useful contribution to semi-supervised histopathology segmentation, especially in low-label regimes. The paper ships code and pre-trained models, uses two public benchmarks, and compares against eight baselines, which is a strength. However, the claimed disentanglement is not enforced by the training objective, and the reported quantitative gains are accompanied by numerical inconsistencies and a lack of statistical significance testing. The significance is therefore conditional on additional evidence that the two student branches actually specialize in stain versus structure, rather than merely providing extra parameters and augmentation diversity.

major comments (3)

[Section 2.3, Eq. (12)] The central claim that CSDS 'explicitly decouples' stain and structural information is not supported by the training objective. No term in L_total penalizes redundancy or enforces independence between the two student representations; the only difference is the augmentation pipeline applied to each student. Both students are trained with the same supervised loss (Eq. 11) and receive pseudo-labels from the same EMA teacher. The reported gains are therefore equally consistent with a simpler explanation: two students with an EMA teacher and diverse augmentations outperform single-student baselines. The authors should supply direct evidence of disentanglement, for example by measuring representation similarity (e.g., CCA/SVCCA) between the two students, probing each student's sensitivity to stain versus structure perturbations, or ablating with both students receiving both augmentation types to isolate augmentation diversity.
[Section 3.4 and Table 1] The text states that 'with only 10% labeled data, CSDS achieves a Dice score of 82.86±1.24', but Table 1 reports 85.06±0.57 for the 10% GlaS setting; 82.86±1.24 is the 5% value. This is a load-bearing misreporting of the headline result. Additionally, the abstract's claim of a 1.4% improvement on CRAG at 10% labeled data is inconsistent with Table 1, where CSDS (79.50) exceeds the strongest baseline FDCL (76.28) by 3.22 Dice points. Please clarify the exact comparison points and correct the numbers.
[Table 1, Section 3.3] No statistical significance tests are reported. Several of the claimed improvements overlap with baseline standard deviations; for example, at 5% labeled GlaS, CSDS (82.86±1.24) overlaps with FDCL (81.64±1.08) and CCVC (80.84±1.75) within one standard deviation. Reporting mean ± std over five folds is insufficient to support the claim of consistent state-of-the-art improvement. The authors should provide paired tests across the five folds (e.g., paired t-test or Wilcoxon signed-rank test) or multiple seeds, and report per-fold differences.

minor comments (6)

[Table 1] The CSDS row uses European decimal commas ('75,25±1,54, 63,00±1,67') while all other rows use decimal points; this is inconsistent and likely to cause reader confusion.
[Section 2.2.2, Eq. (4)] The notation 'G3×3' is introduced as a Gaussian filter but the equation says it can be approximated via average pooling; please specify the kernel size, sigma, and which approximation is used in the experiments.
[Section 3.2] The hyperparameters tau_color, tau_structure, lambda_C, lambda_S, lambda_unsup, and the EMA decay rate are not reported. Without these values, the uncertainty modules and training procedure are not fully reproducible from the manuscript, although the code link helps.
[Section 2.3, Eq. (11)] The notation L_CE+Dice is used without defining whether the cross-entropy and Dice losses are summed or averaged; please define the exact combination.
[Section 3.4] There is a typo: 'Glas dataset' should be 'GlaS dataset'.
[Section 3.5, Figure 2] The qualitative description of 'cleaner, more accurate masks' is not quantified; consider adding boundary-based metrics (e.g., Hausdorff distance) or error maps to support the visual claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CSDS pipeline's reported gains are external benchmark comparisons, and none of its design choices reduce to its own outputs or to load-bearing self-citations.

full rationale

CSDS is evaluated against eight independent SSL baselines on the public GlaS and CRAG test splits, so the central quantitative claim (state-of-the-art Dice/Jaccard under 5% and 10% labels) is externally anchored rather than derived from the method's own assumptions. The disentanglement claim rests on the architectural choice of two students receiving color-augmented vs. elastically-deformed inputs (Eqs. 11-12), but this is a stated design intuition, not a parameter fitted to the test labels; no predicted quantity is algebraically equal to an input by construction. The uncertainty weights, thresholds, and loss coefficients are hand-chosen and mostly unreported, which is a reproducibility concern but not circularity, since they are not optimized against the test sets. Several references involve overlapping authors (Le et al. 2025; Pham et al. 2025; Nguyen et al. 2025), but they are cited only as examples of prior teacher-student/consistency methods and as a comparison baseline; they do not supply a load-bearing premise or a forced uniqueness argument. The paper's limitation that no explicit independence penalty enforces disentanglement is a correctness/evidence gap, not a circularity, because the paper does not claim to derive the disentanglement from a fitted identity. Honest non-finding: no circular step is exhibited, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central empirical claim relies on four domain assumptions typical of SSL, plus a set of unreported hyperparameters. No invented entities are introduced. The two-student disentanglement premise is the most load-bearing, since no constraint in the loss enforces it.

free parameters (6)

tau_color
Threshold for the binary color-diversity mask in Eq. 5; chosen by hand, value not reported.
tau_structure
Threshold for the edge magnitude mask in Eq. 9; chosen by hand, value not reported.
lambda_C
Modulation strength for color uncertainty in Eq. 6; range [0,1], value not reported.
lambda_S
Modulation strength for structure uncertainty in Eq. 10; range [0,1], value not reported.
lambda_unsup
Weight of the unsupervised loss in Eq. 13; value not reported.
EMA decay
Teacher update uses an exponential moving average, but the decay coefficient is not reported.

assumptions (4)

domain assumption Color jittering and histogram matching expose the color student only to stain variation, while elastic deformation exposes the structure student only to shape variation.
This is the load-bearing disentanglement premise; no objective verifies it. Invoked in Section 2.1 and used in Eq. 12.
domain assumption Unlabeled images are drawn from the same distribution as labeled images, so teacher pseudo-labels are informative.
Standard semi-supervised consistency assumption; required for Eq. 12 to transfer labels.
domain assumption Shannon entropy of teacher softmax probabilities is a valid per-pixel uncertainty measure that correlates with labeling error.
Equations 1-2 and the uncertainty modulation in Eqs. 6 and 10 depend on this; not validated in the paper.
domain assumption EMA-averaging both student weights produces teacher pseudo-labels that are better and more stable than either student alone.
Used in Section 2.3 and evaluated only by ablation in Table 3, not theoretically grounded.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning Disentangled Stain and Structural Representations for Semi-Supervised Histopathology Segmentation." pith.science (2026). https://pith.science/paper/KZYPPBEC

@misc{pith2026250703923,
  author       = {Pith},
  title        = {Pith review of: Learning Disentangled Stain and Structural Representations for Semi-Supervised Histopathology Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KZYPPBEC}},
  note         = {Machine review of arXiv:2507.03923}
}

read the original abstract

Accurate gland segmentation in histopathology images is essential for cancer diagnosis and prognosis. However, significant variability in Hematoxylin and Eosin (H&E) staining and tissue morphology, combined with limited annotated data, poses major challenges for automated segmentation. To address this, we propose Color-Structure Dual-Student (CSDS), a novel semi-supervised segmentation framework designed to learn disentangled representations of stain appearance and tissue structure. CSDS comprises two specialized student networks: one trained on stain-augmented inputs to model chromatic variation, and the other on structure-augmented inputs to capture morphological cues. A shared teacher network, updated via Exponential Moving Average (EMA), supervises both students through pseudo-labels. To further improve label reliability, we introduce stain-aware and structure-aware uncertainty estimation modules that adaptively modulate the contribution of each student during training. Experiments on the GlaS and CRAG datasets show that CSDS achieves state-of-the-art performance in low-label settings, with Dice score improvements of up to 1.2% on GlaS and 0.7% on CRAG at 5% labeled data, and 0.7% and 1.4% at 10%. Our code and pre-trained models are available at https://github.com/hieuphamha19/CSDS.

Figures

Figures reproduced from arXiv: 2507.03923 by the authors.

**Figure 1.** Overview of our Color-Structure Dual-Student (CSDS) framework with dual students for color and structure, and a shared EMA-updated teacher. Output uncertainty maps are used to adaptively weight pseudo-labels during training. Overall Pipeline. We propose a Color-Structure Dual-Student (CSDS) framework, an extension of the Mean Teacher paradigm Tarvainen and Valpola (2017). Our framework is built upon the intuition t… view at source ↗

**Figure 2.** Qualitative results for different semi-supervised methods on 10% labeled data from two datasets. Rows 1 to 3 correspond to GlaS, while Rows 4 to 5 correspond to CRAG. The red boxes emphasize the differences in the results. 3.5 Qualitative Results [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 16 canonical work pages

[1]

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. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2613--2622, 2021

work page 2021
[2]

Tailoring automated data augmentation to h&e-stained histopathology

Khrystyna Faryna, Jeroen van der Laak, and Geert Litjens. Tailoring automated data augmentation to h&e-stained histopathology. In Mattias Heinrich, Qi Dou, Marleen de Bruijne, Jan Lellmann, Alexander Schläfer, and Floris Ernst, editors, Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, volume 143 of Proceedings of Machine Learnin...

work page 2021
[3]

Albertson, Benjamin J

Alessandro Ferrero, Elham Ghelichkhan, Hamid Manoochehri, Man Minh Ho, Daniel J. Albertson, Benjamin J. Brintz, Tolga Tasdizen, Ross T. Whitaker, and Beatrice S. Knudsen. Histoem: A pathologist-guided and explainable workflow using histogram embedding for gland classification. Modern Pathology, 37 0 (4): 0 100447, 2024. ISSN 0893-3952

work page 2024
[4]

Mild-net: Minimal information loss dilated network for gland instance segmentation in colon histology images

Simon Graham, Hao Chen, Jevgenij Gamper, Qi Dou, Pheng-Ann Heng, David Snead, Yee Wah Tsang, and Nasir Rajpoot. Mild-net: Minimal information loss dilated network for gland instance segmentation in colon histology images. Medical Image Analysis, 52: 0 199--211, 2019. ISSN 1361-8415

work page 2019
[5]

Tran Quoc Khanh Le, Nguye Lan Vi Vu, Ha-Hieu Pham, Xuan-Loc Huynh, Tien-Huy Nguyen, Minh Huu Nhat Le, Quan Nguyen, and Hien D. Nguyen. Hdc: Hierarchical distillation for multi-level noisy consistency in semi-supervised fetal ultrasound segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, pages 5322--5331...

work page 2025
[6]

Semi-supervised medical image segmentation via cross teaching between cnn and transformer

Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In Ender Konukoglu, Bjoern Menze, Archana Venkataraman, Christian Baumgartner, Qi Dou, and Shadi Albarqouni, editors, Proceedings of The 5th International Conference on Medical Imaging with Deep Lear...

work page 2022
[7]

Semi-supervised histopathology image segmentation with feature diversified collaborative learning

Thanh-Huy Nguyen, Nguyen Lan Vi Vu, Hoang-Thien Nguyen, Quang-Vinh Dinh, Xingjian Li, and Min Xu. Semi-supervised histopathology image segmentation with feature diversified collaborative learning. In Junde Wu, Jiayuan Zhu, Min Xu, and Yueming Jin, editors, Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare, volume 281 of Procee...

work page 2025
[8]

Fetal-bcp: Addressing empirical distribution gap in semi-supervised fetal ultrasound segmentation

Ha-Hieu Pham, Le Tran Quoc Khanh, Hoang-Thien Nguyen, Nguyen Lan Vi Vu, Quang-Vinh Dinh, Thanh-Huy Nguyen, Xingjian Li, and Min Xu. Fetal-bcp: Addressing empirical distribution gap in semi-supervised fetal ultrasound segmentation. In 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1--4, 2025. doi:10.1109/ISBI60581.2025.10980925

arXiv 2025

Show all 18 references

[9]

C. E. Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27 0 (3): 0 379--423, 1948

1948
[10]

Pluim, Hao Chen, Xiaojuan Qi, Pheng-Ann Heng, Yun Bo Guo, Li Yang Wang, Bogdan J

Korsuk Sirinukunwattana, Josien P.W. Pluim, Hao Chen, Xiaojuan Qi, Pheng-Ann Heng, Yun Bo Guo, Li Yang Wang, Bogdan J. Matuszewski, Elia Bruni, Urko Sanchez, Anton Böhm, Olaf Ronneberger, Bassem Ben Cheikh, Daniel Racoceanu, Philipp Kainz, Michael Pfeiffer, Martin Urschler, Da...

2017
[11]

Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel

Kihyuk Sohn, David Berthelot, Chun-Liang Li, Zizhao Zhang, Nicholas Carlini, Ekin D. Cubuk, Alex Kurakin, Han Zhang, and Colin Raffel. Fixmatch: simplifying semi-supervised learning with consistency and confidence. In Proceedings of the 34th International Conference on Neural ...

2020
[12]

Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation

Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3097--3107, 2024

2024
[13]

Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results

Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 1195–1204, Re...

2017
[14]

Conflict-based cross-view consistency for semi-supervised semantic segmentation

Zicheng Wang, Zhen Zhao, Xiaoxia Xing, Dong Xu, Xiangyu Kong, and Luping Zhou. Conflict-based cross-view consistency for semi-supervised semantic segmentation. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19585--19595, 2023

2023
[15]

Revisiting representation learning of color information: Color medical image segmentation incorporating quaternion

Bicheng Xia, Bangcheng Zhan, Mingkui Shen, and Hejun Yang. Revisiting representation learning of color information: Color medical image segmentation incorporating quaternion. Knowledge-Based Systems, 306: 0 112707, 2024. ISSN 0950-7051

2024
[16]

Structural uncertainty estimation for medical image segmentation

Bing Yang, Xiaoqing Zhang, Huihong Zhang, Sanqian Li, Risa Higashita, and Jiang Liu. Structural uncertainty estimation for medical image segmentation. Medical Image Analysis, 103: 0 103602, 2025. ISSN 1361-8415

2025
[17]

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

Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Dinggang Shen, Tianming Liu, Terry M. Peters, Lawrence H. Staib, Caroline Essert, Sean Zhou, Pew-Thian Yap, and Ali...

2019
[18]

A review of uncertainty estimation and its application in medical imaging

Ke Zou, Zhihao Chen, Xuedong Yuan, Xiaojing Shen, Meng Wang, and Huazhu Fu. A review of uncertainty estimation and its application in medical imaging. Meta-Radiology, 1 0 (1): 0 100003, 2023. ISSN 2950-1628

2023

Pith tools

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

[1] [1]

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. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2613--2622, 2021

work page 2021

[2] [2]

Tailoring automated data augmentation to h&e-stained histopathology

Khrystyna Faryna, Jeroen van der Laak, and Geert Litjens. Tailoring automated data augmentation to h&e-stained histopathology. In Mattias Heinrich, Qi Dou, Marleen de Bruijne, Jan Lellmann, Alexander Schläfer, and Floris Ernst, editors, Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, volume 143 of Proceedings of Machine Learnin...

work page 2021

[3] [3]

Albertson, Benjamin J

Alessandro Ferrero, Elham Ghelichkhan, Hamid Manoochehri, Man Minh Ho, Daniel J. Albertson, Benjamin J. Brintz, Tolga Tasdizen, Ross T. Whitaker, and Beatrice S. Knudsen. Histoem: A pathologist-guided and explainable workflow using histogram embedding for gland classification. Modern Pathology, 37 0 (4): 0 100447, 2024. ISSN 0893-3952

work page 2024

[4] [4]

Mild-net: Minimal information loss dilated network for gland instance segmentation in colon histology images

Simon Graham, Hao Chen, Jevgenij Gamper, Qi Dou, Pheng-Ann Heng, David Snead, Yee Wah Tsang, and Nasir Rajpoot. Mild-net: Minimal information loss dilated network for gland instance segmentation in colon histology images. Medical Image Analysis, 52: 0 199--211, 2019. ISSN 1361-8415

work page 2019

[5] [5]

Tran Quoc Khanh Le, Nguye Lan Vi Vu, Ha-Hieu Pham, Xuan-Loc Huynh, Tien-Huy Nguyen, Minh Huu Nhat Le, Quan Nguyen, and Hien D. Nguyen. Hdc: Hierarchical distillation for multi-level noisy consistency in semi-supervised fetal ultrasound segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, pages 5322--5331...

work page 2025

[6] [6]

Semi-supervised medical image segmentation via cross teaching between cnn and transformer

Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In Ender Konukoglu, Bjoern Menze, Archana Venkataraman, Christian Baumgartner, Qi Dou, and Shadi Albarqouni, editors, Proceedings of The 5th International Conference on Medical Imaging with Deep Lear...

work page 2022

[7] [7]

Semi-supervised histopathology image segmentation with feature diversified collaborative learning

Thanh-Huy Nguyen, Nguyen Lan Vi Vu, Hoang-Thien Nguyen, Quang-Vinh Dinh, Xingjian Li, and Min Xu. Semi-supervised histopathology image segmentation with feature diversified collaborative learning. In Junde Wu, Jiayuan Zhu, Min Xu, and Yueming Jin, editors, Proceedings of The First AAAI Bridge Program on AI for Medicine and Healthcare, volume 281 of Procee...

work page 2025

[8] [8]

Fetal-bcp: Addressing empirical distribution gap in semi-supervised fetal ultrasound segmentation

Ha-Hieu Pham, Le Tran Quoc Khanh, Hoang-Thien Nguyen, Nguyen Lan Vi Vu, Quang-Vinh Dinh, Thanh-Huy Nguyen, Xingjian Li, and Min Xu. Fetal-bcp: Addressing empirical distribution gap in semi-supervised fetal ultrasound segmentation. In 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1--4, 2025. doi:10.1109/ISBI60581.2025.10980925

arXiv 2025