pith. sign in

arxiv: 2510.16450 · v2 · submitted 2025-10-18 · 💻 cs.CV

Instance-Aware Pseudo-Labeling and Class-Focused Contrastive Learning for Weakly Supervised Domain Adaptive Segmentation of Electron Microscopy

Shan Xiong , Jiabao Chen , Ye Wang , Jialin Peng This is my paper

Pith reviewed 2026-05-18 06:22 UTC · model grok-4.3

classification 💻 cs.CV
keywords weakly supervised domain adaptationelectron microscopymitochondria segmentationinstance-aware pseudo-labelingcontrastive learningcenter detectionself-training
0
0 comments X

The pith

Detection-guided pseudo-labels improve EM domain-adaptive segmentation

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

The paper develops a weakly supervised domain adaptation method for segmenting mitochondria in various electron microscopy images using only sparse point labels on the target domain. It employs a multitask framework that performs segmentation and center detection simultaneously, with cross-teaching between them and class-focused cross-domain contrastive learning. The core innovation is the instance-aware pseudo-label selection strategy that uses detection to choose reliable and diverse pseudo-labels from unlabeled regions for self-training, avoiding simple pixel-wise filtering. Validations show this outperforms existing methods and narrows the gap to fully supervised performance.

Core claim

We introduce a multitask learning framework that jointly conducts segmentation and center detection with a novel cross-teaching mechanism and class-focused cross-domain contrastive learning. We introduce segmentation self-training with a novel instance-aware pseudo-label (IPL) selection strategy. Unlike existing methods that typically rely on pixel-wise pseudo-label filtering, the IPL semantically selects reliable and diverse pseudo-labels with the help of the detection task.

What carries the argument

The instance-aware pseudo-label (IPL) selection strategy, which uses the center detection task to semantically select reliable and diverse pseudo-labels from unlabeled target domain regions.

If this is right

  • Outperforms existing UDA and WDA methods on challenging datasets.
  • Significantly narrows the performance gap with the supervised upper bound.
  • Achieves substantial improvements over other UDA techniques even without point labels.
  • Effectively utilizes incomplete and imprecise point annotations via multitask learning.

Where Pith is reading between the lines

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

  • The IPL approach could be adapted for other instance segmentation tasks in medical or scientific imaging with limited labels.
  • Multitask learning combining detection and segmentation may help in general for improving pseudo-label quality in domain adaptation.
  • This method suggests a way to reduce annotation costs in cross-domain biological image analysis.

Load-bearing premise

The instance-aware pseudo-label (IPL) selection strategy, guided by the center detection task, can reliably identify accurate and diverse pseudo-labels from unlabeled image regions in the target domain.

What would settle it

A direct comparison on target domain images where full annotations are available showing that IPL-selected pseudo-labels have high error rates or low diversity would falsify the reliability of the selection strategy.

Figures

Figures reproduced from arXiv: 2510.16450 by Jiabao Chen, Jialin Peng, Shan Xiong, Ye Wang.

Figure 1
Figure 1. Figure 1: Weakly-Supervised Domain Adaptation (WDA) for mitochondria segmentation in EM images. Sparse center points on a few object instances are utilized as the weak labels on target training data. sumption that no labels are available in the target domain, thus eliminating the annotation burden on the target domain. However, despite significant research efforts, a substantial performance gap persists between the … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed method for weakly-supervised cross-domain adaptation. Under a multitask learning framework, an auxiliary center detection task is utilized to achieve instance-aware pseudo-label selection for the self-training on the segmentation head. class-focused conservative learning through pixel-to-prototype alignment is introduced to conduct feature-level domain alignment and improve the com… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison results. Green: true positives; Red: false positives; White: false negatives [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of the ablated versions of our proposed approach. TABLE III PERFORMANCE COMPARISON USING DIFFERENT PSEUDO-LABEL SELECTION STRATEGIES FOR THE SEGMENTATION ON EPFL HIPPOCAMPUS→MITOEM-R. Pseudo-Labeling for Segmentation Dice (%) PQ (%) Threshold-based strategy 91.5 73.5 Entropy-based strategy [12] 91.6 73.6 Instance-aware based strategy (Ours) 92.6 75.6 further led to a significant performan… view at source ↗
Figure 5
Figure 5. Figure 5: Robustness to location deviations of the sparse point annotation. The performance of our model (15%) on EPFL Hippocampus→MitoEM￾R is evaluated when randomly displacing the sparse point annotations by d pixels from their respective centers. TABLE VI INFLUENCE OF THE SAMPLE DEVIATIONS OF THE SPARSE POINT ANNOTATION. THE PERFORMANCE OF OUR MODEL (15%) ON EPFL HIPPOCAMPUS→MITOEM-R WITH DIFFERENT SAMPLING WAS E… view at source ↗
read the original abstract

Annotation-efficient segmentation of the numerous mitochondria instances from various electron microscopy (EM) images is highly valuable for biological and neuroscience research. Although unsupervised domain adaptation (UDA) methods can help mitigate domain shifts and reduce the high costs of annotating each domain, they typically have relatively low performance in practical applications. Thus, we investigate weakly supervised domain adaptation (WDA) that utilizes additional sparse point labels on the target domain, which require minimal annotation effort and minimal expert knowledge. To take full use of the incomplete and imprecise point annotations, we introduce a multitask learning framework that jointly conducts segmentation and center detection with a novel cross-teaching mechanism and class-focused cross-domain contrastive learning. While leveraging unlabeled image regions is essential, we introduce segmentation self-training with a novel instance-aware pseudo-label (IPL) selection strategy. Unlike existing methods that typically rely on pixel-wise pseudo-label filtering, the IPL semantically selects reliable and diverse pseudo-labels with the help of the detection task. Comprehensive validations and comparisons on challenging datasets demonstrate that our method outperforms existing UDA and WDA methods, significantly narrowing the performance gap with the supervised upper bound. Furthermore, under the UDA setting, our method also achieves substantial improvements over other UDA techniques.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes a multitask weakly supervised domain adaptation (WDA) framework for mitochondria instance segmentation in electron microscopy (EM) images. It jointly trains segmentation and center detection heads with a cross-teaching mechanism, adds class-focused cross-domain contrastive learning, and introduces an instance-aware pseudo-label (IPL) selection strategy that uses detection outputs to choose reliable and diverse pseudo-labels from unlabeled target-domain regions instead of pixel-wise filtering. The central claim is that this approach outperforms prior UDA and WDA methods on challenging datasets while substantially narrowing the gap to a fully supervised upper bound; the method is also shown to improve results under the pure UDA setting.

Significance. If the performance claims hold under rigorous controls, the work would be significant for annotation-efficient biomedical image analysis. Sparse point labels on the target domain are far cheaper than dense masks, and the IPL strategy that couples detection with segmentation offers a concrete way to exploit unlabeled regions without the usual pitfalls of noisy pseudo-labels. The multitask cross-teaching and contrastive components are technically coherent and could generalize to other instance segmentation tasks with domain shift.

major comments (2)
  1. [Method (IPL selection) and Experiments] The central performance claims rest on the IPL selection strategy (described in the method section and abstract). The paper provides no quantitative evaluation of center-detection accuracy on target-domain instances (e.g., center localization error, precision/recall of detected centers, or failure cases for small/dense mitochondria). Without such analysis it is impossible to verify that the auxiliary detection head remains reliable enough under domain shift to avoid systematically biased pseudo-label selection.
  2. [Abstract and Experimental Results] The abstract asserts that the method 'significantly narrow[s] the performance gap with the supervised upper bound' and outperforms existing UDA/WDA methods, yet the provided text contains no details on the exact datasets, train/test splits, evaluation metrics (Dice, AJI, etc.), number of runs, or statistical significance tests. These omissions make it impossible to assess whether the reported gains are robust or sensitive to post-hoc hyper-parameter choices.
minor comments (2)
  1. [Method] Notation for the cross-teaching loss and the class-focused contrastive term should be introduced with explicit equations and variable definitions in the method section to improve readability.
  2. [Figures] Figure captions and axis labels in the qualitative results should explicitly state which domain (source/target) and which method each panel corresponds to.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. The comments highlight important aspects of the IPL strategy and experimental reporting that we have addressed in the revision.

read point-by-point responses

  1. Referee: [Method (IPL selection) and Experiments] The central performance claims rest on the IPL selection strategy (described in the method section and abstract). The paper provides no quantitative evaluation of center-detection accuracy on target-domain instances (e.g., center localization error, precision/recall of detected centers, or failure cases for small/dense mitochondria). Without such analysis it is impossible to verify that the auxiliary detection head remains reliable enough under domain shift to avoid systematically biased pseudo-label selection.

    Authors: We agree that quantitative analysis of the center-detection head on the target domain would provide direct evidence for the reliability of IPL selection. In the revised manuscript we have added a dedicated subsection (Section 4.4) reporting center localization error (in pixels), precision/recall of detected centers, and qualitative discussion of failure cases on small or densely packed mitochondria. These results, obtained under the same domain-shift conditions as the main experiments, show that the cross-teaching mechanism keeps detection accuracy sufficiently high to avoid systematic bias in pseudo-label selection. revision: yes

  2. Referee: [Abstract and Experimental Results] The abstract asserts that the method 'significantly narrow[s] the performance gap with the supervised upper bound' and outperforms existing UDA/WDA methods, yet the provided text contains no details on the exact datasets, train/test splits, evaluation metrics (Dice, AJI, etc.), number of runs, or statistical significance tests. These omissions make it impossible to assess whether the reported gains are robust or sensitive to post-hoc hyper-parameter choices.

    Authors: We acknowledge the omissions. The revised abstract now explicitly states the datasets (MitoEM and the additional EM volumes), train/test splits, metrics (Dice, AJI, and PQ), and that all results are averaged over three independent runs with standard deviation. In the experimental section we have added a new table summarizing mean and std across runs together with paired t-test p-values against the strongest baselines, confirming that the reported improvements are statistically significant and not sensitive to post-hoc choices. revision: yes

Circularity Check

0 steps flagged

Minor self-citation in related work but central multitask framework and IPL strategy are independently proposed and empirically validated

full rationale

The paper introduces a multitask segmentation-plus-center-detection framework with cross-teaching and a novel instance-aware pseudo-label selection strategy that uses detection outputs to filter reliable pseudo-labels. Performance claims rest on experimental comparisons against UDA/WDA baselines on EM datasets rather than any mathematical derivation that loops back to fitted parameters or self-defined quantities. No equations reduce predictions to inputs by construction, and the method is presented as a set of algorithmic choices whose effectiveness is tested externally. This is a standard empirical contribution with at most incidental self-citation that does not bear the load of the core claims.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review is based solely on the abstract; full details on any hyperparameters, modeling choices, or background assumptions are unavailable. The approach relies on standard deep learning assumptions plus the novel IPL mechanism.

axioms (2)
  • domain assumption Joint training of segmentation and center detection tasks via cross-teaching produces mutual performance gains.
    The multitask framework depends on this interaction being effective for the target domain.
  • ad hoc to paper Instance-aware pseudo-label selection using detection outputs yields more reliable and diverse labels than pixel-wise filtering.
    This is the core novel IPL strategy introduced in the abstract.

pith-pipeline@v0.9.0 · 5762 in / 1401 out tokens · 40663 ms · 2026-05-18T06:22:56.786713+00:00 · methodology

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

49 extracted references · 49 canonical work pages

  1. [1]

    This study develops an effective WDA method that sig- nificantly outperforms UDA methods and demonstrates comparable performance to supervised methods with minimal annotation efforts and knowledge

    work page

  2. [2]

    Rather than using pixel-wise pseudo-label selection, this study proposes a simple yet effective instance-aware JOURNALS TEMPLATE 3 Fig. 2. Overview of the proposed method for weakly-supervised cross-domain adaptation. Under a multitask learning framework, an auxiliary center detection task is utilized to achieve instance-aware pseudo-label selection for t...

    work page

  3. [3]

    A class-focused contrastive learning approach has been introduced to effectively learn domain-invariant features. II. RELATEDWORK A. Unsupervised Domain Adaptation Deep learning models trained on a specific domain often suffer from significant performance degradation when tested on datasets with shifted distributions. Although foundation models like the S...

    work page

  4. [4]

    The goal of the WDA is to learn a segmentation model that adapts well to the target domain

    is the kernel bandwidth. The goal of the WDA is to learn a segmentation model that adapts well to the target domain. Model overview. An overview of our model is demon- strated in Fig. 2, in which we conduct multitask learning with a cross-task teaching mechanism. Our model takes an encoder- decoder architecturef D ◦f E with a segmentation headf S and a re...

    work page

  5. [5]

    This dataset was split into two subsets, each of which contains 165 image slices of size 768×1,024, for training and testing, respectively

    in a resolution of 5×5×5nm 3, representing tissues from the mouse CA1 hippocampus region. This dataset was split into two subsets, each of which contains 165 image slices of size 768×1,024, for training and testing, respectively. MitoEM-R Cortex Data. The MitoEM Dataset [13] con- strains two subsets. The MitoEM-R subset was scanned using a multi-beam scan...

    work page

  6. [6]

    are 2.5D methods that take multiple slices as input, while others, including our model, simply conduct 2D segmentation. We also compare our method with typical foundation models, including SAM [27], Med-SA [28], and their interactive ver- sions, i.e., SAM (Interact), and Med-SA (Interact), which take center-points of all mitochondria instances as user pro...

    work page

  7. [7]

    By comparing our full model with Model III in Table II, we have noticed a performance drop of 1.5% in the Dice coefficient when class-focused contrastive learning was removed

    auxiliary detection with pseudo-labeling, 2) instance-aware pseudo-labeling for segmentation, 3) class-focused contrastive learning. By comparing our full model with Model III in Table II, we have noticed a performance drop of 1.5% in the Dice coefficient when class-focused contrastive learning was removed. By further removing instance-aware pseudo-labeli...

    work page

  8. [8]

    Segmen- tation in large-scale cellular electron microscopy with deep learning: A literature survey,

    A. Aswath, A. Alsahaf, B. N. Giepmans, and G. Azzopardi, “Segmen- tation in large-scale cellular electron microscopy with deep learning: A literature survey,”Medical Image Analysis, p. 102920, 2023

    work page 2023

  9. [9]

    Learning for structured prediction using approximate subgradient descent with working sets,

    A. Lucchi, Y . Li, and P. Fua, “Learning for structured prediction using approximate subgradient descent with working sets,” inProceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2013, pp. 1987–1994

    work page 2013

  10. [10]

    Call to action to properly utilize electron microscopy to measure organelles to monitor disease,

    K. Neikirk, E.-G. Lopez, A. G. Marshall, A. Alghanem, E. Krystofiak, B. Kula, N. Smith, J. Shao, P. Katti, and A. O. Hinton Jr, “Call to action to properly utilize electron microscopy to measure organelles to monitor disease,”European Journal of Cell Biology, p. 151365, 2023

    work page 2023

  11. [11]

    Mitochondria in disease: changes in shapes and dynamics,

    B. C. Jenkins, K. Neikirk, P. Katti, S. M. Claypool, A. Kirabo, M. R. McReynolds, and A. Hinton, “Mitochondria in disease: changes in shapes and dynamics,”Trends in Biochemical Sciences, 2024

    work page 2024

  12. [12]

    Fear memory-associated synaptic and mitochondrial changes revealed by deep learning-based processing of electron mi- croscopy data,

    J. Liu, J. Qi, X. Chen, Z. Li, B. Hong, H. Ma, G. Li, L. Shen, D. Liu, Y . Konget al., “Fear memory-associated synaptic and mitochondrial changes revealed by deep learning-based processing of electron mi- croscopy data,”Cell Reports, vol. 40, no. 5, 2022

    work page 2022

  13. [13]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 2015, pp. 234–241

    work page 2015

  14. [14]

    Cs-net: Instance-aware cellular segmentation with hierarchical dimension-decomposed convolutions and slice-attentive learning,

    J. Peng and Z. Luo, “Cs-net: Instance-aware cellular segmentation with hierarchical dimension-decomposed convolutions and slice-attentive learning,”Knowledge-Based Systems, vol. 232, p. 107485, 2021

    work page 2021

  15. [15]

    Adaptive template transformer for mitochondria segmentation in elec- tron microscopy images,

    Y . Pan, N. Luo, R. Sun, M. Meng, T. Zhang, Z. Xiong, and Y . Zhang, “Adaptive template transformer for mitochondria segmentation in elec- tron microscopy images,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 21 474–21 484

    work page 2023

  16. [16]

    A review of deep learning in medical imaging: Imaging traits, technol- ogy trends, case studies with progress highlights, and future promises,

    S. K. Zhou, H. Greenspan, C. Davatzikos, J. S. Duncan, B. Van Gin- neken, A. Madabhushi, J. L. Prince, D. Rueckert, and R. M. Summers, “A review of deep learning in medical imaging: Imaging traits, technol- ogy trends, case studies with progress highlights, and future promises,” Proceedings of the IEEE, vol. 109, no. 5, pp. 820–838, 2021

    work page 2021

  17. [17]

    Mask rearranging data augmentation for 3d mitochondria segmentation,

    Q. Chen, M. Li, J. Li, B. Hu, and Z. Xiong, “Mask rearranging data augmentation for 3d mitochondria segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Inter- vention. Springer, 2022, pp. 36–46

    work page 2022

  18. [18]

    Evidential uncertainty-guided mitochondria segmentation for 3d em images,

    R. Shi, L. Duan, T. Huang, and T. Jiang, “Evidential uncertainty-guided mitochondria segmentation for 3d em images,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 5, 2024, pp. 4847–4855

    work page 2024

  19. [19]

    Weakly-supervised cross-domain segmentation of electron microscopy with sparse point annotation,

    D. Qiu, S. Xiong, J. Yi, and J. Peng, “Weakly-supervised cross-domain segmentation of electron microscopy with sparse point annotation,” IEEE Transactions on Big Data, 2024

    work page 2024

  20. [20]

    Mitoem dataset: Large-scale 3d mitochondria instance segmentation from em images,

    D. Wei, Z. Lin, D. Franco-Barranco, N. Wendt, X. Liu, W. Yin, X. Huang, A. Gupta, W.-D. Jang, X. Wanget al., “Mitoem dataset: Large-scale 3d mitochondria instance segmentation from em images,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2020, pp. 66–76

    work page 2020

  21. [21]

    A theory of learning from different domains,

    S. Ben-David, J. Blitzer, K. Crammer, A. Kulesza, F. Pereira, and J. W. Vaughan, “A theory of learning from different domains,”Machine Learning, vol. 79, pp. 151–175, 2010

    work page 2010

  22. [22]

    Generalized out-of-distribution detection: A survey,

    J. Yang, K. Zhou, Y . Li, and Z. Liu, “Generalized out-of-distribution detection: A survey,”International Journal of Computer Vision, vol. 132, no. 12, pp. 5635–5662, 2024

    work page 2024

  23. [23]

    Mitochondrial heterogeneity and home- ostasis through the lens of a neuron,

    G. Pekkurnaz and X. Wang, “Mitochondrial heterogeneity and home- ostasis through the lens of a neuron,”Nature Metabolism, vol. 4, no. 7, pp. 802–812, 2022

    work page 2022

  24. [24]

    Domain adaptation for medical image analysis: a survey,

    H. Guan and M. Liu, “Domain adaptation for medical image analysis: a survey,”IEEE Transactions on Biomedical Engineering, vol. 69, no. 3, pp. 1173–1185, 2021

    work page 2021

  25. [25]

    Unsupervised mitochondria segmentation in em images via domain adaptive multi-task learning,

    J. Peng, J. Yi, and Z. Yuan, “Unsupervised mitochondria segmentation in em images via domain adaptive multi-task learning,”IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 6, pp. 1199–1209, 2020

    work page 2020

  26. [26]

    Self-supervised augmentation consistency for adapting semantic segmentation,

    N. Araslanov and S. Roth, “Self-supervised augmentation consistency for adapting semantic segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 15 384–15 394

    work page 2021

  27. [27]

    Uncertainty-aware label rectification for domain adaptive mitochondria segmentation,

    S. Wu, C. Chen, Z. Xiong, X. Chen, and X. Sun, “Uncertainty-aware label rectification for domain adaptive mitochondria segmentation,” in24th International Conference on Medical Image Computing and Computer Assisted Intervention. Springer, 2021, pp. 191–200

    work page 2021

  28. [28]

    Class-aware feature alignment for domain adaptative mitochondria segmentation,

    D. Yin, W. Huang, Z. Xiong, and X. Chen, “Class-aware feature alignment for domain adaptative mitochondria segmentation,” inInterna- tional Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, pp. 238–248

    work page 2023

  29. [29]

    Domain adaptive mitochondria segmentation via enforcing inter-section consistency,

    W. Huang, X. Liu, Z. Cheng, Y . Zhang, and Z. Xiong, “Domain adaptive mitochondria segmentation via enforcing inter-section consistency,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2022, pp. 89–98

    work page 2022

  30. [30]

    Wda-net: Weakly-supervised domain adap- tive segmentation of electron microscopy,

    D. Qiu, J. Yi, and J. Peng, “Wda-net: Weakly-supervised domain adap- tive segmentation of electron microscopy,” in2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2022, pp. 1132–1137

    work page 2022

  31. [31]

    Weakly-supervised domain adaptive semantic segmentation with prototypical contrastive learning,

    A. Das, Y . Xian, D. Dai, and B. Schiele, “Weakly-supervised domain adaptive semantic segmentation with prototypical contrastive learning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 15 434–15 443

    work page 2023

  32. [32]

    Bi-directional contrastive learning for domain adaptive semantic segmentation,

    G. Lee, C. Eom, W. Lee, H. Park, and B. Ham, “Bi-directional contrastive learning for domain adaptive semantic segmentation,” in European Conference on Computer Vision. Springer, 2022, pp. 38– 55

    work page 2022

  33. [33]

    Learning to adapt structured output space for semantic segmentation,

    Y . Tsai, W. Hung, S. Schulter, K. Sohn, M. Yang, and M. Chandraker, “Learning to adapt structured output space for semantic segmentation,” inIEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 7472–7481

    work page 2018

  34. [34]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026. JOURNALS TEMPLATE 11

    work page 2023

  35. [35]

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

    J. Wu, Z. Wang, M. Hong, W. Ji, H. Fu, Y . Xu, M. Xu, and Y . Jin, “Medical sam adapter: Adapting segment anything model for medical image segmentation,”Medical image analysis, vol. 102, p. 103547, 2025

    work page 2025

  36. [36]

    Learning transferable features with deep adaptation networks,

    M. Long, Y . Cao, J. Wang, and M. Jordan, “Learning transferable features with deep adaptation networks,” inInternational Conference on Machine Learning. PMLR, 2015, pp. 97–105

    work page 2015

  37. [37]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, and H. Larochelle, “Domain-adversarial training of neural networks,”Journal of Machine Learning Research, vol. 17, no. 1, pp. 2096–2030, 2016

    work page 2096

  38. [38]

    Contrast, stylize and adapt: Unsupervised contrastive learning framework for domain adaptive semantic segmentation,

    T. Li, S. Roy, H. Zhou, H. Lu, and S. Lathuili `ere, “Contrast, stylize and adapt: Unsupervised contrastive learning framework for domain adaptive semantic segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4869–4879

    work page 2023

  39. [39]

    Cycada: Cycle-consistent adversarial domain adapta- tion,

    J. Hoffman, E. Tzeng, T. Park, J.-Y . Zhu, P. Isola, K. Saenko, A. Efros, and T. Darrell, “Cycada: Cycle-consistent adversarial domain adapta- tion,” inInternational Conference on Machine Learning. PMLR, 2018, pp. 1989–1998

    work page 2018

  40. [40]

    Domain adaptation for semantic segmentation via class-balanced self-training,

    Y . Zou, Z. Yu, B. Kumar, and J. Wang, “Domain adaptation for semantic segmentation via class-balanced self-training,” inEuropean Conference on Computer Vision, 2018, pp. 289–305

    work page 2018

  41. [41]

    Domain adaptive semantic segmentation using weak labels,

    S. Paul, Y .-H. Tsai, S. Schulter, A. K. Roy-Chowdhury, and M. Chan- draker, “Domain adaptive semantic segmentation using weak labels,” in 16th European Conference on Computer Vision. Springer, 2020, pp. 571–587

    work page 2020

  42. [42]

    Dawn: Domain-adaptive weakly supervised nuclei segmentation via cross-task interactions,

    Y . Zhang, Y . Wang, Z. Fang, H. Bian, L. Cai, Z. Wang, and Y . Zhang, “Dawn: Domain-adaptive weakly supervised nuclei segmentation via cross-task interactions,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 5, pp. 4753–4767, 2024

    work page 2024

  43. [43]

    Domain adaptive box- supervised instance segmentation network for mitosis detection,

    Y . Li, Y . Xue, L. Li, X. Zhang, and X. Qian, “Domain adaptive box- supervised instance segmentation network for mitosis detection,”IEEE Transactions on Medical Imaging, vol. 41, no. 9, pp. 2469–2485, 2022

    work page 2022

  44. [44]

    Single-image crowd counting via multi-column convolutional neural network,

    Y . Zhang, D. Zhou, S. Chen, S. Gao, and Y . Ma, “Single-image crowd counting via multi-column convolutional neural network,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 589–597

    work page 2016

  45. [45]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks,

    D.-H. Leeet al., “Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks,” inWorkshop on challenges in representation learning at International Conference on Machine Learning, vol. 3, no. 2, 2013, p. 896

    work page 2013

  46. [46]

    A dataset and a technique for generalized nuclear segmen- tation for computational pathology,

    N. Kumar, R. Verma, S. Sharma, S. Bhargava, A. Vahadane, and A. Sethi, “A dataset and a technique for generalized nuclear segmen- tation for computational pathology,”IEEE Transactions on Medical Imaging, vol. 36, no. 7, pp. 1550–1560, 2017

    work page 2017

  47. [47]

    Panoptic segmentation,

    A. Kirillov, K. He, R. Girshick, C. Rother, and P. Doll ´ar, “Panoptic segmentation,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 9404–9413

    work page 2019

  48. [48]

    Understanding the behaviour of contrastive loss,

    F. Wang and H. Liu, “Understanding the behaviour of contrastive loss,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 2495–2504

    work page 2021

  49. [49]

    The genetics and pathology of mitochondrial disease,

    C. L. Alston, M. C. Rocha, N. Z. Lax, D. M. Turnbull, and R. W. Taylor, “The genetics and pathology of mitochondrial disease,”The Journal of pathology, vol. 241, no. 2, pp. 236–250, 2017

    work page 2017