Back

Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images

Huisi WuZhaoze WangYouyi SongLin YangJing Qin

article2022CVPR100 citations

Proposes a cross-patch dense contrastive learning framework that pairs patch- and pixel-level representations within a mean-teacher architecture to segment cellular nuclei accurately from limited annotated histopathology images.

Listen

Accurate segmentation of cellular nuclei in digital pathology images is essential for computer-assisted cancer diagnosis and analyzing tumor microenvironments. However, training high-performing deep learning models typically requires large volumes of pixel-level annotations from medical specialists. Acquiring these expert labels is extremely costly, labor-intensive, and prone to inter-observer variability, creating a major operational bottleneck for scaling artificial intelligence in digital pathology.

The article demonstrates a semi-supervised learning framework that achieves high segmentation accuracy using minimal labeled data paired with abundant unlabeled images. The main objective is to evaluate whether aligning representations across both image patches and individual pixels—supported by prediction-level regularization—can effectively extract structural knowledge from unannotated histological images.

The authors established a teacher-student deep learning architecture trained on two public benchmark datasets: the 2018 Data Science Bowl (DSB) and the Multi-Organ Nucleus Segmentation (MoNuSeg) dataset. The method evaluates inter-patch feature disparities to categorize regions into foreground-dominated, background-dominated, and mixed patches. Contrastive learning is applied across patches and densely across pixels, pulling similar features together and pushing dissimilar features apart. This representation learning is further paired with consistency regularization and entropy minimization to stabilize predictions and enforce classification confidence.

The experimental findings show that the proposed framework consistently outperforms existing semi-supervised approaches across all testing conditions. When evaluated in an extreme low-label setting with only 1/32 of training images annotated, the framework achieved Dice similarity coefficients of 87.49% on DSB and 75.97% on MoNuSeg, surpassing the leading baseline by over 1 percentage point on each benchmark. Visual analyses confirmed that the method produces superior boundaries, accurate counts, and fewer false positive or false negative errors compared to competing models. Furthermore, ablation experiments confirmed that combining patch-level disparity matching with consistency and entropy regularization creates a reinforcing cycle of higher-quality pseudo-labels and sharper feature representations.

These results indicate that healthcare and digital pathology organizations can significantly reduce expert data-labeling costs and project turnaround times without sacrificing model reliability. By requiring only a fraction of manual annotations, AI deployment in clinical research and diagnosis becomes substantially more feasible and cost-effective.

Based on these findings, development teams should consider adopting dual patch-and-pixel contrastive semi-supervised strategies when building medical segmentation pipelines. Before full clinical deployment, additional pilot validations are recommended across more diverse tissue types and external histopathology datasets. Finally, while confidence in the framework is high given its consistent cross-dataset performance, practitioners should note limitations in segmenting extremely small nuclei or structures with very low background contrast.

Hong Kong Polytechnic UniversityShenzhen Universitycellularcontrastive learninghigh resolution imagecross-patch dense contrastive learningcellular nuclei segmentationhistopathologic images
Wu et al (2022).pdf
Cover for Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images

Abstract

We study the semi-supervised learning problem, using a few labeled data and a large amount of unlabeled data to train the network, by developing a cross-patch dense contrastive learning framework, to segment cellular nuclei in histopathologic images. This task is motivated by the expensive burden on collecting labeled data for histopathologic image segmentation tasks. The key idea of our method is to align features of teacher and student networks, sampled from cross-image in both patch- and pixel-levels, for enforcing the intra-class compactness and inter-class separability of features that as we shown is helpful for extracting valuable knowledge from unlabeled data. We also design a novel optimization framework that combines consistency regularization and entropy minimization techniques, showing good property in eviction of gradient vanishing. We assess the proposed method on two publicly available datasets, and obtain positive results on extensive experiments, outperforming the state-of-the-art methods. Codes are available at https://github.com/zzw-szu/CDCL.

Topics

cellularcontrastive learninghigh resolution image

Keywords

cross-patch dense contrastive learningcellular nuclei segmentationhistopathologic imagessemi-supervised segmentationtumor microenvironment analysismedical image segmentationcontrastive lossinter-patch feature disparity

Table of Contents

  • 1. Introduction
  • 2. Related Work
  • 2.1. Cellular Nuclei Segmentation in Histopathologic Images
  • 2.2. Semi-supervised Semantic Segmentation
  • 2.3. Contrastive Learning
  • 3. Method
  • 3.1. Cross-patch Dense Contrastive Learning
  • 3.2. Consistency Regularization
  • 3.3. Entropy Minimization
  • 4. Experiments
  • 4.1. Datasets
  • 4.2. Implementation Details
  • 4.3. Comparison with State-of-the-art Methods
  • 4.4. Ablation Studies
  • 5. Conclusion
  • Acknowledgments
  • References

Knowls

  1. Knowl 1 — Mean-teacher semi-supervised segmentation objective

    model/method

    The proposed Cross-patch Dense Contrastive Learning (CDCL) framework uses a student network and a teacher network with the same extractor–classifier–projector architecture. The student receives strongly augmented inputs, while the teacher receives weakly augmented versions of the same unlabeled images. Labeled images train the student with pixel-wise cross-entropy; unlabeled images contribute contrastive, consistency, and entropy-minimization losses.

    The student minimizes the weighted objective

    L=wsupLsup+wcontrLcontr+wconsLcons+wentLent,L=w_{\mathrm{sup}}L_{\mathrm{sup}}+w_{\mathrm{contr}}L_{\mathrm{contr}}+w_{\mathrm{cons}}L_{\mathrm{cons}}+w_{\mathrm{ent}}L_{\mathrm{ent}},

    where LsupL_{\mathrm{sup}} is supervised cross-entropy, LcontrL_{\mathrm{contr}} is cross-patch dense contrastive loss, LconsL_{\mathrm{cons}} is teacher–student consistency loss, LentL_{\mathrm{ent}} is student entropy loss, and each ww is its loss weight. The teacher is not optimized by back-propagation; its parameters are an exponential moving average of the student parameters:

    θit=αθi−1t+(1−α)θis,\theta_i^{t}=\alpha\theta_{i-1}^{t}+(1-\alpha)\theta_i^{s},

    where θit\theta_i^{t} and θis\theta_i^{s} are teacher and student parameters at iteration ii, respectively, and α∈[0,1]\alpha\in[0,1] is the teacher update coefficient. This temporal averaging is used to obtain more stable teacher predictions and pseudo-labels.

  2. Knowl 2 — Cross-patch dense sampling of positive and negative pairs

    algorithm

    For an unlabeled image, the student and teacher produce low-dimensional projected feature maps. The image is divided into fixed-size patches, and corresponding student and teacher patches form positive patch pairs. Pixels at identical spatial positions inside corresponding patches form positive pixel pairs.

    Negative sampling is based on pseudo-label-derived patch composition rather than random pixel selection. For a patch containing NN pixels, let NfN_f be the number of pixels whose pseudo-label is the foreground class. Its foreground score is FS=Nf/NFS=N_f/N. Patches are categorized as foreground-dominant (FDP) when FS≥0.7FS\geq0.7, background-dominant (BDP) when FS≤0.3FS\leq0.3, and mixed (MP) otherwise. An FDP and a BDP are treated as a structurally dissimilar negative patch pair. All pixel features across the two selected patches are then considered in a many-to-many, cross-patch negative pairing scheme rather than using only corresponding pixels. A candidate negative pixel pair is retained only when its two pseudo-labels differ, which removes pseudo-label-based false negatives.

    The method maintains a background-dominant feature bank (BDB) and a foreground-dominant feature bank (FDB) containing patch features from the current and preceding batches. For a query from an FDP, the BDB supplies negative candidates; for a query from a BDP, the FDB supplies them. This cross-patch construction simultaneously aligns corresponding student–teacher patches and pixels while contrasting features drawn from patches with strongly different foreground/background composition.

  3. Knowl 3 — Pseudo-label-guided pixel and implicit patch contrastive loss

    equation

    For a projected query feature vector qq, let k+k^+ be its corresponding positive feature, let k−k^- range over candidate negative features in a feature bank FBF_B, let τ>0\tau>0 be the temperature, and let y^q\hat y_q and y^k−\hat y_{k^-} be their pseudo-labels. CDCL uses the masked InfoNCE loss

    ℓcontr(q)=−log⁡sim⁡(q,k+)sim⁡(q,k+)+∑k−∈FBJq,k−sim⁡(q,k−),\ell_{\mathrm{contr}}(q)=-\log\frac{\operatorname{sim}(q,k^+)}{\operatorname{sim}(q,k^+)+\sum_{k^-\in F_B}J_{q,k^-}\operatorname{sim}(q,k^-)},

    with

    sim⁡(q,k)=exp⁡(qTk∥q∥ ∥k∥ τ),Jq,k−=1[y^q≠y^k−].\operatorname{sim}(q,k)=\exp\left(\frac{q^{\mathsf T}k}{\lVert q\rVert\,\lVert k\rVert\,\tau}\right), \qquad J_{q,k^-}=\mathbf{1}[\hat y_q\neq\hat y_{k^-}].

    Here qq, k+k^+, and k−k^- are feature vectors, ∥⋅∥\lVert\cdot\rVert is the Euclidean norm, and 1[⋅]\mathbf{1}[\cdot] is one when its condition is true and zero otherwise. The mask excludes negative candidates whose pseudo-label agrees with the query.

    For a student patch feature map Φs\Phi_s with NN spatial positions, and its corresponding teacher patch Φt\Phi_t, the pixel-level patch loss is

    ℓcontrΦ(Φs)=1N∑h,wℓcontr(ϕsh,w),\ell_{\mathrm{contr}}^{\Phi}(\Phi_s)=\frac{1}{N}\sum_{h,w}\ell_{\mathrm{contr}}(\phi_s^{h,w}),

    where ϕsh,w\phi_s^{h,w} is the student feature at spatial position (h,w)(h,w) and the corresponding teacher feature ϕth,w\phi_t^{h,w} is its positive counterpart. The final contrastive loss averages these patch losses over only the foreground-dominant and background-dominant student patches:

    Lcontr=1NsB,F∑i=1NsB,FℓcontrΦ(Φsi),L_{\mathrm{contr}}=\frac{1}{N_s^{B,F}}\sum_{i=1}^{N_s^{B,F}}\ell_{\mathrm{contr}}^{\Phi}(\Phi_s^i),

    where NsB,FN_s^{B,F} is the number of BDPs and FDPs among the NsΦN_s^{\Phi} student patches. Thus, patch-level discrimination is imposed implicitly through the dense pixel losses, while corresponding features are pulled together and features from different pseudo-label classes are pushed apart.

  4. Knowl 4 — Prediction consistency and entropy minimization

    equation

    For an unlabeled image, let PusP_u^s and PutP_u^t denote the student and teacher class-probability maps, respectively. The teacher pseudo-label at each pixel is the class with maximum teacher probability:

    y^ut=arg⁡max⁡(Put).\hat y_u^t=\arg\max(P_u^t).

    The student is trained to match this teacher target with cross-entropy:

    Lcons=H(Pus,y^ut),L_{\mathrm{cons}}=H(P_u^s,\hat y_u^t),

    where HH is pixel-wise cross-entropy. Independently, the entropy of the student prediction is minimized:

    Lent=−1N∑n=1N∑c=1CPus,n,clog⁡Pus,n,c,L_{\mathrm{ent}}=-\frac{1}{N}\sum_{n=1}^{N}\sum_{c=1}^{C}P_u^{s,n,c}\log P_u^{s,n,c},

    where NN is the number of pixels, CC is the number of classes, and Pus,n,cP_u^{s,n,c} is the student probability that pixel nn belongs to class cc. Consistency regularization directly updates the classifier toward stable teacher predictions, while entropy minimization encourages confident predictions; their improved pseudo-labels in turn guide the contrastive pair selection.

  5. Knowl 5 — Training configuration and datasets

    experimental setup

    Experiments use the DSB nuclei dataset, containing 670 images from brightfield and fluorescence modalities, and MoNuSeg, containing 30 public training images and 14 test images from H&E-stained multi-organ tissue. DSB is randomly split into training, validation, and test sets with ratios 7:1:27{:}1{:}2; 20% of the public MoNuSeg training images are used for validation. MoNuSeg images are cropped into non-overlapping 250×250250\times250 sub-images. All training inputs are augmented and resized to 320×320320\times320 pixels. Experiments use labeled fractions of 1/321/32, 1/161/16, and 1/81/8 of the training data, with the remainder unlabeled.

    The segmentation backbone is DenseUNet with an ImageNet-pretrained DenseNet-161 backbone. The extractor has 256 output channels, and the projector is a fully connected layer, ReLU, and fully connected layer that reduces features to 128 channels. For an input of size h×wh\times w, a contrastive patch covers h/8×w/8h/8\times w/8 image pixels and corresponds to a h/64×w/64h/64\times w/64 feature patch. The fixed loss weights are wsup=1w_{\mathrm{sup}}=1, wcontr=0.1w_{\mathrm{contr}}=0.1, and went=0.01w_{\mathrm{ent}}=0.01; wconsw_{\mathrm{cons}} increases from 0 to 1 along a Gaussian schedule. The unsupervised branch starts at epoch 6, the teacher EMA coefficient is α=0.999\alpha=0.999, the initial learning rate is 0.00010.0001 with polynomial decay, and Adam is used with batch size 8 for labeled and unlabeled images. Training generally converges within 80 epochs.

    Random flipping, random cropping, rotation in [−15,15][-15,15] degrees, Gaussian blur, color jitter, and grayscale conversion are used as data augmentation. Feature banks store patch features from the current and previous batches, with gradient checkpointing used to control memory consumption.

  6. Knowl 6 — Limited-label benchmark performance

    data/table

    The proposed method was compared with SupOnly, TCSMv2, CutMix, GCT, CCT, and CAC under identical segmentation backbones, environments, and augmentations. The following representative entries report all five metrics for CDCL, CAC, and the fully supervised upper-bound model. DC and JC are Dice and Jaccard coefficients; ACC, SP, and SE are accuracy, specificity, and sensitivity. All values are percentages.

    Could not parse LaTeX table

    CDCL achieves the best reported Dice score among the compared semi-supervised methods at every label fraction on both datasets. Relative to CAC, its Dice improvement is 1.09 percentage points on DSB and 1.18 points on MoNuSeg with only 1/321/32 labeled data. At 1/321/32 supervision, it approaches the fully supervised Dice scores of 90.46% on DSB and 79.97% on MoNuSeg, obtaining 87.49% and 75.97%, respectively.

  7. Knowl 7 — Ablation evidence for sampling and auxiliary losses

    data/table

    Ablations use only 1/321/32 labeled training data and report Dice coefficients on DSB and MoNuSeg. SupOnly is DenseUNet trained only with labeled images. Scheme 1 uses random pixel-wise contrastive sampling; Scheme 2 uses pseudo-label-guided pixel-wise sampling; Scheme 3 uses the proposed pixel–patch sampling with contrastive loss. Schemes 4 and 5 add consistency or entropy loss individually, Scheme 6 uses consistency plus entropy without contrastive loss, and CDCL uses all three unlabeled objectives.

    Could not parse LaTeX table

    Random negative sampling performs worse than supervised-only training, whereas pseudo-label-guided sampling raises Dice to 85.54% on DSB and 73.69% on MoNuSeg. Replacing pixel-only sampling with cross-patch dense sampling further raises Dice to 86.26% and 74.49%. Adding both prediction-level losses to the contrastive objective produces the final scores of 87.49% and 75.97%, showing that feature-level contrastive learning and prediction-level regularization are complementary.

  8. Knowl 8 — Qualitative segmentation improvements

    empirical result

    Visual comparisons under 1/321/32 labeled supervision show that the proposed method more accurately captures both the global distribution of nuclei—object count and location—and local object details such as shape and size than the competing semi-supervised methods. Its predictions are visually close to those from the 100%-labeled FullSup model, with fewer over-predicted and under-predicted pixels. Ablation feature maps also show sharper separation between target and background features after pseudo-label-guided cross-patch alignment, consistent with the improvement in Dice score.

  9. Knowl 9 — Failure modes and scope limitation

    limitation

    The method still fails on cases containing extremely small nuclei and cases with extremely low contrast between nuclei and surrounding tissue. Although it performs well on most challenging examples with very limited annotation, the paper does not establish robustness beyond the two evaluated histopathology datasets; it identifies evaluation on additional datasets and integration into tumor-microenvironment analysis as future work.

Coverage note — The benchmark rows for competing methods other than CAC were condensed because the paper's main quantitative conclusions are captured by the proposed method, the strongest Dice competitor, the fully supervised reference, and the complete ablation results.

References

  1. 1.Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2209–2218, 2019. 1
  2. 2.Inigo Alonso, Alberto Sabater, David Ferstl, Luis Montesano, and Ana C Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8219–8228, 2021. 5
  3. 3.Juan C Caicedo, Allen Goodman, Kyle W Karhohs, Beth A Cimini, Jeanelle Ackerman, Marzieh Haghighi, CherKeng Heng, Tim Becker, Minh Doan, Claire McQuin, et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods, 16(12):1247–1253, 2019. 5
  4. 4.Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local features for medical image segmentation with limited annotations. arXiv preprint arXiv:2006.10511, 2020. 2
  5. 5.Minghao Chen, Hongyang Xue, and Deng Cai. Domain adaptation for semantic segmentation with maximum squares loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2090–2099, 2019. 5
  6. 6.Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597–1607. PMLR, 2020. 2, 3, 4
  7. 7.Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2
  8. 8.Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613–2622, 2021. 2
  9. 9.Yiqi Chen, Xuanya Li, Kai Hu, Zhineng Chen, and Xieping Gao. Nuclei segmentation in histopathology images using rotation equivariant and multi-level feature aggregation neural network. In 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 549–554. IEEE, 2020. 1
  10. 10.Zhengyang Feng, Qianyu Zhou, Guangliang Cheng, Xin Tan, Jianping Shi, and Lizhuang Ma. Semi-supervised semantic segmentation via dynamic self-training and class-balanced curriculum. arXiv preprint arXiv:2004.08514, 1(2):5, 2020. 1, 2
  11. 11.Geoff French, Samuli Laine, Timo Aila, Michal Mackiewicz, and Graham Finlayson. Semi-supervised semantic segmentation needs strong, varied perturbations. arXiv preprint arXiv:1906.01916, 2019. 2, 6
  12. 12.Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2
  13. 13.Yves Grandvalet, Yoshua Bengio, et al. Semi-supervised learning by entropy minimization. CAP, 367:281–296, 2005. 5
  14. 14.Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. arXiv preprint arXiv:2006.07733, 2020. 2
  15. 15.Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. 5
  16. 16.Zhanghan Ke, Di Qiu, Kaican Li, Qiong Yan, and Rynson WH Lau. Guided collaborative training for pixel-wise semi-supervised learning. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIII 16, pages 429–445. Springer, 2020. 2, 6
  17. 17.Zhanghan Ke, Daoye Wang, Qiong Yan, Jimmy Ren, and Rynson WH Lau. Dual student: Breaking the limits of the teacher in semi-supervised learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6728–6736, 2019. 1
  18. 18.Anna Khoreva, Rodrigo Benenson, Jan Hosang, Matthias Hein, and Bernt Schiele. Simple does it: Weakly supervised instance and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 876–885, 2017. 1
  19. 19.Neeraj Kumar, Ruchika Verma, Deepak Anand, Yanning Zhou, Omer Fahri Onder, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng, Jiahui Li, Zhiqiang Hu, et al. A multi-organ nucleus segmentation challenge. IEEE transactions on medical imaging, 39(5):1380–1391, 2019. 5
  20. 20.Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised semantic segmentation with directional context-aware consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1205–1214, 2021. 1, 2, 3, 6
  21. 21.Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 1
  22. 22.Xiaomeng Li, Hao Chen, Xiaojuan Qi, Qi Dou, Chi-Wing Fu, and Pheng-Ann Heng. H-denseunet: hybrid densely connected unet for liver and tumor segmentation from ct volumes. IEEE transactions on medical imaging, 37(12):2663–2674, 2018. 5, 6
  23. 23.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 Transactions on Neural Networks and Learning Systems, 32(2):523–534, 2020. 1, 2, 6
  24. 24.Dongnan Liu, Donghao Zhang, Yang Song, Chaoyi Zhang, Fan Zhang, Lauren O’Donnell, and Weidong Cai. Nuclei segmentation via a deep panoptic model with semantic feature fusion. In IJCAI, pages 861–868, 2019. 2
  25. 25.Weizhe Liu, David Ferstl, Samuel Schulter, Lukas Zebedin, Pascal Fua, and Christian Leistner. Domain adaptation for semantic segmentation via patch-wise contrastive learning. arXiv preprint arXiv:2104.11056, 2021. 5
  26. 26.Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. 1, 2
  27. 27.Shivang Naik, Scott Doyle, Shannon Agner, Anant Madabhushi, Michael Feldman, and John Tomaszewski. Automated gland and nuclei segmentation for grading of prostate and breast cancer histopathology. In 2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, pages 284–287. IEEE, 2008. 2
  28. 28.Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4
  29. 29.Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE transactions on systems, man, and cybernetics, 9(1):62–66, 1979. 2
  30. 30.Yassine Ouali, Celine Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674–12684, 2020. 2, 6
  31. 31.Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Greg Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32:8026–8037, 2019. 5
  32. 32.Shan E Ahmed Raza, Linda Cheung, Muhammad Shaban, Simon Graham, David Epstein, Stella Pelengaris, Michael Khan, and Nasir M Rajpoot. Micro-net: A unified model for segmentation of various objects in microscopy images. Medical image analysis, 52:160–173, 2019. 2
  33. 33.Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211–252, 2015. 5
  34. 34.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. arXiv preprint arXiv:2001.07685, 2020. 5
  35. 35.Jost Tobias Springenberg. Unsupervised and semi-supervised learning with categorical generative adversarial networks. arXiv preprint arXiv:1511.06390, 2015. 2
  36. 36.Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. arXiv preprint arXiv:1703.01780, 2017. 1, 2, 3, 5
  37. 37.Paul Vernaza and Manmohan Chandraker. Learning random-walk label propagation for weakly-supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7158–7166, 2017. 1
  38. 38.Wenguan Wang, Tianfei Zhou, Fisher Yu, Jifeng Dai, Ender Konukoglu, and Luc Van Gool. Exploring cross-image pixel contrast for semantic segmentation. arXiv preprint arXiv:2101.11939, 2021. 1, 2, 3
  39. 39.Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3024–3033, 2021. 2
  40. 40.Yicheng Wu, Minfeng Xu, Zongyuan Ge, Jianfei Cai, and Lei Zhang. Semi-supervised left atrium segmentation with mutual consistency training. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 297–306. Springer, 2021. 5
  41. 41.Zhenda Xie, Yutong Lin, Zheng Zhang, Yue Cao, Stephen Lin, and Han Hu. Propagate yourself: Exploring pixel-level consistency for unsupervised visual representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16684–16693, 2021. 2
  42. 42.Bingchao Zhao, Xin Chen, Zhi Li, Zhiwen Yu, Su Yao, Lixu Yan, Yuqian Wang, Zaiyi Liu, Changhong Liang, and Chu Han. Triple u-net: Hematoxylin-aware nuclei segmentation with progressive dense feature aggregation. Medical Image Analysis, 65:101786, 2020. 1, 2
  43. 43.Meng Zhao, Hao Wang, Ying Han, Xiaokang Wang, Hong-Ning Dai, Xuguo Sun, Jin Zhang, and Marius Pedersen. Seens: Nuclei segmentation in pap smear images with selective edge enhancement. Future Generation Computer Systems, 114:185–194, 2021. 1
  44. 44.Yuanyi Zhong, Bodi Yuan, Hong Wu, Zhiqiang Yuan, Jian Peng, and Yu-Xiong Wang. Pixel contrastive-consistent semi-supervised semantic segmentation. arXiv preprint arXiv:2108.09025, 2021. 1
  45. 45.Yanning Zhou, Omer Fahri Onder, Qi Dou, Efstratios Tsougenis, Hao Chen, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Information Processing in Medical Imaging, pages 682–693. Springer, 2019. 2

Citation

MLA
Wu, H., et al. “Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 11656–65, https://doi.org/10.1109/CVPR52688.2022.01137.
APA
Wu, H., Wang, Z., Song, Y., Yang, L., & Qin, J. (2022). Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11656–11665. https://doi.org/10.1109/CVPR52688.2022.01137
Chicago
Wu, H., Z. Wang, Y. Song, L. Yang, and J. Qin. 2022. “Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images”. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11656–65. https://doi.org/10.1109/CVPR52688.2022.01137.
Harvard
Wu, H. et al. (2022) “Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images”, 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 11656–11665. Available at: https://doi.org/10.1109/CVPR52688.2022.01137.
Vancouver
1. Wu H, Wang Z, Song Y, Yang L, Qin J (2022) Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp 11656–11665

BibTeX

@inproceedings{Wu_2022, title={Cross-patch Dense Contrastive Learning for Semi-supervised Segmentation of Cellular Nuclei in Histopathologic Images}, url={http://dx.doi.org/10.1109/CVPR52688.2022.01137}, DOI={10.1109/cvpr52688.2022.01137}, booktitle={2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, publisher={IEEE}, author={Wu, Huisi and Wang, Zhaoze and Song, Youyi and Yang, Lin and Qin, Jing}, year={2022}, month=June, pages={11656–11665} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF
License: IEEE