Signet Ring Cell Detection With a Semi-supervised Learning Framework

Chaofu Wang; Hongsheng Li; Jiahui Li; Qian Da; Qi Duan; Shuang Yang; Xiaodi Huang; Xiaoqun Yang; Zhiqiang Hu

arxiv: 1907.03954 · v1 · pith:SVBJIUSInew · submitted 2019-07-09 · 💻 cs.CV

Signet Ring Cell Detection With a Semi-supervised Learning Framework

Jiahui Li , Shuang Yang , Xiaodi Huang , Qian Da , Xiaoqun Yang , Zhiqiang Hu , Qi Duan , Chaofu Wang

show 1 more author

Hongsheng Li

This is my paper

Pith reviewed 2026-05-25 01:00 UTC · model grok-4.3

classification 💻 cs.CV

keywords signet ring cell detectionsemi-supervised learningself-trainingcooperative-trainingpathology imagesadenocarcinomamedical image analysisobject detection

0 comments

The pith

A semi-supervised framework combines self-training and cooperative-training to detect signet ring cells accurately from both labeled and unlabeled pathology images.

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

The paper develops a semi-supervised learning method for spotting signet ring cells, a rare form of cancer cell whose early identification improves survival rates. Visual inspection by pathologists is slow and risks missing cells, so an automated solution would reduce labor and omissions. Self-training fills gaps in incomplete annotations while cooperative-training labels additional unlabeled image regions. The combination lets the system draw on both labeled and unlabeled clinical data. Experiments on large real datasets confirm the approach produces accurate detections suitable for clinical use.

Core claim

The authors present a semi-supervised learning framework for signet ring cell detection that applies self-training to handle incomplete annotations and adapts cooperative-training to explore unlabeled regions. This design makes fuller use of both labeled and unlabeled data. On large real clinical pathology datasets the framework delivers accurate detection and can be applied directly in clinical trials.

What carries the argument

The semi-supervised learning framework that integrates self-training for incomplete annotations with cooperative-training for unlabeled regions.

If this is right

The method improves detection by exploiting unlabeled data in addition to limited labeled examples.
The resulting detector reaches accuracy levels ready for use in clinical trials.
Release of the dataset supports further work on signet ring cell problems.
The framework demonstrates that combining the two training techniques yields better results than either alone.

Where Pith is reading between the lines

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

The same combination of self-training and cooperative-training could extend to detection tasks for other rare cell types where full annotation is costly.
Wider adoption might reduce the annotation burden on pathologists when building new detection models.
If the error-control property holds across datasets, the approach could support screening programs for cancers with low incidence but high risk.
Testing the framework on different staining protocols or magnification levels would check robustness beyond the reported experiments.

Load-bearing premise

Cooperative-training can reliably explore and label unlabeled regions in pathology images without introducing errors that degrade overall detection performance.

What would settle it

A direct comparison on the same clinical dataset where adding the cooperative-training step increases false positives or misses more cells than self-training alone.

Figures

Figures reproduced from arXiv: 1907.03954 by Chaofu Wang, Hongsheng Li, Jiahui Li, Qian Da, Qi Duan, Shuang Yang, Xiaodi Huang, Xiaoqun Yang, Zhiqiang Hu.

**Figure 1.** Figure 1: Signet ring cells are overcrowded and of various appearances. Cells in green rectangles are signet ring cells, which are the indicator of signet ring cell carcinoma. Cells in yellow rectangles are also signet ring cells but are missed by pathologists in crowd regions, during long time tedious annotation. We propose a self-training method to deal with the challenge of incomplete annotations. We observe that… view at source ↗

**Figure 2.** Figure 2: Overview of our semi-supervised framework: Initial fully-supervised training, self-training, and cooperative-training [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: 3 regions of over 2, 000 × 2, 000 pixels are randomly cropped from each WSI and bounding boxes for high confident independent signet ring cells are annotated. 2.1 Dataset Our dataset is collected from several highly ranked hospitals, and consists of H&E stained images captured at 40× magnification. Containing 127 (21 positive + 106 negative) whole slide images (WSIs), this dataset covers a large number of … view at source ↗

**Figure 4.** Figure 4: Signet ring cell detector: image → 3-class mask → cell instance mask → cell box prediction [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 5.** Figure 5: Pipeline of our self-training strategy to self-correct imperfectly annotated images. The next-round model is trained on annotations from the previous round, and iteratively adjusts annotations towards higher quality. Initially we draw inscribed ellipse in each rectangle as ground truth inner region. The gray regions are edge mask and the white regions are the inner regions. The green arrow points to a gr… view at source ↗

**Figure 6.** Figure 6: Pipeline of the cooperative-training strategy. For new unlabeled images, each model is trained on predictions from the other model, so that we can reduce the possibility that one model get stuck in its local minimum , and allow the two models to support each others. In this way we will gradually obtain higher annotation quality on unlabeled images without any manual interventions. The yellow arrow points … view at source ↗

**Figure 7.** Figure 7: Example detection results in test images of over 2, 000×2, 000 pixels. Our semisupervised learning framework can obtain cell edges of each signet ring cell as shown in yellow polygon. Comparing self-training with self-training-extra, unlabeled images with selfgenerated annotations can harm the performance, especially in easy mode where the false positive rises to 5.22. Visually we find that there are too… view at source ↗

read the original abstract

Signet ring cell carcinoma is a type of rare adenocarcinoma with poor prognosis. Early detection leads to huge improvement of patients' survival rate. However, pathologists can only visually detect signet ring cells under the microscope. This procedure is not only laborious but also prone to omission. An automatic and accurate signet ring cell detection solution is thus important but has not been investigated before. In this paper, we take the first step to present a semi-supervised learning framework for the signet ring cell detection problem. Self-training is proposed to deal with the challenge of incomplete annotations, and cooperative-training is adapted to explore the unlabeled regions. Combining the two techniques, our semi-supervised learning framework can make better use of both labeled and unlabeled data. Experiments on large real clinical data demonstrate the effectiveness of our design. Our framework achieves accurate signet ring cell detection and can be readily applied in the clinical trails. The dataset will be released soon to facilitate the development of the area.

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.

Desk Editor's Note private letter to a colleague

This is the first paper on automatic signet ring cell detection, adapting self-training plus cooperative-training to incomplete pathology annotations, but the abstract gives zero numbers or validation details.

read the letter

The main point is that the work claims to be the first automatic detector for signet ring cells and shows how to combine self-training for missing labels with cooperative-training for unlabeled image regions. That framing of the clinical problem is clear and the adaptation itself is straightforward. The paper correctly notes that early detection matters for this rare cancer and that full supervision is impractical on large slides. If the experiments later demonstrate gains over a plain supervised baseline on real clinical data, the combination could be a practical starting point for others facing similar annotation gaps. The methods section appears to describe standard adaptations without new theory, which is fine for an application paper. The soft spot is the complete absence of metrics, splits, or error analysis in the abstract. No precision, recall, or comparison numbers are supplied, so the claim that the framework “achieves accurate” detection cannot be checked. The stress-test worry about cooperative-training injecting undetected false positives on a rare, variable class is worth watching; single-modality slides often lack the independent views that co-training assumes, and the abstract does not mention disagreement thresholds or correction steps. If the full paper does not address that risk with ablations, the central claim weakens. This paper is mainly for medical-image researchers who already work with semi-supervised methods and want to see them applied to a new detection task. A reader looking for a ready-to-use pipeline or strong empirical evidence will find little here yet. It deserves peer review once the results and validation details are in place, because the task is genuinely new and the clinical motivation is solid; without those sections it would be a desk reject.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a semi-supervised learning framework for signet ring cell detection in pathology images. It combines self-training to address incomplete annotations with an adaptation of cooperative-training to explore unlabeled regions. The central claim is that this combination makes better use of both labeled and unlabeled data, and that experiments on large real clinical data demonstrate the framework's effectiveness in achieving accurate detection, with plans to release the dataset.

Significance. If the experimental results hold with proper quantitative validation, the work would be significant for computational pathology: it targets an important clinical problem (early detection of a rare adenocarcinoma with poor prognosis) where expert annotations are scarce and expensive, and shows how existing semi-supervised techniques can be combined to leverage abundant unlabeled slides.

major comments (2)

[Abstract] Abstract: the claim that the framework 'achieves accurate signet ring cell detection' and 'demonstrate[s] the effectiveness of our design' supplies no quantitative metrics (e.g., precision, recall, F1), no baselines (supervised or otherwise), no validation splits, and no error analysis, rendering the central empirical claim unsupported.
[Abstract] Abstract (paragraph on cooperative-training adaptation): the description states only that cooperative-training 'is adapted' without specifying view construction, disagreement threshold, or any error-correction step. In a setting where the positive class is rare and morphologically variable, this omission directly undermines the assumption that pseudo-labeling of unlabeled regions will not introduce net performance degradation via error propagation.

minor comments (1)

[Abstract] Abstract, final sentence: 'clinical trails' is a typographical error and should read 'clinical trials'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We agree that it can be strengthened with quantitative details and more specificity on the method adaptation, and will revise it accordingly while keeping the full technical descriptions in the main text.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the framework 'achieves accurate signet ring cell detection' and 'demonstrate[s] the effectiveness of our design' supplies no quantitative metrics (e.g., precision, recall, F1), no baselines (supervised or otherwise), no validation splits, and no error analysis, rendering the central empirical claim unsupported.

Authors: The abstract is intended as a high-level summary. Quantitative results (F1 scores, precision/recall, supervised baselines, validation splits on the clinical dataset, and error analysis) appear in Section 4. To address the concern directly, we will revise the abstract to report the key metrics and briefly note the evaluation protocol. revision: yes
Referee: [Abstract] Abstract (paragraph on cooperative-training adaptation): the description states only that cooperative-training 'is adapted' without specifying view construction, disagreement threshold, or any error-correction step. In a setting where the positive class is rare and morphologically variable, this omission directly undermines the assumption that pseudo-labeling of unlabeled regions will not introduce net performance degradation via error propagation.

Authors: The concrete adaptations (view construction for the two networks, disagreement threshold for pseudo-label selection, and error-correction via self-training filtering) are specified in Section 3.2. We acknowledge the abstract's brevity leaves this implicit. We will add a short clause in the abstract describing these elements at a high level. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical adaptation of existing semi-supervised methods with no derivation chain

full rationale

The paper describes an adaptation of self-training and cooperative-training for signet ring cell detection on pathology images. No equations, derivations, or parameter-fitting steps are presented that could reduce to inputs by construction. Claims rest on experimental results on clinical data rather than any self-definitional, fitted-prediction, or self-citation load-bearing structure. The approach is presented as a practical combination of known techniques without renaming known results or smuggling ansatzes via citation in a manner that creates circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility into specific hyperparameters or modeling choices; standard semi-supervised assumptions are implicit but not enumerated.

axioms (1)

domain assumption Pseudo-labels generated during self-training and cooperative-training improve model performance without systematic bias
Core premise of the self-training and cooperative-training components described in the abstract

pith-pipeline@v0.9.0 · 5717 in / 1189 out tokens · 21803 ms · 2026-05-25T01:00:18.644438+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · 2 internal anchors

[1]

Biometrics 65(1), 247–256 (2009)

Bandos, A.I., Rockette, H.E., Song, T., Gur, D.: Area under the free-response roc curve (froc) and a related summary index. Biometrics 65(1), 247–256 (2009)

work page 2009
[2]

Bosman, F.T., Carneiro, F., Hruban, R.H., Theise, N.D., et al.: WHO classiﬁ- cation of tumours of the digestive system., pp. 52–53. No. Ed. 4, World Health Organization (2010) 12 J.Li et al

work page 2010
[3]

In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition

Chen, H., Qi, X., Yu, L., Heng, P.A.: Dcan: deep contour-aware networks for ac- curate gland segmentation. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2487–2496 (2016)

work page 2016
[4]

International Journal of Computer Vision 88(2), 303–338 (Jun 2010)

Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International Journal of Computer Vision 88(2), 303–338 (Jun 2010)

work page 2010
[5]

IEEE transactions on pattern analysis and machine intelligence 28(11), 1768–1783 (2006)

Grady, L.: Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence 28(11), 1768–1783 (2006)

work page 2006
[6]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

work page 2016
[7]

Simple Does It: Weakly Supervised Instance and Semantic Segmentation

Hosang, A.K.R.B.J., Schiele, M.H.B.: Weakly supervised semantic labelling and instance segmentation. arXiv preprint arXiv:1603.07485 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[8]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Hu, R., Doll´ ar, P., He, K., Darrell, T., Girshick, R.: Learning to segment every thing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4233–4241 (2018)

work page 2018
[9]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA

Luo, P., Wang, G., Lin, L., Wang, X.: Deep dual learning for semantic image segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. pp. 21–26 (2017)

work page 2017
[10]

Weakly- and Semi-Supervised Learning of a DCNN for Semantic Image Segmentation

Papandreou, G., Chen, L., Murphy, K., Yuille, A.L.: Weakly- and semi-supervised learning of a DCNN for semantic image segmentation. CoRR abs/1502.02734 (2015), http://arxiv.org/abs/1502.02734

work page internal anchor Pith review Pith/arXiv arXiv 2015
[11]

IEEE transactions on medical imaging 36(2), 674–683 (2017)

Rajchl, M., Lee, M.C., Oktay, O., Kamnitsas, K., Passerat-Palmbach, J., Bai, W., Damodaram, M., Rutherford, M.A., Hajnal, J.V., Kainz, B., et al.: Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging 36(2), 674–683 (2017)

work page 2017
[12]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Uniﬁed, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779–788 (2016)

work page 2016
[13]

In: Advances in neural information processing systems

Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detec- tion with region proposal networks. In: Advances in neural information processing systems. pp. 91–99 (2015)

work page 2015
[14]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

work page 2015
[15]

In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition

Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2403–2412 (2018)

work page 2018
[16]

In: Computer Vision and Pattern Recognition, 2009

Zach, C., Niethammer, M., Frahm, J.M.: Continuous maximal ﬂows and wulﬀ shapes: Application to mrfs. In: Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. pp. 1911–1918. IEEE (2009)

work page 2009
[17]

IEEE Transac- tions on Medical Imaging 34(2), 496–506 (2015)

Zhang, X., Liu, W., Dundar, M., Badve, S., Zhang, S.: Towards large-scale histopathological image analysis: Hashing-based image retrieval. IEEE Transac- tions on Medical Imaging 34(2), 496–506 (2015)

work page 2015
[18]

Medical image analysis 26(1), 306–315 (2015)

Zhang, X., Xing, F., Su, H., Yang, L., Zhang, S.: High-throughput histopathological image analysis via robust cell segmentation and hashing. Medical image analysis 26(1), 306–315 (2015)

work page 2015
[19]

In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Zhao, X., Liang, S., Wei, Y.: Pseudo mask augmented object detection. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4061–4070 (2018)

work page 2018

[1] [1]

Biometrics 65(1), 247–256 (2009)

Bandos, A.I., Rockette, H.E., Song, T., Gur, D.: Area under the free-response roc curve (froc) and a related summary index. Biometrics 65(1), 247–256 (2009)

work page 2009

[2] [2]

Bosman, F.T., Carneiro, F., Hruban, R.H., Theise, N.D., et al.: WHO classiﬁ- cation of tumours of the digestive system., pp. 52–53. No. Ed. 4, World Health Organization (2010) 12 J.Li et al

work page 2010

[3] [3]

In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition

Chen, H., Qi, X., Yu, L., Heng, P.A.: Dcan: deep contour-aware networks for ac- curate gland segmentation. In: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. pp. 2487–2496 (2016)

work page 2016

[4] [4]

International Journal of Computer Vision 88(2), 303–338 (Jun 2010)

Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International Journal of Computer Vision 88(2), 303–338 (Jun 2010)

work page 2010

[5] [5]

IEEE transactions on pattern analysis and machine intelligence 28(11), 1768–1783 (2006)

Grady, L.: Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence 28(11), 1768–1783 (2006)

work page 2006

[6] [6]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

work page 2016

[7] [7]

Simple Does It: Weakly Supervised Instance and Semantic Segmentation

Hosang, A.K.R.B.J., Schiele, M.H.B.: Weakly supervised semantic labelling and instance segmentation. arXiv preprint arXiv:1603.07485 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[8] [8]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Hu, R., Doll´ ar, P., He, K., Darrell, T., Girshick, R.: Learning to segment every thing. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4233–4241 (2018)

work page 2018

[9] [9]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA

Luo, P., Wang, G., Lin, L., Wang, X.: Deep dual learning for semantic image segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA. pp. 21–26 (2017)

work page 2017

[10] [10]

Weakly- and Semi-Supervised Learning of a DCNN for Semantic Image Segmentation

Papandreou, G., Chen, L., Murphy, K., Yuille, A.L.: Weakly- and semi-supervised learning of a DCNN for semantic image segmentation. CoRR abs/1502.02734 (2015), http://arxiv.org/abs/1502.02734

work page internal anchor Pith review Pith/arXiv arXiv 2015

[11] [11]

IEEE transactions on medical imaging 36(2), 674–683 (2017)

Rajchl, M., Lee, M.C., Oktay, O., Kamnitsas, K., Passerat-Palmbach, J., Bai, W., Damodaram, M., Rutherford, M.A., Hajnal, J.V., Kainz, B., et al.: Deepcut: Object segmentation from bounding box annotations using convolutional neural networks. IEEE transactions on medical imaging 36(2), 674–683 (2017)

work page 2017

[12] [12]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Redmon, J., Divvala, S., Girshick, R., Farhadi, A.: You only look once: Uniﬁed, real-time object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 779–788 (2016)

work page 2016

[13] [13]

In: Advances in neural information processing systems

Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: Towards real-time object detec- tion with region proposal networks. In: Advances in neural information processing systems. pp. 91–99 (2015)

work page 2015

[14] [14]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

work page 2015

[15] [15]

In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition

Yu, F., Wang, D., Shelhamer, E., Darrell, T.: Deep layer aggregation. In: Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 2403–2412 (2018)

work page 2018

[16] [16]

In: Computer Vision and Pattern Recognition, 2009

Zach, C., Niethammer, M., Frahm, J.M.: Continuous maximal ﬂows and wulﬀ shapes: Application to mrfs. In: Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. pp. 1911–1918. IEEE (2009)

work page 2009

[17] [17]

IEEE Transac- tions on Medical Imaging 34(2), 496–506 (2015)

Zhang, X., Liu, W., Dundar, M., Badve, S., Zhang, S.: Towards large-scale histopathological image analysis: Hashing-based image retrieval. IEEE Transac- tions on Medical Imaging 34(2), 496–506 (2015)

work page 2015

[18] [18]

Medical image analysis 26(1), 306–315 (2015)

Zhang, X., Xing, F., Su, H., Yang, L., Zhang, S.: High-throughput histopathological image analysis via robust cell segmentation and hashing. Medical image analysis 26(1), 306–315 (2015)

work page 2015

[19] [19]

In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Zhao, X., Liang, S., Wei, Y.: Pseudo mask augmented object detection. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 4061–4070 (2018)

work page 2018