Domain Adaptation-based Augmentation for Weakly Supervised Nuclei Detection

Ansh Kapil; Armin Meier; Christos G. Gavriel; G\"unter Schmidt; Nicolas Brieu; Peter D. Caie; Ralf Schoenmeyer

arxiv: 1907.04681 · v1 · pith:FXCBHV35new · submitted 2019-07-10 · 📡 eess.IV · cs.CV

Domain Adaptation-based Augmentation for Weakly Supervised Nuclei Detection

Nicolas Brieu , Armin Meier , Ansh Kapil , Ralf Schoenmeyer , Christos G. Gavriel , Peter D. Caie , G\"unter Schmidt This is my paper

Pith reviewed 2026-05-24 23:28 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords nuclei detectiondomain adaptationweakly supervised learningcomputational pathologyimage-to-image translationstain normalizationpathology images

0 comments

The pith

Unpaired image-to-image translation plus stain normalization turns source-labeled images into training data that works on new target domains for nuclei detection.

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

The paper aims to cut the labeling cost when moving nuclei detection models to new tissue stains or patient cohorts. It starts with already-labeled images on a source domain and applies stain normalization followed by unpaired translation to generate synthetic images that match the appearance of an unlabeled target domain while carrying over the source labels. These synthetic pairs are then used to train a detector that is tested on real target images. Experiments indicate this inter-domain route beats the usual approach of training a fully supervised detector within a single domain. The result matters because new cohorts keep arriving with different stains and the cost of fresh expert labels is high.

Core claim

Performing stain normalization and unpaired image-to-image translation on labeled source images produces synthetic labeled target images that can be used to train a detection network whose performance on real target images exceeds that of fully-supervised intra-domain detectors.

What carries the argument

Unpaired image-to-image translation combined with stain normalization to create synthetic labeled target-domain images for detector training.

If this is right

New target domains can be handled without acquiring fresh labels for them.
Existing labeled source datasets become reusable across different stains and indications.
Detection performance on the target domain improves relative to standard fully supervised training that stays inside one domain.
Labeling effort scales with the number of source domains rather than every new target domain.

Where Pith is reading between the lines

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

The same translation step might reduce labeling needs in other medical imaging tasks that suffer from stain or scanner shifts.
One could measure how much the method degrades when the source and target domains differ more extremely than the paper's test cases.
Combining the generated images with a small number of real target labels might produce further gains that the paper does not test.

Load-bearing premise

The source labels stay accurate enough after translation that a network trained on the synthetic images will still detect nuclei correctly on real target images.

What would settle it

Training a detector on the generated synthetic images and finding that its accuracy on held-out real target images falls below the accuracy of a detector trained on a modest set of manually labeled target images.

Figures

Figures reproduced from arXiv: 1907.04681 by Ansh Kapil, Armin Meier, Christos G. Gavriel, G\"unter Schmidt, Nicolas Brieu, Peter D. Caie, Ralf Schoenmeyer.

**Figure 1.** Figure 1: Proposed method for the weakly supervised detection of nuclei on a target domain (e.g. HE) based on labeled images from a different source domain (e.g. IF). well as the high variability between samples motivates the development of image analysis methods working and reusing information across different domains. This is particularly true for the detection of objects (e.g. nuclei) and regions (e.g. epitheliu… view at source ↗

**Figure 2.** Figure 2: Proposed normalization of HE images. (a) Visually consistent reference images; (b) Input images with high color variability; (c) Results of DC-GMM normalization, resulting in unrealistic patterns in saturated regions; (d) Normalized and realistic looking images obtained with one-to-one domain adaptation between (a) and (c). but realistic synthetic images. The proposed methodology is weakly supervised: nuc… view at source ↗

**Figure 3.** Figure 3: Results of unpaired one-to-one ’inter-domain’ transformation and augmentation between IF and HE normalized images. (a) HE2IF: HE to IF transformation and augmentation. (b) IF2HE: IF to HE transformation and augmentation. on minimum and maximum values. For the more complex normalization of HE images, we combine previous works based on DCGMM [16] and CycleGAN [13]. As shown in [PITH_FULL_IMAGE:figures/full… view at source ↗

**Figure 4.** Figure 4: Detection accuracy with inter-domain (HE&IF2HE - NHE = 0), intra-domain (HE only), and cross-domain (HE&IF2HE - NHE > 0) supervisions for increasing availability of labeled nuclei NHEin the target domain (HE), if (a) an ensemble or (b) solely the best of the IF2HE inter-domain transformation models are considered for generation of the synthetic IF2HE images. domain (HE2IF) and of (ii) HE as unlabeled targe… view at source ↗

**Figure 5.** Figure 5: Comparison analysis between different methodologies for domain adaptation and detection: (a) Inter-domain vs. Intra-domain; (b) Cross-domain vs. Intra-domain; (c) Cross-domain vs. Inter-domain; (d) Ensemble vs. Best single CycleGAN models. More precisely, relative improvement of detection accuracy yielded by the first method vs. the second method, and significance testing as measured by paired Student t-Te… view at source ↗

read the original abstract

The detection of nuclei is one of the most fundamental components of computational pathology. Current state-of-the-art methods are based on deep learning, with the prerequisite that extensive labeled datasets are available. The increasing number of patient cohorts to be analyzed, the diversity of tissue stains and indications, as well as the cost of dataset labeling motivates the development of novel methods to reduce labeling effort across domains. We introduce in this work a weakly supervised 'inter-domain' approach that (i) performs stain normalization and unpaired image-to-image translation to transform labeled images on a source domain to synthetic labeled images on an unlabeled target domain and (ii) uses the resulting synthetic labeled images to train a detection network on the target domain. Extensive experiments show the superiority of the proposed approach against the state-of-the-art 'intra-domain' detection based on fully-supervised learning.

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

Applies stain normalization plus unpaired translation to create synthetic target-domain training images for nuclei detection, but provides no check that the attached source labels stay accurate after translation.

read the letter

The core idea is to take labeled source-domain images, run stain normalization and unpaired image-to-image translation to match a target stain, keep the original point or box labels, and train a detector on the resulting synthetic target images. The abstract claims this inter-domain route beats standard fully supervised intra-domain training. That is the main new angle: a practical pipeline that tries to cut labeling costs when moving across stains in computational pathology. The components themselves are established, so the contribution is the specific combination and the empirical comparison rather than a new algorithm or proof. If the experiments are clean, the pipeline description could still be useful to labs that routinely face stain variation. The soft spot is the missing validation on label fidelity. Unpaired translation optimizes global appearance and cycle consistency but has no built-in term that keeps small, crowded nuclei in the same locations or counts. Without a reported check such as expert re-labeling of the translated patches, landmark consistency metrics, or a controlled test showing that a detector trained on the synthetic data performs well on real target images with known ground truth, the superiority numbers are difficult to trust. The paper does not appear to introduce new equations or free parameters, and the citation pattern follows the usual domain-adaptation references. This work is aimed at computational pathology groups that already run nuclei detectors and need to handle new stains without fresh annotations. A reader who wants a concrete example of inter-domain augmentation might extract the pipeline details and the reported numbers, but anyone focused on rigorous domain-shift evaluation will want the label-preservation evidence filled in. It is worth sending to peer review so the experimental section can be examined directly; the central claim is testable once the methods and any auxiliary checks are visible.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a weakly supervised inter-domain approach for nuclei detection in computational pathology. Labeled source-domain images undergo stain normalization and unpaired image-to-image translation (e.g., CycleGAN-style) to produce synthetic target-domain images that retain the source labels; these synthetic pairs are then used to train a detection network that is evaluated on real target-domain images. The central claim is that extensive experiments demonstrate superiority of this inter-domain method over state-of-the-art intra-domain fully-supervised detection.

Significance. If the assumption that label fidelity is preserved holds, the method could meaningfully reduce annotation burden across diverse stains and cohorts by transferring existing labeled data rather than requiring new target-domain annotations.

major comments (2)

[Abstract (methods description) and Experiments section] The load-bearing assumption that unpaired translation preserves nuclei locations, sizes, and counts with sufficient accuracy for supervised detection training is not quantitatively validated. No section reports expert re-annotation of translated patches, landmark consistency metrics, or comparison of nuclei counts pre- and post-translation, leaving open the possibility that global style optimization introduces local geometric distortions that undermine the attached labels.
[Experiments section (quantitative results)] The superiority claim over intra-domain fully-supervised baselines is presented without sufficient detail on whether the comparison controls for label quality, network architecture, or training protocol; if the synthetic labels are noisier than real target labels, the reported gains may not generalize.

minor comments (2)

[Abstract] Clarify in the abstract and methods whether the approach is truly weakly supervised or relies on full source labels transferred via adaptation.
[Experiments section] Specify the exact datasets, number of images, and baseline implementations used in the experiments for reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We respond to each major comment below and indicate planned revisions.

read point-by-point responses

Referee: [Abstract (methods description) and Experiments section] The load-bearing assumption that unpaired translation preserves nuclei locations, sizes, and counts with sufficient accuracy for supervised detection training is not quantitatively validated. No section reports expert re-annotation of translated patches, landmark consistency metrics, or comparison of nuclei counts pre- and post-translation, leaving open the possibility that global style optimization introduces local geometric distortions that undermine the attached labels.

Authors: We acknowledge that the manuscript does not include direct quantitative validation such as expert re-annotation or landmark consistency metrics for label preservation after translation. The reported performance improvements on real target-domain images provide indirect support for sufficient label fidelity, but we agree this assumption would benefit from explicit checks. In the revision we will add a nuclei count comparison before and after translation. revision: yes
Referee: [Experiments section (quantitative results)] The superiority claim over intra-domain fully-supervised baselines is presented without sufficient detail on whether the comparison controls for label quality, network architecture, or training protocol; if the synthetic labels are noisier than real target labels, the reported gains may not generalize.

Authors: All comparisons used the identical detection network architecture and training protocol. The intra-domain baseline was trained on real target-domain labels while the proposed method used the generated synthetic labels; this difference is inherent to the inter-domain setting. We will revise the experiments section to state these controls explicitly and discuss implications for generalization. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical method with external components

full rationale

The manuscript presents a practical pipeline that applies existing stain normalization and unpaired image-to-image translation (e.g., CycleGAN-style) to attach source labels to synthetic target-domain images, then trains and evaluates a detector empirically. No equations, parameter-fitting steps, or derivations are described that would reduce any claimed result to a self-defined quantity or to a self-citation chain. The superiority claim rests on experimental comparisons against fully-supervised intra-domain baselines rather than on any internal mathematical closure. External techniques are invoked without the paper re-deriving or re-fitting them in a circular manner.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are described. The method implicitly assumes that standard domain-adaptation components preserve label semantics.

pith-pipeline@v0.9.0 · 5693 in / 1092 out tokens · 22475 ms · 2026-05-24T23:28:03.299691+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages · 2 internal anchors

[1]

In: ISBI (2017)

Brieu, N., Schmidt, G.: Learning size adaptive local maxima selection for robust nuclei detection in histopathology images. In: ISBI (2017)

work page 2017
[2]

Nature comm

Carstens, J., et al.: Spatial computation of intratumoral t cells correlates with survival of patients with pancreatic cancer. Nature comm. (2017)

work page 2017
[3]

In: Intl Wksp on Simulation and Synthesis in Medical Imaging (2017)

Chartsias, A., et al.: Adversarial image synthesis for unpaired multi-modal cardiac data. In: Intl Wksp on Simulation and Synthesis in Medical Imaging (2017)

work page 2017
[4]

In: ISBI (2017)

Ciompi, F., et al.: The importance of stain normalization in colorectal tissue clas- siﬁcation with convolutional networks. In: ISBI (2017)

work page 2017
[5]

H¨ ofener, H., et al.: Deep learning nuclei detection: A simple approach can deliver state-of-the-art results. Comput. Medical Imaging and Graphics (2018)

work page 2018
[6]

Hou, L., et al.: Robust histopathology image analysis: To label or to synthesize? In: CVPR (2019)

work page 2019
[7]

Scientiﬁc reports (2018)

Kapil, A., et al.: Deep semi supervised generative learning for automated tumor proportion scoring on nsclc tissue needle biopsies. Scientiﬁc reports (2018)

work page 2018
[8]

DASGAN -- Joint Domain Adaptation and Segmentation for the Analysis of Epithelial Regions in Histopathology PD-L1 Images

Kapil, A., et al.: Dasgan - joint domain adaptation and segmentation for the anal- ysis of epithelial regions in histopathology pd-l1 images. arXiv:1906.11118 (2019)

work page internal anchor Pith review Pith/arXiv arXiv 1906
[9]

Transactions on Medical Imaging (2017)

Kumar, N., et al.: A dataset and a technique for generalized nuclear segmentation for computational pathology. Transactions on Medical Imaging (2017)

work page 2017
[10]

Naylor, P., et al.: Segmentation of nuclei in histopathology images by deep regres- sion of the distance map. Trans. on Medical Imaging (2018)

work page 2018
[11]

In: MIDL (2019)

Qu, H., et al.: Weakly supervised deep nuclei segmentation using points annotation in histopathology images. In: MIDL (2019)

work page 2019
[12]

Nature Biomedical Eng

Rivenson, Y., et al.: Virtual histological staining of unlabelled tissue- autoﬂuorescence images via deep learning. Nature Biomedical Eng. (2019)

work page 2019
[13]

StainGAN: Stain Style Transfer for Digital Histological Images

Shaban, M.T., Baur, C., Navab, N., Albarqouni, S.: Staingan: Stain style transfer for digital histological images. arXiv preprint arXiv:1804.01601 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[14]

Sirinukunwattana, K., et al.: Locality sensitive deep learning for detection and classiﬁcation of nuclei in routine colon cancer histology images. Trans. Medical Imaging (2016)

work page 2016
[15]

Scientiﬁc reports (2017)

Vandenberghe, M.E., Barker, C., et al.: Relevance of deep learning to facilitate the diagnosis of her2 status in breast cancer. Scientiﬁc reports (2017)

work page 2017
[16]

In: MIDL (2018)

Zanjani, F., van der Laak, J., et al.: Histopathology stain-color normalization using deep generative models. In: MIDL (2018)

work page 2018
[17]

arXiv preprint (2017)

Zhu, J.Y., et al.: Unpaired image-to-image translation using cycle-consistent ad- versarial networks. arXiv preprint (2017)

work page 2017

[1] [1]

In: ISBI (2017)

Brieu, N., Schmidt, G.: Learning size adaptive local maxima selection for robust nuclei detection in histopathology images. In: ISBI (2017)

work page 2017

[2] [2]

Nature comm

Carstens, J., et al.: Spatial computation of intratumoral t cells correlates with survival of patients with pancreatic cancer. Nature comm. (2017)

work page 2017

[3] [3]

In: Intl Wksp on Simulation and Synthesis in Medical Imaging (2017)

Chartsias, A., et al.: Adversarial image synthesis for unpaired multi-modal cardiac data. In: Intl Wksp on Simulation and Synthesis in Medical Imaging (2017)

work page 2017

[4] [4]

In: ISBI (2017)

Ciompi, F., et al.: The importance of stain normalization in colorectal tissue clas- siﬁcation with convolutional networks. In: ISBI (2017)

work page 2017

[5] [5]

H¨ ofener, H., et al.: Deep learning nuclei detection: A simple approach can deliver state-of-the-art results. Comput. Medical Imaging and Graphics (2018)

work page 2018

[6] [6]

Hou, L., et al.: Robust histopathology image analysis: To label or to synthesize? In: CVPR (2019)

work page 2019

[7] [7]

Scientiﬁc reports (2018)

Kapil, A., et al.: Deep semi supervised generative learning for automated tumor proportion scoring on nsclc tissue needle biopsies. Scientiﬁc reports (2018)

work page 2018

[8] [8]

DASGAN -- Joint Domain Adaptation and Segmentation for the Analysis of Epithelial Regions in Histopathology PD-L1 Images

Kapil, A., et al.: Dasgan - joint domain adaptation and segmentation for the anal- ysis of epithelial regions in histopathology pd-l1 images. arXiv:1906.11118 (2019)

work page internal anchor Pith review Pith/arXiv arXiv 1906

[9] [9]

Transactions on Medical Imaging (2017)

Kumar, N., et al.: A dataset and a technique for generalized nuclear segmentation for computational pathology. Transactions on Medical Imaging (2017)

work page 2017

[10] [10]

Naylor, P., et al.: Segmentation of nuclei in histopathology images by deep regres- sion of the distance map. Trans. on Medical Imaging (2018)

work page 2018

[11] [11]

In: MIDL (2019)

Qu, H., et al.: Weakly supervised deep nuclei segmentation using points annotation in histopathology images. In: MIDL (2019)

work page 2019

[12] [12]

Nature Biomedical Eng

Rivenson, Y., et al.: Virtual histological staining of unlabelled tissue- autoﬂuorescence images via deep learning. Nature Biomedical Eng. (2019)

work page 2019

[13] [13]

StainGAN: Stain Style Transfer for Digital Histological Images

Shaban, M.T., Baur, C., Navab, N., Albarqouni, S.: Staingan: Stain style transfer for digital histological images. arXiv preprint arXiv:1804.01601 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[14] [14]

Sirinukunwattana, K., et al.: Locality sensitive deep learning for detection and classiﬁcation of nuclei in routine colon cancer histology images. Trans. Medical Imaging (2016)

work page 2016

[15] [15]

Scientiﬁc reports (2017)

Vandenberghe, M.E., Barker, C., et al.: Relevance of deep learning to facilitate the diagnosis of her2 status in breast cancer. Scientiﬁc reports (2017)

work page 2017

[16] [16]

In: MIDL (2018)

Zanjani, F., van der Laak, J., et al.: Histopathology stain-color normalization using deep generative models. In: MIDL (2018)

work page 2018

[17] [17]

arXiv preprint (2017)

Zhu, J.Y., et al.: Unpaired image-to-image translation using cycle-consistent ad- versarial networks. arXiv preprint (2017)

work page 2017