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

Abraham Silva; Ansh Kapil; Guenter Schmidt; Keith Steele; Marlon Rebelatto; Nicolas Brieu; Simon Lanzmich; Tobias Wiestler

arxiv: 1906.11118 · v1 · pith:5RSJEFM6new · submitted 2019-06-26 · 📡 eess.IV · cs.CV

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

Ansh Kapil , Tobias Wiestler , Simon Lanzmich , Abraham Silva , Keith Steele , Marlon Rebelatto , Guenter Schmidt , Nicolas Brieu This is my paper

Pith reviewed 2026-05-25 15:11 UTC · model grok-4.3

classification 📡 eess.IV cs.CV

keywords domain adaptationhistopathology segmentationPD-L1Cytokeratintumor cell scoreunpaired image translationdeep learningepithelial regions

0 comments

The pith

An end-to-end network segments tumor epithelium in PD-L1 images by adapting labels from unpaired Cytokeratin stains.

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

The paper presents a deep learning approach for segmenting tumor epithelium in PD-L1 stained histopathology images. It leverages semi-automatically labeled images from a Cytokeratin stain domain through unpaired image-to-image translation inside a joint network. This setup avoids the need for manual annotations directly on PD-L1 slides or for physical serial sections and re-staining. The method further classifies PD-L1 positive versus negative epithelial regions to compute an automated Tumor Cell score. A sympathetic reader would care because it lowers the cost and labor of quantifying immune markers in cancer tissue samples.

Core claim

We introduce an end-to-end trainable network that jointly segments tumor epithelium on PD-L1 while leveraging unpaired image-to-image translation between CK and PD-L1, therefore completely bypassing the need for serial sections or re-staining of slides. Extending the method to differentiate between PD-L1 positive and negative tumor epithelium regions enables the automated estimation of the PD-L1 Tumor Cell (TC) score. Quantitative experimental results demonstrate the accuracy of our approach against state-of-the-art segmentation methods.

What carries the argument

DASGAN, the joint domain adaptation and segmentation network that performs unpaired image-to-image translation to transfer epithelial labels from the CK domain to the PD-L1 domain.

If this is right

Epithelial segmentation becomes possible on PD-L1 images without any manual labeling in that stain domain.
PD-L1 Tumor Cell scores can be estimated automatically once positive and negative epithelial regions are distinguished.
Analysis of tumor microenvironments can occur without acquiring serial sections or performing additional staining.
The joint training reaches accuracy comparable to supervised methods that require target-domain annotations.

Where Pith is reading between the lines

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

The same joint adaptation-segmentation pattern could transfer to other stain pairs where one domain has easier semi-automatic labeling.
If translation fidelity holds across staining batches, the approach may cut annotation effort for additional cancer biomarkers.
Testing on multi-site datasets with natural staining variation would reveal whether the transferred labels remain reliable at scale.

Load-bearing premise

Unpaired translation between CK and PD-L1 images keeps the spatial layout and identity of epithelial regions intact enough that labels moved from CK remain correct when used to train segmentation on PD-L1.

What would settle it

If a model trained only on labels transferred via the unpaired translation shows markedly lower segmentation accuracy on manually annotated PD-L1 test images than a model trained directly on PD-L1 labels, the central claim does not hold.

Figures

Figures reproduced from arXiv: 1906.11118 by Abraham Silva, Ansh Kapil, Guenter Schmidt, Keith Steele, Marlon Rebelatto, Nicolas Brieu, Simon Lanzmich, Tobias Wiestler.

**Figure 1.** Figure 1: Synthetic (a) and real (b) PD-L1 datasets generated from the semi-automated segmentation of CK images and manual annotations, respectively. (c) DASGAN model for joint domain adaptation and semantic segmentation. NB: the two cycle losses between the real and the cyclic images are not displayed for clarity purposes. and more particularly towards deep semantic segmentation networks [9]. The prerequisite data… view at source ↗

**Figure 2.** Figure 2: Segmentation accuracy on the validation set, for the three baseline models (in gray, blue and yellow) and for the proposed DASGAN model (in orange) under the condition (i) of low availability of manual annotations on real PD-L1 images. F1 scores are reported for each class of interest - epithelium (TC), epithelium positive (TC+), epithelium negative (TC-) and Other, together with their average score Avg. i… view at source ↗

**Figure 3.** Figure 3: Segmentation accuracy (avg. f1 score) on the unseen validation set, for the baseline model trained only on real PD-L1 samples (blue) and the proposed DASGAN model trained on both real and synthetic PD-L1 samples (orange), for increasing availability (i)-(ii)-(iii) of manual annotations on real PD-L1 images. 3 Experiments and Results 3.1 Cytokeratin and PD-L1 Datasets The training set consists of NCK = 56 … view at source ↗

**Figure 4.** Figure 4: (a) Example of negative (red) and positive (blue) epithelial regions as well as of non-epithelial regions (green) segmented by the proposed DASGAN model. (b) Bar plot showing mean and standard deviation of the TC scores estimated by the proposed approach on unseen cases (n=704), on the following TC score bins: T C < 1, 1 <= T C < 10, 10n <= T C < 10(n + 1) for 1 < n < 10 and T C = 100. The inverted histogr… view at source ↗

read the original abstract

The analysis of the tumor environment on digital histopathology slides is becoming key for the understanding of the immune response against cancer, supporting the development of novel immuno-therapies. We introduce here a novel deep learning solution to the related problem of tumor epithelium segmentation. While most existing deep learning segmentation approaches are trained on time-consuming and costly manual annotation on single stain domain (PD-L1), we leverage here semi-automatically labeled images from a second stain domain (Cytokeratin-CK). We introduce an end-to-end trainable network that jointly segment tumor epithelium on PD-L1 while leveraging unpaired image-to-image translation between CK and PD-L1, therefore completely bypassing the need for serial sections or re-staining of slides. Extending the method to differentiate between PD-L1 positive and negative tumor epithelium regions enables the automated estimation of the PD-L1 Tumor Cell (TC) score. Quantitative experimental results demonstrate the accuracy of our approach against state-of-the-art segmentation methods.

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

Joint unpaired translation plus segmentation for PD-L1 TC scoring is a sensible application but rests on an unverified assumption that label transfer preserves epithelial boundaries.

read the letter

The paper's main move is an end-to-end network that translates unpaired CK images to the PD-L1 domain and uses the resulting images plus transferred labels to train epithelium segmentation, then extends it to positive/negative tumor cells for automated TC scoring. This is new as a combined trainable system for this exact workflow, even though the CycleGAN-style translation and segmentation pieces are not original on their own. It does a clean job naming the annotation cost problem in immuno-oncology imaging and showing how a second stain domain can be leveraged without serial sections or re-staining. The framing is practical and the claim of bypassing manual PD-L1 labels is stated clearly. The soft spot is exactly the one the stress-test flags. Standard unpaired translation only enforces cycle consistency and realism; nothing in the description adds a term that keeps epithelial region boundaries or semantic masks aligned after translation. Stain texture differences can easily make the generator shift or expand those regions, so the CK labels stop being valid supervision for the PD-L1 side. The abstract gives no numbers, no ablation on the translation step, and no paired validation that the transferred masks match, which leaves the central claim under-supported. No circularity or invented quantities appear. This is for people already working on stain adaptation or digital pathology pipelines for PD-L1. A reader facing the same annotation bottleneck would get a useful concrete example even if the method needs more checks. It deserves peer review so the experiments and any boundary-preservation details can be looked at properly.

Referee Report

1 major / 1 minor

Summary. The manuscript presents DASGAN, an end-to-end trainable network that jointly performs unpaired image-to-image translation between CK and PD-L1 stain domains and segments tumor epithelium in PD-L1 images. It uses semi-automatically labeled CK images to supervise PD-L1 segmentation without serial sections or re-staining, extends the approach to distinguish PD-L1 positive/negative epithelium for automated TC scoring, and reports quantitative superiority over state-of-the-art segmentation methods.

Significance. If the central claim holds, the work could meaningfully reduce annotation costs in digital pathology by exploiting multiple stain domains for label transfer. The joint training formulation is a potential strength if semantic fidelity is demonstrated.

major comments (1)

[Abstract] Abstract (paragraph on the joint network): the central claim that unpaired translation between CK and PD-L1 domains enables valid label transfer for PD-L1 segmentation rests on the unstated assumption that epithelial region boundaries and semantics are preserved; standard unpaired I2I (CycleGAN-style) losses enforce only cycle-consistency and adversarial realism with no explicit semantic-consistency or boundary-preservation term, rendering the supervision signal potentially misaligned and the joint objective ill-posed. This requires either an auxiliary loss, paired validation set, or quantitative boundary-alignment metric in the methods.

minor comments (1)

[Abstract] Abstract: no dataset sizes, validation splits, or numerical metrics are provided despite the claim of 'quantitative experimental results'; a brief summary of these should appear even in the abstract for a methods paper.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and the opportunity to address this important point on the validity of the label transfer mechanism. We respond to the major comment below.

read point-by-point responses

Referee: [Abstract] Abstract (paragraph on the joint network): the central claim that unpaired translation between CK and PD-L1 domains enables valid label transfer for PD-L1 segmentation rests on the unstated assumption that epithelial region boundaries and semantics are preserved; standard unpaired I2I (CycleGAN-style) losses enforce only cycle-consistency and adversarial realism with no explicit semantic-consistency or boundary-preservation term, rendering the supervision signal potentially misaligned and the joint objective ill-posed. This requires either an auxiliary loss, paired validation set, or quantitative boundary-alignment metric in the methods.

Authors: The joint formulation in DASGAN incorporates the segmentation objective directly into the end-to-end training, so that the segmentation loss on translated images (supervised by CK labels) functions as the semantic-consistency term that encourages preservation of epithelial boundaries. This is described in the methods (Section 3) and is the key distinction from a standard CycleGAN. We agree the abstract paragraph does not make this explicit and will revise it to state that the segmentation loss provides the required semantic guidance. The experimental section already reports quantitative boundary-sensitive metrics (Dice, Hausdorff) on held-out PD-L1 data that serve as the empirical validation of alignment; no paired validation set was available by design. revision: partial

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper introduces a joint segmentation and unpaired I2I translation network (DASGAN) that transfers labels from CK to PD-L1 domains. No equations, fitted parameters, or predictions are shown that reduce by construction to the inputs (no self-definitional loops, no fitted-input-called-prediction, no load-bearing self-citation chains). The central claim rests on the architectural combination and empirical validation against external benchmarks rather than renaming or re-deriving its own assumptions. Standard unpaired translation (e.g., cycle-consistency) is invoked as an external mechanism; any validity issues concern empirical performance, not circular derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review based solely on abstract; no explicit free parameters, axioms, or invented entities are stated. The method implicitly rests on standard deep-learning assumptions about unpaired translation preserving semantic labels.

axioms (1)

domain assumption Unpaired image-to-image translation between CK and PD-L1 domains preserves epithelial region boundaries and PD-L1 positivity information.
This premise is required for labels from the CK domain to train a valid PD-L1 segmenter.

pith-pipeline@v0.9.0 · 5725 in / 1316 out tokens · 32205 ms · 2026-05-25T15:11:10.010763+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Domain Adaptation-based Augmentation for Weakly Supervised Nuclei Detection
eess.IV 2019-07 unverdicted novelty 4.0

Domain adaptation via stain normalization and unpaired translation generates synthetic labeled target images to train nuclei detection networks, reported superior to fully supervised intra-domain baselines.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · cited by 1 Pith paper · 4 internal anchors

[1]

Clinical cancer research 14(16), 5220–5227 (2008)

Al-Shibli, K.I., Donnem, T., Al-Saad, S., Persson, M., Bremnes, R.M., Busund, L.T.: Prognostic eﬀect of epithelial and stromal lymphocyte inﬁltration in non– small cell lung cancer. Clinical cancer research 14(16), 5220–5227 (2008)

work page 2008
[2]

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
[3]

Scientiﬁc reports 7, 46450 (2017)

Cruz-Roa, A., et al.: Accurate and reproducible invasive breast cancer detection in whole-slide images: A deep learning approach for quantifying tumor extent. Scientiﬁc reports 7, 46450 (2017)

work page 2017
[4]

Nature reviews Clinical oncology 14(12), 717 (2017)

Fridman, W.H.e.a.: The immune contexture in cancer prognosis and treatment. Nature reviews Clinical oncology 14(12), 717 (2017)

work page 2017
[5]

In: MICCAI (2018)

Jiang, J., et al.: Tumor-aware, adversarial domain adaptation from ct to mri for lung cancer segmentation. In: MICCAI (2018)

work page 2018
[6]

Scientiﬁc reports (2018)

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

work page 2018
[7]

Scientiﬁc reports 6, 26286 (2016)

Litjens, G., et al.: Deep learning as a tool for increased accuracy and eﬃciency of histopathological diagnosis. Scientiﬁc reports 6, 26286 (2016)

work page 2016
[8]

Spectral Normalization for Generative Adversarial Networks

Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[9]

In: ICCV

Noh, H., Hong, S., Han, B.: Learning deconvolution network for semantic segmen- tation. In: ICCV. pp. 1520–1528 (2015)

work page 2015
[10]

Conditional Image Synthesis With Auxiliary Classifier GANs

Odena, A., Olah, C., Shlens, J.: Conditional image synthesis with auxiliary classiﬁer gans. arXiv preprint arXiv:1610.09585 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016
[11]

Diagnostic pathology 11(1), 95 (2016)

Rebelatto, M.C., et al.: Development of a programmed cell death ligand-1 immuno- histochemical assay validated for analysis of non-small cell lung cancer and head and neck squamous cell carcinoma. Diagnostic pathology 11(1), 95 (2016)

work page 2016
[12]

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
[13]

IEEE trans- actions on medical imaging 37(9), 2126–2136 (2018)

Tellez, D., et al.: Whole-slide mitosis detection in h&e breast histology using phh3 as a reference to train distilled stain-invariant convolutional networks. IEEE trans- actions on medical imaging 37(9), 2126–2136 (2018)

work page 2018
[14]

Self-Attention Generative Adversarial Networks

Zhang, H., Goodfellow, I., Metaxas, D., Odena, A.: Self-attention generative ad- versarial networks. arXiv preprint arXiv:1805.08318 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[15]

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]

Clinical cancer research 14(16), 5220–5227 (2008)

Al-Shibli, K.I., Donnem, T., Al-Saad, S., Persson, M., Bremnes, R.M., Busund, L.T.: Prognostic eﬀect of epithelial and stromal lymphocyte inﬁltration in non– small cell lung cancer. Clinical cancer research 14(16), 5220–5227 (2008)

work page 2008

[2] [2]

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

[3] [3]

Scientiﬁc reports 7, 46450 (2017)

Cruz-Roa, A., et al.: Accurate and reproducible invasive breast cancer detection in whole-slide images: A deep learning approach for quantifying tumor extent. Scientiﬁc reports 7, 46450 (2017)

work page 2017

[4] [4]

Nature reviews Clinical oncology 14(12), 717 (2017)

Fridman, W.H.e.a.: The immune contexture in cancer prognosis and treatment. Nature reviews Clinical oncology 14(12), 717 (2017)

work page 2017

[5] [5]

In: MICCAI (2018)

Jiang, J., et al.: Tumor-aware, adversarial domain adaptation from ct to mri for lung cancer segmentation. In: MICCAI (2018)

work page 2018

[6] [6]

Scientiﬁc reports (2018)

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

work page 2018

[7] [7]

Scientiﬁc reports 6, 26286 (2016)

Litjens, G., et al.: Deep learning as a tool for increased accuracy and eﬃciency of histopathological diagnosis. Scientiﬁc reports 6, 26286 (2016)

work page 2016

[8] [8]

Spectral Normalization for Generative Adversarial Networks

Miyato, T., Kataoka, T., Koyama, M., Yoshida, Y.: Spectral normalization for generative adversarial networks. arXiv preprint arXiv:1802.05957 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[9] [9]

In: ICCV

Noh, H., Hong, S., Han, B.: Learning deconvolution network for semantic segmen- tation. In: ICCV. pp. 1520–1528 (2015)

work page 2015

[10] [10]

Conditional Image Synthesis With Auxiliary Classifier GANs

Odena, A., Olah, C., Shlens, J.: Conditional image synthesis with auxiliary classiﬁer gans. arXiv preprint arXiv:1610.09585 (2016)

work page internal anchor Pith review Pith/arXiv arXiv 2016

[11] [11]

Diagnostic pathology 11(1), 95 (2016)

Rebelatto, M.C., et al.: Development of a programmed cell death ligand-1 immuno- histochemical assay validated for analysis of non-small cell lung cancer and head and neck squamous cell carcinoma. Diagnostic pathology 11(1), 95 (2016)

work page 2016

[12] [12]

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

[13] [13]

IEEE trans- actions on medical imaging 37(9), 2126–2136 (2018)

Tellez, D., et al.: Whole-slide mitosis detection in h&e breast histology using phh3 as a reference to train distilled stain-invariant convolutional networks. IEEE trans- actions on medical imaging 37(9), 2126–2136 (2018)

work page 2018

[14] [14]

Self-Attention Generative Adversarial Networks

Zhang, H., Goodfellow, I., Metaxas, D., Odena, A.: Self-attention generative ad- versarial networks. arXiv preprint arXiv:1805.08318 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[15] [15]

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