Pith. sign in

REVIEW 4 major objections 5 minor 26 references

Domain-Agnostic Learning with Anatomy-Consistent Embedding for Cross-Modality Liver Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read A single end-to-end model can learn a domain-agnostic anatomical representation from labeled CT and unlabeled multi-phasic MR, and this representation lets it outperform prior domain adaptation methods on both known-target and…

desk verdict A sensible extension of MUNIT/DADR to a genuinely new task for medical imaging, but the headline performance claims need per-fold significance tests before they can be trusted. read the letter →

arxiv 1908.10489 v1 pith:3XNUG2D7 submitted 2019-08-27 eess.IV cs.CVcs.LG

classification eess.IVcs.CVcs.LG
keywords domainagnosticlearningadaptationdisentangledrepresentationliversegmentationcross-modalityCTtoMRIadversarialtraininganatomy-consistentembedding
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Liver segmentation usually needs labeled data in each imaging modality, and domain adaptation methods assume the target modality is known in advance. This paper claims that a model called DALACE can instead learn an anatomy-only representation from labeled CT scans that carries over to unlabeled MR images from several phases at once, without knowing the target domain. It reports that DALACE beats three established baselines on both the single-target domain adaptation task (Dice score 0.847 vs 0.806 for the best baseline) and the harder domain-agnostic task (0.794 vs 0.742). A sympathetic reader would care because, if true, one trained model could segment livers in CT and the many contrast phases of MRI without retraining on each new acquisition protocol.

What carries the argument

The load-bearing mechanism is a disentangled representation split into anatomical content and modality style. Anatomy encoders map CT and MR images to feature maps; modality encoders map them to short style vectors; style-based generators with adaptive instance normalization (AdaIN) reconstruct and cross-transfer images from these codes. The Domain-Agnostic Module (DAM) adversarially pushes anatomy codes to be indistinguishable across modalities, while the Anatomy-Preserving Module (APM) adds a U-Net segmentation consistency loss so the anatomy code retains liver structure. These two supervision modules are what distinguish DALACE from prior disentanglement methods.

What would settle it

Take the model trained as in the paper and test it on an MR sequence not used in training, such as T2-weighted or a delayed post-contrast phase, or on multi-phasic MR from a different hospital; if its Dice score falls to near the lower bound or no longer beats DADR, the claim of domain-agnostic anatomy-consistent embedding is falsified for arbitrary target domains.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that disentangling each image into an anatomy code and a modality code, then forcing the anatomy code to be domain-agnostic and anatomy-preserving through two supervised modules, yields a shared embedding that supports liver segmentation across CT and multi-phasic MR. The authors claim this is what allows the model to work in the domain-agnostic setting, where target data come from a mixture of unknown domains, and to beat CycleGAN, TD-GAN, and DADR in both DA and DAL experiments. They further claim that in joint learning, training on both modalities with masks lets DALACE exceed separate fully supervised U-Nets on each modality (0.911 CT and 0.907 MR versus 0.901 and 0.869).

Load-bearing premise

The load-bearing premise is that the multi-phasic MR data from 20 patients at one local medical center spans the variety of unseen MR domains promised by domain-agnostic learning; if the model meets a genuinely different sequence or external site, the reported advantage may not persist.

Editorial extensions

If this is right

  • In the single-target domain adaptation task, DALACE reports a Dice score of 0.847, ahead of CycleGAN at 0.721, TD-GAN at 0.793, and DADR at 0.806.
  • In the domain-agnostic task over three MR phases, DALACE reports 0.794 versus 0.522, 0.719, and 0.742 for the same baselines.
  • In joint learning with labeled CT and MR, DALACE reaches 0.911 on CT and 0.907 on MR, exceeding single-modality fully supervised U-Nets (0.901 and 0.869).
  • Ablations show that removing the disentanglement component drops DA performance from 0.847 to 0.793 and DAL from 0.794 to 0.719.
  • Because segmentation uses only the anatomy code, the model can generate domain-agnostic images and transfer modality rendering between CT and MR phases, which suggests the learned representation is interpretable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the anatomy-modality split is as clean as the visualizations suggest, the same architecture should transfer to other organs or structures, since the segmentation target is carried by the anatomy code rather than by any modality-specific texture.
  • A direct testable extension is to expose DALACE to an unseen MR sequence, such as T2-weighted imaging, and measure whether the Dice score degrades gracefully, which would distinguish true domain-agnostic behavior from memorization of the three training phases.
  • The joint-learning result hints that a single shared model could replace per-modality segmentation pipelines in clinical practice, but that would require validation on larger, multi-site datasets with paired outcomes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents DALACE, an end-to-end disentangled representation learning framework for unsupervised cross-modality liver segmentation from CT to multi-phasic MR. The model uses separate anatomy and modality encoders, style-based generators with AdaIN, a Domain-Agnostic Module (DAM) for adversarial alignment, and an Anatomy-Preserving Module (APM) to encourage semantic consistency through a U-Net segmentation network. Experiments on 130 CT (LiTS) and 20 multi-phasic MR patients with 5-fold cross-validation report improved DSC over CycleGAN, TD-GAN, and DADR for domain adaptation (0.847 vs 0.806), domain-agnostic learning (0.794 vs 0.742), and joint learning (CT 0.911, MR 0.907 vs fully-supervised single-modality U-Nets). Ablation studies and visualizations of domain-agnostic images support the disentanglement claims.

Significance. If the reported results hold, DALACE offers a practical step toward domain-agnostic medical image segmentation by combining disentanglement with explicit anatomy-preserving supervision. The paper clearly formulates the DAL task for medical imaging, includes a useful ablation analysis, and provides qualitative evidence of disentanglement. However, the evidence base is narrow: the MR cohort comprises only 20 patients from a single center, the cross-validation is reported only as mean and standard deviation over five folds with no significance testing, and no external validation is provided. These limitations leave the headline superiority claims under-supported, though the underlying method and evaluation framework are sound and the identified gaps appear addressable.

major comments (4)
  1. [§4.2–4.4, Tables 1–3] Report per-fold DSC values and paired statistical significance tests (e.g., paired t-test or Wilcoxon signed-rank test over the five folds) for every claimed improvement. The reported differences are small relative to the dispersion: DA 0.847±0.041 vs DADR 0.806±0.035, DAL 0.794±0.044 vs DADR 0.742±0.045, and joint MR 0.907±0.049 vs MR-trained U-Net 0.869±0.044. With n=5 folds, these gaps could easily arise from fold-to-fold variability, so the abstract's claim that DALACE 'outperformed' the baselines is not statistically established without test statistics.
  2. [§4.1] Specify whether the 5-fold partitions are made at the patient level or the slice level. If slices, rather than patients, are randomly assigned to folds, the same patient's data can appear in both the training and testing folds, causing data leakage and inflated DSC scores. This is especially important given the small MR cohort (20 patients) and the use of multiple slices per patient; patient-level splitting must be confirmed.
  3. [§4.3] The DAL experiments operationalize 'arbitrary target domains' using only pre-contrast, arterial, and venous phase MR images from 20 patients acquired at a single medical center. This is a narrow slice of possible target domains (the same scanner, site, and patient population). Please temper the generalization claims accordingly and discuss external validation (e.g., different MR sequences, scanners, or sites) as necessary future work; alternatively, provide a held-out patient cohort from a different acquisition setting.
  4. [§6.2, Table 4] The ablation table is difficult to interpret. The text says APM and DAM are 'taken out respectively,' which implies two ablation variants, but Table 4 appears to show three conditions (no modules, DAM only, both). The APM-only condition is not reported, and the row/column labels are ambiguous (the header 'APM DAM DSC for DADSC for DAL' is garbled). Please provide a complete ablation matrix with per-fold DSC values and clearly labeled rows.
minor comments (5)
  1. [§3.3] 'Thhird' is a typo for 'Third' in the optimization description.
  2. [§3.2.2] The variables x_MR^a, \hat{M}_MR^a, x_CT^a, and \hat{M}_CT^a are used in Eq. (8) but not explicitly defined in the text; define them when they are introduced.
  3. [§4.1] Report the total number of slices and the number of slices per patient for both CT and MR to improve reproducibility.
  4. [Figure 7 caption] 'the first row from right to left' is confusing; use 'left to right' if that is the intended ordering, or clarify the ordering explicitly.
  5. [§5.2] The paper states that modality information is 'erased' in domain-agnostic images, but no quantitative measure of domain invariance is provided; consider reporting a domain classifier accuracy on the anatomy codes or domain-agnostic images.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported DA/DAL DSC values come from held-out MR masks, and the DALACE losses enforce consistency rather than fitting the target labels.

full rationale

The paper's derivation chain is self-contained. The DALACE model is defined by explicit losses (Eqs. 1-8): reconstruction, latent-code cycle consistency, cross-domain adversarial alignment, CT-supervised cross-entropy, and pair-wise adversarial anatomy preservation. The DA/DAL experiments train on labeled CT and unlabeled MR, with one MR fold held out for testing and the test masks used only for evaluation (Sections 4.2 and 4.3). The APM uses CT ground-truth masks and an adversarial pair loss to enforce anatomy consistency; this is a domain-alignment mechanism, not a circular definition of the target metric. DADR is cited as both an inspiration and a baseline, but the baseline comparison is an independent evaluation on the same folds; self-citation alone is not load-bearing. The ablation in Section 6.2 identifies the APM-only variant as 'essentially DADR' and reports identical numbers, but this is an explicit ablation check, not a prediction derived from the baseline. Any concern about small margins, fold variance, or single-center data is a statistical/external-validity risk, not circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim relies on several unproven domain assumptions common in deep learning papers: that the disentanglement architecture generalizes, that adversarial training is stable, and that the small, single-center MR dataset is representative. No invented physical entities are introduced. The loss weights are hand-set, not fitted to the test set.

free parameters (3)
  • Loss weight α (reconstruction) = 2.5
    Chosen by hand in Section 3.3; balances image and latent reconstruction losses, affects the learned embedding.
  • Loss weight β (latent) = 0.01
    Chosen by hand in Section 3.3; controls the weight of latent reconstruction relative to image reconstruction.
  • Modality code dimension = 8
    Set in Section 3.3; the length of the modality vector; affects capacity for representing modality information.
assumptions (5)
  • domain assumption The disentanglement architecture (based on MUNIT) can separate anatomy and modality with the given reconstruction and adversarial losses.
    The entire method rests on this assumption; it is not proven but empirically evaluated.
  • domain assumption The adversarial training in DAM and APM converges to a useful equilibrium.
    Adversarial training is known to be unstable; the paper assumes it works with the chosen hyperparameters.
  • domain assumption The segmentation module S, trained on CT anatomical images, provides a valid anatomy-preserving signal for MR via the pair discriminator.
    APM assumes that a CT-trained segmenter can guide MR anatomy preservation without ground truth MR masks.
  • domain assumption The three MR phases from 20 patients represent the target domain distribution adequately.
    Underpins the DAL claim; if this fails, the reported results may not generalize.
  • standard math Standard backpropagation and stochastic optimization work as intended.
    No unusual math is used; relies on conventional deep learning machinery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Domain-Agnostic Learning with Anatomy-Consistent Embedding for Cross-Modality Liver Segmentation." pith.science (2026). https://pith.science/paper/3XNUG2D7

@misc{pith2026190810489,
  author       = {Pith},
  title        = {Pith review of: Domain-Agnostic Learning with Anatomy-Consistent Embedding for Cross-Modality Liver Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3XNUG2D7}},
  note         = {Machine review of arXiv:1908.10489}
}
read the original abstract

Domain Adaptation (DA) has the potential to greatly help the generalization of deep learning models. However, the current literature usually assumes to transfer the knowledge from the source domain to a specific known target domain. Domain Agnostic Learning (DAL) proposes a new task of transferring knowledge from the source domain to data from multiple heterogeneous target domains. In this work, we propose the Domain-Agnostic Learning framework with Anatomy-Consistent Embedding (DALACE) that works on both domain-transfer and task-transfer to learn a disentangled representation, aiming to not only be invariant to different modalities but also preserve anatomical structures for the DA and DAL tasks in cross-modality liver segmentation. We validated and compared our model with state-of-the-art methods, including CycleGAN, Task Driven Generative Adversarial Network (TD-GAN), and Domain Adaptation via Disentangled Representations (DADR). For the DA task, our DALACE model outperformed CycleGAN, TD-GAN ,and DADR with DSC of 0.847 compared to 0.721, 0.793 and 0.806. For the DAL task, our model improved the performance with DSC of 0.794 from 0.522, 0.719 and 0.742 by CycleGAN, TD-GAN, and DADR. Further, we visualized the success of disentanglement, which added human interpretability of the learned meaningful representations. Through ablation analysis, we specifically showed the concrete benefits of disentanglement for downstream tasks and the role of supervision for better disentangled representation with segmentation consistency to be invariant to domains with the proposed Domain-Agnostic Module (DAM) and to preserve anatomical information with the proposed Anatomy-Preserving Module (APM).

Figures

Figures reproduced from arXiv: 1908.10489 by the authors.

Figure 1
Figure 1. Schematic diagram of the domain adaptation task and the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (Best viewed in color) The end-to-end DALACE pipeline to learn domain-agnostic anatomical embeddings. The solid line shows [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Domain-Agnostic Module, which encourages the em [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Examples of images from different modalities, from left [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Two examples of DA task for cross-modality liver seg [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Two sets of examples of domain-agnostic images. In [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: CT images are transferred to multi-phasic MR images in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

  1. [1]

    Almahairi, S

    A. Almahairi, S. Rajeshwar, A. Sordoni, P. Bachman, and A. Courville. Augmented cyclegan: Learning many-to-many mappings from unpaired data. In International Conference on Machine Learning, pages 195–204, 2018

  2. [2]

    X. Chen, Y . Duan, R. Houthooft, J. Schulman, I. Sutskever, and P. Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In Advances in neural information processing systems , pages 2172–2180, 2016

  3. [3]

    Christ, F

    P. Christ, F. Ettlinger, F. Gr ¨un, J. Lipkova, and G. Kaissis. Lits-liver tumor segmentation challenge. ISBI and MICCAI, 2017

  4. [4]

    N. C. Codella, C.-C. Lin, A. Halpern, M. Hind, R. Feris, and J. R. Smith. Collaborative human-ai (chai): Evidence- based interpretable melanoma classification in dermoscopic images. In Understanding and Interpreting Machine Learn- ing in Medical Image Computing Applications , pages 97–

  5. [5]

    Q. Dou, C. Ouyang, C. Chen, H. Chen, B. Glocker, X. Zhuang, and P.-A. Heng. Pnp-adanet: Plug-and-play adversarial domain adaptation network with a benchmark at cross-modality cardiac segmentation. arXiv preprint arXiv:1812.07907, 2018

  6. [6]

    L. H. Gilpin, D. Bau, B. Z. Yuan, A. Bajwa, M. Specter, and L. Kagal. Explaining explanations: An overview of inter- pretability of machine learning. In 2018 IEEE 5th Interna- tional Conference on data science and advanced analytics (DSAA), pages 80–89. IEEE, 2018

  7. [7]

    Higgins, D

    I. Higgins, D. Amos, D. Pfau, S. Racaniere, L. Matthey, D. Rezende, and A. Lerchner. Towards a definition of dis- entangled representations. arXiv preprint arXiv:1812.02230, 2018

  8. [8]

    Higgins, L

    I. Higgins, L. Matthey, A. Pal, C. Burgess, X. Glorot, M. Botvinick, S. Mohamed, and A. Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR, 2(5):6, 2017

Show all 26 references
  1. [9]

    Huang, M.-Y

    X. Huang, M.-Y . Liu, S. Belongie, and J. Kautz. Multimodal unsupervised image-to-image translation. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 172–189, 2018

  2. [10]

    Jiang, Y .-C

    J. Jiang, Y .-C. Hu, N. Tyagi, P. Zhang, A. Rimner, G. S. Mageras, J. O. Deasy, and H. Veeraraghavan. Tumor-aware, adversarial domain adaptation from ct to mri for lung cancer segmentation. In International Conference on Medical Im- age Computing and Computer-Assisted Interven...

  3. [11]

    Karras, S

    T. Karras, S. Laine, and T. Aila. A style-based generator ar- chitecture for generative adversarial networks. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4401–4410, 2019

  4. [12]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

  5. [13]

    X. Li, N. C. Dvornek, Y . Zhou, J. Zhuang, P. Ventola, and J. S. Duncan. Efficient interpretation of deep learning models using graph structure and cooperative game theory: Applica- tion to asd biomarker discovery. InInternational Conference on Information Processing in Medical...

  6. [14]

    Locatello, S

    F. Locatello, S. Bauer, M. Lucic, G. Raetsch, S. Gelly, B. Sch ¨olkopf, and O. Bachem. Challenging common as- sumptions in the unsupervised learning of disentangled rep- resentations. In International Conference on Machine Learn- ing, pages 4114–4124, 2019

  7. [15]

    M. Long, Y . Cao, J. Wang, and M. I. Jordan. Learning transferable features with deep adaptation networks. arXiv preprint arXiv:1502.02791, 2015

  8. [16]

    Y . Luo, L. Zheng, T. Guan, J. Yu, and Y . Yang. Taking a closer look at domain shift: Category-level adversaries for semantics consistent domain adaptation. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2507–2516, 2019

  9. [17]

    M. R. Oliva and S. Saini. Liver cancer imaging: role of ct, mri, us and pet. Cancer imaging, 4(Spec No A):S42, 2004

  10. [18]

    X. Peng, Z. Huang, X. Sun, and K. Saenko. Domain agnostic learning with disentangled representations. In International Conference on Machine Learning, pages 5102–5112, 2019

  11. [19]

    P. Z. Ramirez, A. Tonioni, and L. Di Stefano. Exploiting se- mantics in adversarial training for image-level domain adap- tation. In 2018 IEEE International Conference on Image Processing, Applications and Systems (IPAS), pages 49–54. IEEE, 2018

  12. [20]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convo- lutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention , pages 234–241. Springer, 2015

  13. [21]

    Tzeng, J

    E. Tzeng, J. Hoffman, T. Darrell, and K. Saenko. Simultane- ous deep transfer across domains and tasks. In Proceedings of the IEEE International Conference on Computer Vision , pages 4068–4076, 2015

  14. [22]

    Tzeng, J

    E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversar- ial discriminative domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion, pages 7167–7176, 2017

  15. [23]

    Wang and W

    M. Wang and W. Deng. Deep visual domain adaptation: A survey. Neurocomputing, 312:135–153, 2018

  16. [24]

    J. Yang, N. C. Dvornek, F. Zhang, J. Chapiro, M. Lin, and J. S. Duncan. Unsupervised domain adaptation via disen- tangled representations: Application to cross-modality liver segmentation. arXiv preprint arXiv:1907.13590, 2019

  17. [25]

    Zhang, S

    Y . Zhang, S. Miao, T. Mansi, and R. Liao. Task driven gener- ative modeling for unsupervised domain adaptation: Appli- cation to x-ray image segmentation. InInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention, pages 599–607. Springer, 2018

  18. [26]

    J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros. Unpaired image- to-image translation using cycle-consistent adversarial net- works. In Proceedings of the IEEE international conference on computer vision, pages 2223–2232, 2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.