{"id":"9edb286d-21bf-4ecd-b56f-ab6adcd6c369","arxiv_id":"1908.10489","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DALACE disentangles anatomy from modality to adapt liver segmentation from CT to multi-phase unlabeled MRI, beating CycleGAN, TD-GAN, and DADR.","lead":"A deep learning model called DALACE learns to segment livers in CT scans and transfers that ability to unlabeled MRI scans across multiple phases. In a small study with 20 MRI patients, it outperformed three existing domain-adaptation methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported outperformance rests on 5-fold mean DSC differences comparable to fold variance; without significance testing or external validation, superiority over DADR and fully-supervised U-Net is not established.","rationale":"The reader's weakest assumption concerns external validity: the multi-phasic MR data from 20 patients at one center may not represent arbitrary target domains. This is a valid readiness concern, but the more immediate load-bearing issue is internal statistical support for the central claim. The report does not provide significance testing, confidence intervals, or per-fold results, and the DSS differences are comparable to the reported standard deviations. If those differences are not robust across folds, the entire 'outperforms' claim is not supported, regardless of external generalizability. I do not see circularity in the architecture, and the ablation study is a useful check, but it does not resolve the statistical question. The reader's conditional verdict should remain unchanged: acceptance should require either a statistically grounded comparison on this dataset or additional external validation on unseen MR sequences from other sites.","tokens_in":11603,"tokens_out":5830,"duration_ms":67545,"concrete_test":"Request or reconstruct the per-fold DSC values for DALACE and each baseline in Tables 1-3, then run a paired Wilcoxon signed-rank test (or permutation test) on the five fold differences for (i) DALACE vs DADR in DA, (ii) DALACE vs DADR in DAL, and (iii) DALACE vs MR-trained U-Net in joint MR. If any two-sided p-value is >=0.05, or if the median difference is not consistently positive across folds, the claim of outperformance should be weakened to a numerical improvement in this sample rather than an established advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DALACE outperforms CycleGAN, TD-GAN, and DADR in DA/DAL and exceeds fully-supervised single-modality U-Nets in joint learning. The evidence is mean DSC over five cross-validation folds on MR data from 20 patients at one site (Section 4.1). The reported margins are small relative to the reported dispersion: DA +0.041 over DADR (0.847 vs 0.806, std 0.041 vs 0.035), DAL +0.052 (0.794 vs 0.742, std 0.044 vs 0.045), and joint MR +0.038 over an MR-trained U-Net (0.907 vs 0.869, std 0.049 vs 0.044). With n=5 folds and no paired significance test, these differences cannot be distinguished from fold-to-fold noise. The DAL generalization claim is further unsecured because the target 'arbitrary domains' are represented only by three MR phases from the same 20-patient single-center dataset, and the cross-validation split is not specified as patient-level, leaving open the possibility of slice-level leakage. Unless per-fold scores and a paired test are provided, the headline claim of superiority is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":11999,"tokens_out":6511,"duration_ms":61025,"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":[{"comment":"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.","section":"§4.2–4.4, Tables 1–3"},{"comment":"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.","section":"§4.1"},{"comment":"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.","section":"§4.3"},{"comment":"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.","section":"§6.2, Table 4"}],"minor_comments":[{"comment":"'Thhird' is a typo for 'Third' in the optimization description.","section":"§3.3"},{"comment":"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.","section":"§3.2.2"},{"comment":"Report the total number of slices and the number of slices per patient for both CT and MR to improve reproducibility.","section":"§4.1"},{"comment":"'the first row from right to left' is confusing; use 'left to right' if that is the intended ordering, or clarify the ordering explicitly.","section":"Figure 7 caption"},{"comment":"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.","section":"§5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's comparisons include the authors' own prior DADR method as a baseline; while this is standard practice, the novelty over DADR is somewhat incremental (the DAM and APM modules). The DAL task formulation and the joint learning experiment are the most novel contributions. The statistical and data-composition issues raised in the major comments are the main barriers to acceptance; they are addressable within the scope of the manuscript. I do not see evidence of circular evaluation or data leakage by design, but the split-level ambiguity must be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the DALACE liver segmentation paper. The short version: the task formulation is genuinely new for medical imaging, and the architecture is a sensible extension of MUNIT/DADR with two well-motivated modules. But the headline numbers do not establish superiority — the reported gains are within fold-to-fold noise and no significance testing is provided.\n\nWhat's new: the paper brings Domain Agnostic Learning to cross-modality medical segmentation, where the target is a mix of unlabeled domains (multi-phasic MR) rather than a single known one. The Domain-Agnostic Module and Anatomy-Preserving Module are reasonable additions, and the ablations in Table 4 do show each module contributes something. The visualization of domain-agnostic images is a nice addition, and the authors are honest that the upper bounds are probably conservative.\n\nWhere the soft spots are: the evidence. The MR cohort is 20 patients from one center, and the three phases are all from the same acquisition protocol family. More importantly, the 5-fold cross-validation means n=5 for each comparison; the differences between DALACE and DADR (0.041 DA, 0.052 DAL) are smaller than or comparable to the reported standard deviations, and no paired test is reported. The text also never says the split is patient-level, so slice-level leakage is a possibility that needs to be ruled out. Without per-fold scores and a paired test, phrases like \"outperforms\" are not supported. External validation on a different site or different MR sequence is also missing. The joint-learning claim that DALACE beats fully-supervised U-Nets has the same problem: the MR margin is 0.038 on a std of about 0.05.\n\nThese are fixable for a revision, and they don't make the method implausible. The architecture is coherent, and the ablation is consistent with the intended mechanism. I would not reject it; I'd send it to serious referees with the expectation that the authors add per-fold results, significance tests, and a clearer statement on data partitioning. Code/data release would help a lot.\n\nWho is it for: people working on unsupervised domain adaptation in medical imaging, especially liver segmentation, and anyone interested in disentanglement as a tool for cross-modality generalization. It's worth a reading group discussion about evaluation standards more than about the architecture.\n\nRecommendation: accept for peer review, with the caveat that the current evidence is suggestive, not conclusive.","headline":"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.","tokens_in":12388,"tokens_out":3602,"would_cite":true,"duration_ms":35797,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["domain agnostic learning","domain adaptation","disentangled representation","liver segmentation","cross-modality segmentation","CT to MRI","adversarial training","anatomy-consistent embedding"],"falsifier":"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.","tokens_in":11439,"feed_emoji":"🩻","tokens_out":6045,"duration_ms":58583,"temperature":0.7,"pith_summary":"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.","feed_headline":"Anatomy-style split lifts liver segmentation on CT and MRI","feed_subtitle":"One model segments liver in CT and three MRI phases, beating dedicated adaptation baselines on Dice score.","key_machinery":"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.","core_discovery":"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).","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the CycleGAN baseline whose DA Dice score of 0.721 and DAL Dice score of 0.522 are compared against DALACE.","marker":"[26]"},{"why":"Supplies the task-driven generative baseline, essentially DALACE without disentanglement, with DA Dice score 0.793 and DAL Dice score 0.719.","marker":"[25]"},{"why":"Supplies the prior disentanglement model DADR; removing the APM from DALACE reduces it to a DADR-like model with DA 0.806 and DAL 0.742.","marker":"[24]"},{"why":"Defines the Domain Agnostic Learning task that the paper adopts for medical segmentation.","marker":"[18]"},{"why":"Provides the style-content disentanglement and style-based generator architecture that DALACE adapts.","marker":"[9]"},{"why":"Supplies adaptive instance normalization (AdaIN), the mechanism for injecting modality codes into the generators.","marker":"[11]"},{"why":"Supplies the U-Net segmentation network used in the Anatomy-Preserving Module and in supervised comparisons.","marker":"[20]"}],"fun_headline_variants":["Split anatomy and modality for robust liver segmentation across CT and MRI","Disentangled embedding beats dedicated adaptation in liver segmentation","Anatomy-consistent sharing lifts liver Dice in CT and multi-phasic MR","One anatomy code for all modalities: liver segmentation without domain labels","Supervised disentanglement improves cross-modality liver segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Split anatomy and modality for robust liver segmentation across CT and MRI","Disentangled embedding beats dedicated adaptation in liver segmentation","Anatomy-consistent sharing lifts liver Dice in CT and multi-phasic MR","One anatomy code for all modalities: liver segmentation without domain labels","Supervised disentanglement improves cross-modality liver segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1381,"prompt_tokens":1024,"completion_tokens":357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":271}},"tokens_in":640,"tokens_out":357,"duration_ms":4310,"temperature":1.0,"reasoning_tokens":271,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:41:46.147588+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CycleGAN baseline whose DA Dice score of 0.721 and DAL Dice score of 0.522 are compared against DALACE."},{"cited_title":"Zhang, S","cited_arxiv_id":null,"evidence_quote":"Supplies the task-driven generative baseline, essentially DALACE without disentanglement, with DA Dice score 0.793 and DAL Dice score 0.719."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Domain Agnostic Learning task that the paper adopts for medical segmentation."},{"cited_title":"Huang, M.-Y","cited_arxiv_id":null,"evidence_quote":"Provides the style-content disentanglement and style-based generator architecture that DALACE adapts."},{"cited_title":"Karras, S","cited_arxiv_id":null,"evidence_quote":"Supplies adaptive instance normalization (AdaIN), the mechanism for injecting modality codes into the generators."}],"review_version":1}