{"id":"346317cc-b939-4d87-a775-dfaa043340dd","arxiv_id":"1908.07344","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Unpaired style transfer from bSSFP to synthetic LGE images lets a cascaded U-net segment real LGE MRI without labeled LGE training data, with high Dice scores on the MS-CMRSeg test set.","lead":"The authors segment cardiac structures in late-gadolinium enhanced (LGE) MRI without using any labeled LGE images, by first translating labeled bSSFP images into synthetic LGE images and then fine-tuning a cascaded U-net on these synthetic images. The method reports average Dice scores of 0.92 for the left ventricle, 0.83 for the myocardium, and 0.88 for the right ventricle on a challenge test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on the unverified assumption that MUNIT preserves cardiac anatomy exactly; the paper provides no quantitative check of label fidelity in synthetic LGE, and bSSFP contains no infarct information, so the reported MYO Dice may reflect bSSFP shape priors rather than true domain…","rationale":"The reader's weakest assumption identified the same load-bearing point: anatomy preservation by MUNIT is assumed, not shown. My stress-test confirms and sharpens this concern. The paper's reported success depends on synthetic LGE images carrying correct labels from bSSFP. Yet no metric verifies that content codes are truly anatomy-invariant rather than appearance-contaminated. In addition, the specific clinical challenge of LGE, namely infarct-enhanced myocardium with heterogeneous intensity and blurry borders, cannot be synthesized from bSSFP because that information is absent in the source modality. This explains the lower MYO score and suggests the fine-tuned network may be exploiting bSSFP-like boundaries rather than adapting to true LGE appearance. The absence of error bars on the 40-subject test results weakens the evidence but does not invalidate the strong absolute scores. The paper has external benchmark support, but no released code or independent reproduction. The conditional verdict remains appropriate, provided the requested anatomy-preservation check is performed; hence the reader's verdict should not change.","tokens_in":6860,"tokens_out":4978,"duration_ms":54213,"concrete_test":"Select 20 of the 150 synthetic LGE images uniformly at random; have two expert cardiologists, blinded to the source bSSFP labels, manually segment LV, MYO, and RV in each; compute Dice against the inherited bSSFP labels. If the median MYO Dice is below 0.85, the translation pipeline does not preserve anatomy to the accuracy required for fine-tuning, and the reported test performance must be re-attributed, for example to bSSFP shape priors rather than to unsupervised style transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is Section 2.1's claim that the MUNIT content code 'preserves the underlying anatomical structure' while the style code is swapped. This is asserted by citation to [8], not demonstrated. The synthetic LGE images are generated from 30 labelled bSSFP slices (Section 3.1); each bSSFP label map is transferred unchanged to all 5 stochastic translations. If the content code leaks bSSFP-specific appearance (e.g., edge sharpness, blood-pool homogeneity) or the decoder adds or removes structures, the synthetic labels are systematically wrong. The paper offers only qualitative examples (Fig. 5) and no quantitative anatomy-preservation metric. Moreover, because bSSFP has no late-gadolinium enhancement, the synthetic LGE cannot reproduce the heterogeneous, bright infarcted myocardium that the Introduction identifies as the main difficulty; the fine-tuned network may therefore rely on clean bSSFP-like boundaries that are absent in real LGE. This is consistent with MYO Dice (0.83) being markedly lower than LV (0.92) and RV (0.88). The reported test numbers come from 40 subjects but no error bars or per-subject distributions are given, so it is unknown whether the synthetic fine-tuning gain over the bSSFP-only U-net (Table 1 rows 2-3) is statistically significant. Thus the central claim 'accurate LGE segmentation without labeled LGE data' is only as strong as the unvalidated label-fidelity assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully automatic, two-stage framework for segmenting cardiac structures in late-gadolinium enhanced (LGE) MRI without using labelled LGE images during training. First, an MUNIT-based multi-modal translation network, trained on unpaired bSSFP and LGE images, converts annotated bSSFP images into multiple synthetic LGE-style images while keeping the underlying anatomy fixed. Second, a cascaded two-U-net segmentation network is pre-trained on labelled bSSFP images and then fine-tuned on the synthetic LGE images plus the original bSSFP labels. At inference, dense CRF and morphological post-processing are applied. The authors evaluate on the MS-CMRSeg 2019 challenge data: on a 5-volume validation set the cascaded fine-tuned network improves Dice over a bSSFP-only U-net and a registration baseline, and on the 40-subject test set they report average Dice of 0.92 for the left ventricle, 0.83 for the myocardium, and 0.88 for the right ventricle.","tokens_in":7177,"tokens_out":5567,"duration_ms":53977,"significance":"The result, if supported, is practically significant: it removes the need for costly LGE annotations by transferring anatomical knowledge from bSSFP, which is easier to label, and it is evaluated on an external challenge test set rather than an in-house dataset. The paper also ships a clear experimental pipeline, uses the official MUNIT implementation, and includes a quantitative comparison to both a registration-based method and a single U-net on the validation set. However, the main claim rests on the untested assumption that the synthetic LGE images preserve the bSSFP anatomy exactly, and the statistical evidence for the central performance gain is thin (five validation images, no error bars, no significance tests, and no test-set comparators).","major_comments":[{"comment":"The central claim that fine-tuning on synthetic LGE images yields accurate real-LGE segmentation depends on the assertion in §2.1 that MUNIT preserves underlying anatomical structure when swapping style codes. This is currently supported only by a citation to [8] and by qualitative examples in Fig. 5. Because the synthetic training labels are taken unchanged from the source bSSFP labels, any leakage of bSSFP-specific appearance into the content code or any decoder-introduced artifact will produce systematically wrong labels. Please add a quantitative anatomy-preservation check, for example by segmenting the synthetic LGE images with a network trained on the same bSSFP labels and measuring Dice with the original masks, or by measuring contour/landmark agreement between input bSSFP and synthetic LGE across all 150 generated images.","section":"§2.1, §3.1, Fig. 5"},{"comment":"The validation experiments reported in Table 1 are computed on only five LGE volumes, and the paper gives no per-subject numbers, error bars, or significance tests. The statement in §3.3 that fine-tuning with synthetic LGE 'significantly improves' the MYO Dice by about 15% is therefore not statistically supported; with n=5, a single favorable volume could drive the apparent improvement. Please report the per-volume Dice/ASD distributions for every method and a paired statistical test (e.g., Wilcoxon signed-rank or paired t-test) on the validation set.","section":"§3.3, Table 1"},{"comment":"The final test-set results on 40 subjects are reported as absolute Dice and ASD values (LV 0.92, MYO 0.83, RV 0.88) but are not compared with any alternative method on the same test set. Without a test-set comparison, the reader cannot judge whether the proposed method is actually competitive for LGE segmentation, and no per-subject distribution is given, so the variability of the test performance is unknown. Please provide test-set results for at least one baseline (e.g., U-net (FT) or MIRTK), or report the challenge leaderboard/ranking, together with per-subject statistics.","section":"§3.3, test-set results"},{"comment":"The introduction identifies the heterogeneous, enhanced infarcted myocardium as the main challenge in LGE segmentation, but the bSSFP source domain contains no late-gadolinium enhancement information. Consequently, the synthetic LGE images cannot reproduce the bright, heterogeneous scarred myocardium that the method is ultimately asked to segment, and the lower MYO Dice (0.83 test, 0.816 validation) compared with LV and RV is consistent with the network relying on bSSFP-like boundary cues rather than true LGE appearance. This is not fatal for the proposed pipeline, but the claim of accurate LGE segmentation should be qualified as applying primarily to myocardium with bSSFP-like contrast; the paper should discuss this limitation explicitly and, if feasible, stratify the evaluation by the presence or extent of myocardial scar.","section":"§1, §3.3"}],"minor_comments":[{"comment":"The sentence 'the posterior modelled by the network p(yb|xbl) matches p(yl|xl) when image space and label space are shared' is imprecise: the image spaces are not literally shared after translation, and the equality is the goal of adaptation rather than a condition. Please rephrase.","section":"§2.2"},{"comment":"Several hyperparameters (edge loss weight λ=0.5, class weights 0.2:0.25:0.3:0.25, five synthetic samples per bSSFP image) are introduced without sensitivity analysis; a brief justification or a reference for each choice would make the pipeline more reproducible.","section":"§3.2"},{"comment":"Figure 5 is only mentioned as 'supplemental material' in §3.1 and is not referenced by number in the main text; please cite it explicitly.","section":"Supplemental Material"},{"comment":"The N/A entries for the U-net ASD values should be clarified: if the network produced no prediction in some volumes, the corresponding Dice scores should be treated consistently (e.g., as zero) or those volumes should be excluded with the criterion stated.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue and the external challenge evaluation is a definite strength. My recommendation of major revision is driven by the absence of quantitative evidence for the load-bearing label-fidelity assumption and by the statistically weak validation; neither issue appears unfixable within the manuscript's scope. I would not reject, because the method is clearly described and the reported test numbers are encouraging, but the claims need to be tempered and supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, well-engineered paper that shows you can segment LGE cardiac MR without labeled LGE data by translating labeled bSSFP images into synthetic LGE and fine-tuning a cascaded U-net. The result is real, externally evaluated on the MS-CMRSeg 2019 test set (40 subjects), and the numbers are credible: 0.92 LV, 0.83 MYO, 0.88 RV Dice. What is new is not any single component—MUNIT, U-net, auto-context, CRF are all established—but the specific recipe and demonstration that it works on a public challenge. That is a meaningful integrative contribution for cardiac MRI and for cross-modal medical image segmentation generally.\n\nThe paper does some things well. The method is clearly specified, the training procedure is reproducible enough from the text (though no code is released), and the comparisons to registration and single-U-net baselines are appropriate. The external test set is a real strength; the central claim is not forced by construction.\n\nThe soft spots are real but not fatal. The load-bearing assumption is that MUNIT preserves anatomy exactly when swapping style. The paper cites [8] for this and shows qualitative examples, but gives no quantitative label-fidelity check. That matters because bSSFP has no late-gadolinium enhancement, so synthetic LGE cannot reproduce the heterogeneous bright infarcted myocardium that makes real LGE hard. The lower MYO Dice (0.83 vs 0.92 LV) is consistent with that concern. A second issue: the baseline comparisons and the fine-tuning gain are demonstrated on five validation images, with no error bars or significance tests. The test numbers are from an ensemble of four models, but you cannot tell from the paper whether the fine-tuning gain over the bSSFP-only U-net is statistically robust. No code is provided.\n\nNone of this changes my read that the paper deserves a serious referee. The central result is plausible, externally benchmarked, and directly useful to anyone working on unsupervised domain adaptation in medical imaging. The right revision would add a quantitative anatomy-preservation metric on synthetic LGE (e.g., Dice between bSSFP labels propagated and a human label on real LGE, or a segmentation probe), per-subject error bars on the validation set, and ideally code or a clear reproducibility note. I would take the test-set numbers as real but would be cautious about overclaiming that the synthetic images preserve all clinically relevant structure.\n\nRecommendation: send to peer review, with the above points as required revisions.","headline":"Solid integrative paper showing unsupervised LGE cardiac segmentation by translating labeled bSSFP to synthetic LGE; the central claim is plausible and externally benchmarked, but the anatomy-preservation assumption needs quantitative support.","tokens_in":7705,"tokens_out":2415,"would_cite":true,"duration_ms":22475,"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":"Synthetic LGE images generated from bSSFP scans, with zero labelled LGE data, yield accurate cardiac segmentation.","keywords":["cardiac MRI segmentation","late gadolinium enhancement","unsupervised domain adaptation","multi-modal image translation","style transfer","cascaded U-net","synthetic training data","unpaired image translation"],"falsifier":"Take a batch of synthetic LGE images generated from bSSFP inputs, have an expert segment the three cardiac structures on them, and compute Dice against the original bSSFP ground-truth labels; if myocardium Dice on this synthetic-to-source check falls well below the 0.83 reported on real LGE, the translation is not preserving anatomy and the fine-tuning labels are systematically corrupted.","tokens_in":6679,"feed_emoji":"🫀","tokens_out":7683,"duration_ms":70531,"temperature":0.7,"pith_summary":"This paper tries to establish that a cardiac MRI segmentation network can learn to segment late-gadolinium enhanced (LGE) images without ever seeing a labelled LGE image during training. The idea is to translate easily annotated balanced steady-state free precession (bSSFP) scans into realistic, diverse synthetic LGE images with an unpaired style-transfer network, then fine-tune a segmentation network pre-trained on labelled bSSFP data using those synthetic images with their original labels. On the held-out test set the resulting pipeline reports an average Dice score of 0.92 for the left ventricle, 0.83 for the myocardium, and 0.88 for the right ventricle. The significance, if the claim holds, is that annotation effort can be shifted from a modality with blurry, ambiguous borders to one where borders are clear, and the same recipe may transfer to other modalities.","feed_headline":"No labeled LGE images? Style transfer still segments the heart","feed_subtitle":"An unpaired translation network turns labeled bSSFP scans into synthetic LGE training data, preserving anatomy.","key_machinery":"The load-bearing mechanism is MUNIT's disentangled representation: a content encoder extracts anatomy, a style encoder captures modality-specific appearance, and the decoder re-renders the content with a sampled LGE style. Because the style space is trained to match a standard Gaussian, one bSSFP image yields many synthetic LGE images, providing data diversity. On the segmentation side, the cascaded U-net and its composite loss — weighted cross-entropy plus a Sobel-filter edge loss — carry the argument that combining shape priors with appearance handles the blurry infarct boundaries typical of LGE.","core_discovery":"The central claim is that an unsupervised, unpaired image-translation step can close the appearance gap between bSSFP and LGE enough for labels to transfer. Using MUNIT, each bSSFP image is decomposed into a domain-invariant content code and a domain-specific style code; sampling style codes from a Gaussian prior produces multiple LGE-stylized versions of the same anatomy. These synthetic LGE images, paired with the original bSSFP labels, are used to fine-tune a two-stage cascaded U-net that first produces coarse probability maps and then refines them by combining image appearance with the predicted shape prior. The paper reports that this synthetic fine-tuning improves myocardium Dice by roughly 15 percentage points over training on bSSFP alone and outperforms both a registration-based transfer baseline and a single fine-tuned U-net.","pith_inferences":["A testable extension the paper leaves implicit is measuring anatomy-preservation fidelity directly: compare expert outlines of synthetic LGE images against the original bSSFP labels to isolate how much of the reported Dice comes from translation quality rather than the cascaded segmenter.","The method's ceiling is probably set by the content code's invariance; if the scarred, enhanced myocardium carries appearance information that the encoder routes into style, the transferred labels will systematically misrepresent infarct regions, so per-region Dice on scar versus remote myocardium would be a sharper evaluation than whole-structure Dice.","The same pipeline could be pointed at other unpaired modality pairs, such as CT to MRI or different MRI sequences, whenever one modality is easy to annotate and the other is not; the paper itself names brain MRI as future work."],"forward_implications":["If the reported Dice scores hold, LGE cardiac segmentation can be automated in settings where no LGE annotations exist, as long as bSSFP labels are available.","The synthetic LGE images act as a data-augmentation engine: sampling different style codes from the same anatomy increases training diversity and can improve generalization.","The cascaded shape-plus-appearance design adds robustness on apical and basal slices, where single-pass networks and registration methods tend to fail.","The same two-stage recipe — unpaired translation followed by fine-tuning a pretrained segmenter — is proposed as a general route for other cross-modality tasks with one labelled modality."],"supporting_citations":[{"why":"Supplies the multi-modal translation network whose disentangled content/style codes generate synthetic LGE images.","marker":"[7]"},{"why":"Provides the U-net architecture used as the backbone of both stages of the cascaded segmentation network.","marker":"[6]"},{"why":"Represents the supervised LGE segmentation baseline this work aims to surpass without LGE labels.","marker":"[5]"},{"why":"Supports the claim that disentangled representations preserve anatomical structure across modalities.","marker":"[8]"},{"why":"Provides the dense CRF used in post-processing to refine the predicted segmentation masks.","marker":"[12]"},{"why":"Motivates the auto-context design in which the second U-net consumes the first network's probability maps as shape prior.","marker":"[10]"}],"fun_headline_variants":["Style transfer turns bSSFP into synthetic LGE for cardiac segmentation","Unlabeled LGE? Style transfer from bSSFP still segments hearts","Synthetic LGE via style transfer: unsupervised cardiac MR segmentation","bSSFP style transfer produces fake LGE to train heart segmenter"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that translating a bSSFP image into LGE style preserves the anatomy pixel for pixel; if the translator changes shape or introduces artifacts in the enhanced myocardium, the synthetic labels mislead the segmentation network.","fun_headline_variants_meta":{"raw":{"variants":["Style transfer turns bSSFP into synthetic LGE for cardiac segmentation","Unlabeled LGE? Style transfer from bSSFP still segments hearts","Synthetic LGE via style transfer: unsupervised cardiac MR segmentation","bSSFP style transfer produces fake LGE to train heart segmenter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1463,"prompt_tokens":898,"completion_tokens":565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":489}},"tokens_in":514,"tokens_out":565,"duration_ms":6082,"temperature":1.0,"reasoning_tokens":489,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:19:22.559743+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a batch of synthetic LGE images generated from bSSFP inputs, have an expert segment the three cardiac structures on them, and compute Dice against the original bSSFP ground-truth labels; if myocardium Dice on this synthetic-to-source check falls well below the 0.83 reported on real LGE, the translation is not preserving anatomy and the fine-tuning labels are systematically corrupted.","supporting_citations":[{"cited_title":"Multimodal unsupervised image-to-image translation","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-modal translation network whose disentangled content/style codes generate synthetic LGE images."},{"cited_title":"U-net: Convolutional networks for biomedical image segmen- tation","cited_arxiv_id":null,"evidence_quote":"Provides the U-net architecture used as the backbone of both stages of the cascaded segmentation network."},{"cited_title":"Cardiac Segmentation from LGE MRI Using Deep Neural Network Incorporating Shape and Spatial Priors","cited_arxiv_id":null,"evidence_quote":"Represents the supervised LGE segmentation baseline this work aims to surpass without LGE labels."},{"cited_title":"Unsupervised deformable registration for multi-modal images via disentan- gled representations","cited_arxiv_id":null,"evidence_quote":"Supports the claim that disentangled representations preserve anatomical structure across modalities."},{"cited_title":"Efficient inference in fully connected CRFs with gaussian edge potentials","cited_arxiv_id":null,"evidence_quote":"Provides the dense CRF used in post-processing to refine the predicted segmentation masks."},{"cited_title":"Auto-context and its application to high-level vision tasks and 3D brain image segmentation","cited_arxiv_id":null,"evidence_quote":"Motivates the auto-context design in which the second U-net consumes the first network's probability maps as shape prior."}],"review_version":1}