{"id":"b5652ed0-f690-4127-99c6-1d402f2e0b8d","arxiv_id":"1908.05094","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A shape-transfer GAN converts bSSFP cine images into synthetic LGE images while preserving heart shape, letting a segmentor trained on synthetic LGE segment real LGE images without LGE labels.","lead":"This paper trains a neural network to outline the heart muscle in late gadolinium enhancement (LGE) MRI scans without using any labeled LGE images, by first converting easier-to-segment cine MRI images into synthetic LGE images. The method reports accurate-looking segmentation scores on a 40-patient test set, though three failed cases were left out of the final statistics.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core assumption that synthetic LGE reproduces infarct enhancement is untested; the 40-patient claim rests on unshown GAN outputs and excludes three failures.","rationale":"The reader's verdict is CONDITIONAL and I agree. The most load-bearing uncertainty is not a mathematical contradiction but an empirical one: the whole architecture routes LGE segmentation through synthetic images whose pathology content is never verified. The shape-preservation loss in Eq. 7 is the only label-driven term in the LGE direction, and it uses bSSFP masks, so it cannot distinguish a synthetic LGE image that contains a real infarct from one that merely looks LGE-like while remaining segmentable from non-pathological cues. The paper's own ablation (Table 1) is encouraging: Shape-Transfer GAN clearly beats No-Shape GAN on the 5 validation patients, which suggests the shape loss helps. But 5 patients is small, no generated image is displayed, and the 40-patient table excludes three failures without reporting their Dice. A single quantitative check on the 5 validation patients can determine whether synthetic LGE actually carries hyperenhanced tissue; if it does not, the method's success must be attributed to something other than the stated transfer mechanism and the central claim is not established. The right verdict remains CONDITIONAL, not REJECT, because the conditional evidence is positive and the check is readily runnable.","tokens_in":6377,"tokens_out":7187,"duration_ms":80509,"concrete_test":"Use the 5 MS-CMRSeg validation patients that have LGE ground truth: generate synthetic LGE from their bSSFP slices with the trained G1, then compare, within the myocardial mask, the fraction of hyperenhanced pixels (intensity > mean + 2 SD of remote myocardium) in the synthetic images vs the real LGE images of the same patients. If the synthetic images have no comparable bright tail, the segmentor has not been trained on infarct appearance and the claimed generalization to real LGE is not supported by the intended mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the synthetic LGE images being representative of real LGE in precisely the regard that makes LGE segmentation hard: bright, heterogeneous infarcted myocardium. Nothing in the objective guarantees this. Eq. (7) is a cross-entropy between the bSSFP mask mx and the segmentor's output on G1(x); it only enforces that the synthetic image be segmentable into the same shape. The cycle loss and adversarial loss constrain global realism and structure, but not the intensity statistics of pathological myocardium. The generator can therefore satisfy all terms by emitting LGE-like textures that S can segment from shape/boundary cues, while omitting or relocating hyperenhanced tissue. No generated image is shown and no quantitative measure of translation quality is reported. The consequence is visible in Table 2: three test patients are excluded without criteria, and the 5-patient validation in Table 1 cannot establish that S is using real LGE appearance. If the excluded cases are failures caused by absent or atypical enhancement, the reported 0.847/0.776/0.686 overstate the method's direct applicability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Shape-Transfer GAN for segmenting the left ventricle, right ventricle, and myocardium in late gadolinium enhancement (LGE) cardiac MRI without using any LGE segmentation labels. The method learns a bidirectional bSSFP-to-LGE image translation with cycle-consistency and an additional shape-preservation loss: a segmentation network S is trained on synthetic LGE images produced from bSSFP inputs, supervised by bSSFP ground-truth masks, so that after training S can be applied directly to real LGE images. The authors report an ablation study on 5 validation patients and final test results on 40 patients with Dice scores of 0.847 (LV), 0.776 (RV), and 0.686 (myocardium), after excluding three failure cases, and compare against GMM-based and deep-learning baselines.","tokens_in":6564,"tokens_out":5667,"duration_ms":55259,"significance":"If the central claim is validated, the work would be practically valuable: it offers a way to train an LGE segmenter using only bSSFP labels plus unpaired LGE images, which is relevant to clinical settings where LGE annotations are scarce. The paper's strengths include a clear problem formulation, an explicit objective combining adversarial, cycle-consistency, and shape-preservation terms, and an ablation study showing that each component (adversarial learning, then shape preservation) improves over a U-Net baseline on the small validation set. The method is falsifiable and the experimental design is straightforward to reproduce in principle. However, the published evidence does not yet substantiate the key assumption that the synthetic LGE images preserve the pathological hyperenhanced appearance that makes real LGE segmentation difficult, and the reporting of the test set is incomplete.","major_comments":[{"comment":"The dataset description is internally inconsistent. The text states that data were collected from 45 patients, that ground truth was provided for bSSFP and T2 images of 35 patients, that LGE ground truth was provided for 5 patients for validation, and then that 'the rest 40 patients are used for test.' Since 35 + 5 = 40, this leaves only 5 patients from the 45-patient collection, not 40. Please clarify the total number of patients, how the 40-patient test set is formed, whether the 35 bSSFP-labeled patients overlap with the test set, and how the 5 validation patients relate to the 40 test patients. This is necessary to interpret the headline Dice scores.","section":"Section 3.1"},{"comment":"The test metrics are reported after excluding three failure cases, but no exclusion criterion is stated. Excluding failures without a transparent rule can arbitrarily inflate the reported Dice scores, and the reader cannot tell whether the excluded cases correspond to missing or atypical hyperenhancement, which is precisely the failure mode most relevant to LGE segmentation. Please describe the exclusion rule, report the metrics with and without the excluded cases, and state how many patients were used for each reported number.","section":"Section 3.2, Table 2"},{"comment":"The shape-preservation loss only enforces that the synthetic LGE image G1(x) is segmentable into the bSSFP mask mx; it does not constrain the intensity distribution of pathological, hyperenhanced myocardium, which is the main source of difficulty in real LGE segmentation. The cycle-consistency and adversarial losses constrain global structure and image realism, but they do not guarantee that the synthetic images reproduce the bright, heterogeneous appearance of infarcted tissue. Without any visual examples of generated LGE images or a quantitative analysis of translation fidelity (e.g., intensity statistics in infarct versus remote myocardium, or a domain-distance measure), the central claim that the segmentor trained on synthetic LGE generalizes to real LGE is not established. Please add such an analysis.","section":"Section 2.3, Eq. (7)"},{"comment":"The segmentation network S is co-trained with the generator G1 through the shape-preservation loss, and the same S is then applied to real LGE images at test time. This creates a potential co-adaptation: G1 may learn to produce synthetic images that are easy for S to segment using boundary or shape cues inherited from bSSFP, without S learning to handle the actual intensity/texture of real LGE. The ablation on 5 validation patients (Table 1) is too small to rule out this degenerate solution. I recommend reporting a direct comparison of synthetic and real LGE appearance (for example, a reader study or feature-space distances) and, if possible, an analysis of segmentation performance stratified by the presence of hyperenhancement.","section":"Sections 2.3 and 2.4"}],"minor_comments":[{"comment":"There are several typos: 'Its worth' should be 'It's worth', 'to be segment' should be 'to segment', and 'LEG images' in the caption of Figure 3 should be 'LGE images'.","section":"Abstract and Introduction"},{"comment":"In the second term of LGAN2, the argument of G2 should be y, not x, to be consistent with the text and with the expectation over y: the generated bSSFP image is G2(y), not G2(x).","section":"Section 2.1, Eq. (2)"},{"comment":"The formula for the Jaccard index in Eq. (9) incorrectly includes a factor of 2; the Jaccard index is |A ∩ B| / |A ∪ B|.","section":"Section 3.1, Evaluation Metrics"},{"comment":"The label 'Resetnet Block' should read 'Resnet Block'.","section":"Figure 2"},{"comment":"The text cites Goodfellow et al. as 2016, but reference [4] is the 2014 GAN paper; please correct the year or the reference.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a challenge- or workshop-style paper. The central idea is plausible and the ablation is suggestive, but the dataset accounting and the failure-case exclusion need to be fixed, and the lack of any generated-image evidence is a substantive gap for a method whose entire argument rests on the fidelity of synthetic LGE images. If the authors can supply the missing translation-quality analysis and clarify the test set, the paper could become acceptable; as it stands, the main claim is not fully evidenced."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take on arXiv:1908.05094. The idea: train a CycleGAN to turn bSSFP cine images into fake LGE images, embed a segmentor into the generator, and use bSSFP ground-truth masks via a shape-preservation cross-entropy loss so the segmentor learns from synthetic LGE without any real LGE labels. The ablation is the best part: U-Net trained directly on bSSFP fails, adding the GAN transfer helps, and adding the shape-preservation loss helps again. That is a clean demonstration that each component contributes.\n\nWhat is genuinely new here is the specific combination: an embedded segmentor inside a cycle-consistent translation network with a shape-preservation term for unpaired cross-modality cardiac segmentation. It builds on standard CycleGAN and U-Net, but the application and the loss design are a reasonable extension.\n\nThe soft spots are real but not fatal. The biggest one: nothing in the loss guarantees that the fake LGE images actually reproduce the pathological hyperenhancement that makes real LGE segmentation hard. Eq. (7) only forces the segmentor's output on G1(x) to match the bSSFP mask; it does not constrain the generator to produce infarct-like brightness or texture. The authors show no generated images and no quantitative measure of translation fidelity, so we have to take on faith that the synthetic LGE is representative in the exact respect that matters. If the generator simply paints LGE-like textures while preserving shape, the segmentor can learn to find boundaries rather than pathology—which may explain the modest myocardium Dice (0.686) and the need to exclude three failure cases without stated criteria. That exclusion is a selection bias that should be disclosed and justified.\n\nOther issues are minor: the dataset description says 45 patients but 35+5=40, the Jaccard formula in Eq. (9) is miswritten, and the citation numbering is inconsistent. No code or model is released, which is not fatal but limits reproducibility.\n\nThe central argument holds up in the sense that the method is plausible and the ablation supports the components. But the paper overstates direct applicability given the excluded failures and untested synthetic quality. It deserves a serious referee rather than a desk reject, but the authors should be asked to show generated LGE examples, define failure exclusion criteria, and correct the minor errors.\n\nWho is this for: researchers working on cross-modality cardiac segmentation or unpaired image translation with embedded task networks. It is a useful workshop-level contribution.\n\nMy recommendation: send it to peer review with a request for major revision focusing on synthetic-quality evaluation and transparent reporting.","headline":"Plausible cross-modality segmentation pipeline with a useful ablation, but the headline test numbers rest on untested synthetic-image quality and post-hoc exclusion of failures.","tokens_in":7105,"tokens_out":2663,"would_cite":false,"duration_ms":25924,"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":"By turning bSSFP cine images into synthetic LGE images with a shape-preserving GAN, this paper trains an LGE segmentor without using any LGE labels and reports Dice of 0.847, 0.776, and 0.686 on a 40-patient test set.","keywords":["cardiac magnetic resonance imaging","late gadolinium enhancement","myocardial segmentation","cross-modality image translation","generative adversarial network","cycle consistency","shape preservation","unpaired image-to-image translation"],"falsifier":"Give the trained segmentor held-out real LGE slices with known masks and stratify Dice by whether a slice contains a bright infarct; if accuracy stays high on healthy slices but collapses on infarct slices while the shape-preservation loss is satisfied, the transfer of pathological appearance, not overall anatomy, is what failed.","tokens_in":6174,"feed_emoji":"🫀","tokens_out":13427,"duration_ms":131629,"temperature":0.7,"pith_summary":"The paper tries to establish that a shape-preserving generative adversarial network (GAN) can train a segmentor for late gadolinium enhancement (LGE) cardiac MRI using only labels from balanced steady-state free precession (bSSFP) cine images, whose boundaries are clearer and easier to segment, and never using LGE labels. The method translates bSSFP slices into synthetic LGE images, enforces that each synthetic image keeps the known myocardium shape, and trains an embedded segmentation network on those synthetic pairs. After training, that segmentation network is applied directly to real LGE images. On a 40-patient test set, with three failure cases excluded, the paper reports Dice scores of 0.847 for the left ventricle, 0.776 for the right ventricle, and 0.686 for the myocardium. If the claim holds, LGE segmentation becomes possible in settings where only cine labels are available.","feed_headline":"Style-transfer GAN segments heart MRI without late-enhancement labels","feed_subtitle":"It learns LGE appearance from bSSFP cine scans, then segments real LGE hearts from that learned shape.","key_machinery":"The load-bearing object is the Shape-Transfer GAN objective \\(L_{\\mathrm{total}} = L_{\\mathrm{GAN}} + \\lambda_1 L_{\\mathrm{cyc}} + \\lambda_2 L_{\\mathrm{shape}}\\), with \\(L_{\\mathrm{shape}}(S,G_1) = \\mathbb{E}_x[\\mathrm{CE}(m_x, S(G_1(x)))]\\). This shape-preservation term forces the segmentor's prediction on a fake LGE image to match the known bSSFP mask, anchoring anatomy during unpaired style transfer. Cycle consistency stabilizes the translation, the discriminators enforce LGE-like intensity and texture, and the same \\(S\\) that serves as the shape regularizer becomes the deployed LGE segmentor.","core_discovery":"The central discovery is that unpaired style transfer, constrained by a segmentation network that must reproduce the source-domain mask, produces a usable target-domain segmentor. Concretely, the generator \\(G_1\\) maps bSSFP slices \\(x\\) to fake LGE slices, and the segmentor \\(S\\) is trained with cross-entropy against the bSSFP mask \\(m_x\\) on those fake slices; after training, \\(S\\) is deployed unchanged on real LGE slices. The paper's evidence includes reported Dice scores on the test set and an ablation in which the shape-preservation term lifts validation myocardium Dice from 0.303 to 0.607.","pith_inferences":["The same cycle-plus-shape recipe could be pointed at other unlabeled cardiac MRI targets, such as T2-weighted images, reusing the same bSSFP masks; the paper does not test this.","A slice-level comparison of segmentor Dice on synthetic versus real LGE would separate style fidelity from anatomy transfer, something the aggregate Dice averages do not reveal.","The three excluded failures argue that a deployable version would need an automatic check for when the shape prior is violated, since the network itself would not flag low confidence.","If bright infarct regions are not faithfully generated, average Dice can stay moderate while diseased myocardium is systematically under-segmented; stratifying test Dice by infarct presence would settle this."],"forward_implications":["A label-free LGE segmentor can be built from bSSFP labels alone, which addresses the shortage of manual LGE annotations in practice.","At inference the method is a single forward pass on real LGE slices; there is no per-patient registration, iterative optimization, or LGE label fine-tuning.","The ablation attributes most of the transfer gain to the shape-preservation term: on the five-patient validation set, myocardium Dice rises from 0.303 without it to 0.607 with it.","The method does not outperform a deep network trained on real LGE labels (myocardium Dice 0.686 vs 0.812 in the paper's comparison), so supervised labels still add accuracy when they exist.","The test-set averages cover 37 of 40 patients; the three excluded failure cases are not part of the reported Dice numbers."],"supporting_citations":[{"why":"Supplies the cycle-consistent unpaired translation mechanism that turns bSSFP images into fake LGE images without paired data.","marker":"[5]"},{"why":"Supplies the adversarial learning objective used to make the generated images realistic.","marker":"[4]"},{"why":"Provides the supervised SRSCN comparison in the paper, marking the accuracy level reached only with real LGE labels.","marker":"[1]"},{"why":"Provides the GMM-based comparison methods that the paper matches without their iterative per-patient optimization.","marker":"[2]"}],"fun_headline_variants":["Shape-transfer GAN segments LGE MRI without LGE labels","Learn LGE shape from bSSFP, then segment: no labels needed","Unpaired GAN transfers bSSFP shape to segment LGE hearts","Segments late-enhancement MRI using bSSFP shape, zero LGE labels","Generator learns LGE look; segmentor uses it: no LGE masks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that fake LGE images produced from bSSFP are faithful enough in the clinically relevant places, especially bright diseased myocardium, that a segmentor trained on them transfers to real LGE; the shape-preservation loss only checks that the segmentor's output matches the bSSFP mask, not that the fake image reproduces real LGE pathology.","fun_headline_variants_meta":{"raw":{"variants":["Shape-transfer GAN segments LGE MRI without LGE labels","Learn LGE shape from bSSFP, then segment: no labels needed","Unpaired GAN transfers bSSFP shape to segment LGE hearts","Segments late-enhancement MRI using bSSFP shape, zero LGE labels","Generator learns LGE look; segmentor uses it: no LGE masks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000869,"raw_usage":{"total_tokens":3719,"prompt_tokens":851,"completion_tokens":2868,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":2768}},"tokens_in":467,"tokens_out":2868,"duration_ms":18147,"temperature":1.0,"reasoning_tokens":2768,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:23:29.226629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give the trained segmentor held-out real LGE slices with known masks and stratify Dice by whether a slice contains a bright infarct; if accuracy stays high on healthy slices but collapses on infarct slices while the shape-preservation loss is satisfied, the transfer of pathological appearance, not overall anatomy, is what failed.","supporting_citations":[{"cited_title":"Unpaired image-to-image translation using cycle-consistent adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the cycle-consistent unpaired translation mechanism that turns bSSFP images into fake LGE images without paired data."},{"cited_title":"Generative adversarial nets","cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial learning objective used to make the generated images realistic."},{"cited_title":"Fully Automatic Left Atrium Segmentation From Late Gadolinium Enhanced Magnetic Resonance Imaging Using a Dual Fully Convolutional Neural Network,","cited_arxiv_id":null,"evidence_quote":"Provides the GMM-based comparison methods that the paper matches without their iterative per-patient optimization."}],"review_version":1}