{"id":"3a5bacc3-f889-4d5f-b897-85bebb8b5b0a","arxiv_id":"2506.15549","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CLAIM combines AHA-guided scar mask generation with joint diffusion and segmentation training, raising scar Dice from 58.89 to 63.53 on the EMIDEC test set.","lead":"This paper introduces CLAIM, a framework that uses the clinical AHA 17-segment heart model to guide a diffusion model in synthesizing realistic myocardial scar images for cardiac MRI, and couples it with a jointly trained scar segmentation network. The work matters because scarce labeled scar data is a major bottleneck for automated cardiac MRI analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed Dice improvement over baselines rests on a single 10-case split with no variance estimates; the reported margins are within expected patient-level noise, so the superiority claim is not statistically supported.","rationale":"The reader's weakest_assumption concerns the SMILE registration step; this is a legitimate methodological risk, but the more load-bearing weakness is the statistical fragility of the quantitative evidence. The main claim in the abstract and in the reader's strongest_claim is that CLAIM achieves higher Dice than baselines; this rests entirely on Table 2, which reports single numbers from a 10-case test set with no error bars or significance tests. The reader's rationale also mentions this issue, so there is partial agreement, but the formal weakest_assumption field does not identify it. I considered whether the joint-training mechanism is the most load-bearing concern: Section 2.1 describes L1 for the diffusion model and L_seg for the segmentation network, but never specifies how L_seg gradients reach the generator, making the 'joint' optimization unclear. However, even if that mechanism is imperfect, the downstream Dice claims could still hold; the statistical concern directly invalidates the headline numeric claim unless addressed. The code link is a positive sign and makes the proposed test feasible. A conditional acceptance is appropriate: the verdict does not change, but the authors should add repeated-seed statistics and confidence intervals, and ideally compare with published EMIDEC results.","tokens_in":10666,"tokens_out":8077,"duration_ms":92777,"concrete_test":"Re-run the three key configurations (Baseline, LeFusion+SMILE, CLAIM(J)+SMILE) on the identical 57/10 EMIDEC split with 5 different random seeds and report mean +/- std Dice over the 10 test patients. Then compute patient-level bootstrap 95% confidence intervals (10,000 resamples) for each method's Dice. If the confidence intervals for CLAIM(J)+SMILE overlap with those of LeFusion+SMILE or Baseline, the claimed superiority in Section 3.5 is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is Table 2: CLAIM(J)+SMILE reaches 63.53 Dice vs 58.89 for the no-augmentation baseline and 62.63 for LeFusion+SMILE on the EMIDEC test set. But the test set is only 10 pathological cases (Section 3.2), and no standard deviations, confidence intervals, or repeated-seed results are reported. Cardiac LGE scar Dice varies substantially across patients and scans; a 4.6-point edge over baseline and a 0.9-point edge over LeFusion+SMILE are within the range of sampling noise for n=10. The paper states that experiments were 're-implemented and repeated' but never quantifies the variation across repetitions. Without significance testing or bootstrap intervals, the abstract's claim that CLAIM 'achieves higher Dice similarity with real scar distributions compared to baseline models' is not empirically established. The out-of-domain results in Fig. 5 are also presented without variance estimates and show non-monotonic behavior with increasing synthetic data, further undermining the robustness claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CLAIM, a framework for synthesizing myocardial scar images from LGE cardiac MRI, built around the SMILE module, which generates scar masks conditioned on the AHA 17-segment model via registration to an atlas template. A diffusion model produces synthetic pathological images from normal or pathological inputs, and a joint training scheme alternately optimizes the generator and an nnUNet-based scar segmentation model. Experiments on EMIDEC and a private out-of-domain dataset compare CLAIM and its joint-training variant against LeFusion with DiffMask or SMILE and against a no-augmentation baseline, reporting Dice improvements and qualitatively more anatomical masks and bull's-eye volume distributions.","tokens_in":10846,"tokens_out":2841,"duration_ms":37237,"significance":"If the reported gains are reproducible, the work is a useful contribution: it is the first to inject AHA segment knowledge into a diffusion-based scar synthesis pipeline, it addresses a real data-scarcity problem, and it provides code and an out-of-domain evaluation. The main strengths are the clinically interpretable mask generation and the joint-training formulation, which is a sensible way to couple synthesis and downstream segmentation. However, the headline quantitative claim currently rests on a single split of 10 test cases without variance or significance information, and the registration-dependent anatomical claim is not directly validated. These issues are load-bearing for the stated conclusions.","major_comments":[{"comment":"The central quantitative claim—that CLAIM achieves higher Dice than baselines—is supported only by a single run on 10 EMIDEC test cases. Table 2 reports one Dice value per setting with no standard deviations, confidence intervals, or significance tests, even though the text says experiments were re-implemented and repeated. The largest margin over the no-augmentation baseline is 63.53 versus 58.89, and the margin over LeFusion+SMILE is 0.90 Dice points; both are within plausible patient-level and training-run noise for n=10 cardiac LGE cases. Please report results averaged over multiple training seeds with subject-level bootstrap intervals or paired significance tests, or soften the abstract's claim that CLAIM 'achieves higher Dice similarity with real scar distributions compared to baseline models.'","section":"Section 3.5, Table 2"},{"comment":"SMILE's anatomical coherence depends on the accuracy of rigid plus non-rigid demons registration between each subject's myocardium and the AHA template, but no quantitative validation of this registration is provided. If registration is inaccurate—for example in hearts with large infarcts or different slice spacing—the generated masks will be placed in incorrect AHA segments, undermining the claim of clinically plausible scar patterns. Please report registration quality (e.g., Dice overlap of transformed myocardium masks, visual inspection rates) and, if possible, quantify how often synthetic masks fall outside the myocardium or outside the intended AHA segment after transformation back to subject space.","section":"Section 2.2, Steps 1 and 4"},{"comment":"The out-of-domain evaluation shows non-monotonic Dice as synthetic data increases and is presented without variance estimates. The text attributes the drop on the private dataset to scar volume differences between EMIDEC and the private data, but this explanation is not tested. To support the claimed robustness against domain shift, please provide repeated-run statistics for Fig. 5 and, ideally, a direct comparison of synthetic versus private-data scar volume distributions, or discuss why the non-monotonic behavior does not undermine the augmentation recommendation.","section":"Section 3.5, Fig. 5"},{"comment":"The AHA atlas is used both to generate synthetic scar masks in SMILE and to evaluate per-segment scar volume in the bull's-eye plot analysis. This creates a shared coordinate system that can inflate apparent agreement between predicted and ground-truth segment-level volumes. The evaluation should be framed as an internal consistency check, not an independent validation, or should use a separate anatomical parcellation for evaluation. This issue does not invalidate the voxel-level Dice results, but it weakens the volume-distribution claim.","section":"Section 3.5, Fig. 6 and Section 2.2"}],"minor_comments":[{"comment":"The phrasing 'CLAIM: Clinically-Guided LGE Augmentation for Realistic and Diverse Myocardial Scar Synthesis and Segmentation framework' is grammatically awkward; consider removing the duplicated 'framework' after the acronym expansion.","section":"Abstract"},{"comment":"The text says both CLAIM and CLAIM (J) were trained for '50,000 epochs,' but 50,000 iterations is likely intended for a diffusion model; please clarify the training schedule and distinguish epochs from iterations.","section":"Section 3.2"},{"comment":"The caption contains a typo, 'In-domian,' which should be 'In-domain.'","section":"Fig. 5 caption"},{"comment":"The notation P, N', N\", and P' is introduced only in the caption; please define it in the main text before the table is referenced, and consider using a more visually distinct notation for N' and N\".","section":"Table 2"},{"comment":"The reference to 'their own plotted results (Figure A2)' from LeFusion is unclear because Figure A2 is not included in this manuscript; please either include the figure or describe the comparison in words.","section":"Section 3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and within scope for a medical imaging journal, and the code release is a plus. The main concern is statistical: the reported Dice differences are plausible but not established without variance or significance testing on the 10-case test set. I would also ask for registration validation and a clearer separation between the generation and evaluation uses of the AHA atlas. These are fixable within the manuscript's scope; hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the SMILE module is a genuinely useful idea — conditioning scar synthesis on the AHA 17-segment model is a clinical prior that shows up in the masks and in the downstream segmentation numbers. The joint-training loop is a reasonable extension of LeFusion. But the central quantitative claim is not yet established: everything rests on one 10-case split with no error bars, and the reported margins are exactly the size you'd expect from patient-level noise.\n\nWhat the paper does well: it identifies a real bottleneck (scar data scarcity), builds a concrete solution (atlas-registered mask generation with a diffusion inpainter), and ships code. The bull's-eye plot evaluation is a thoughtful way to look at regional scar volume, and the authors are honest about the out-of-domain degradation — they admit in Sec. 3.5 that joint training doesn't always help on external data and attribute it to volume differences.\n\nSoft spots, in proportion: (1) Statistics. Table 2 shows consistent monotonic gains for CLAIM and CLAIM(J) across all synthetic subsets, which is encouraging, but the paper says experiments were 're-implemented and repeated' without ever reporting the spread. For a 4.6-point Dice edge over baseline on n=10 pathological cases, that's not enough. The fix is straightforward: repeated seeds, confidence intervals, and a paired significance test. (2) The AHA atlas is used both to generate and to evaluate segment-level scar volumes. That's not circular for the overall Dice on real test images, but it does align the bull's-eye evaluation with the generative prior, so treat that figure as partially self-consistency rather than independent validation. Minor. (3) The out-of-domain results are noisy and non-monotonic; the 'improved robustness' contribution in the intro is stronger than the evidence in Fig. 5.\n\nWho deserves this paper: a cardiac imaging researcher working on augmentation or scar segmentation will get a clear, reproducible method description. A serious referee should get it too — the core idea is sound and the code release makes verification feasible. My recommendation: send to peer review and push hard on the statistics. If the variance estimates don't hold up, the paper still has value as a method paper; if they do, the clinical prior point is worth making.\n\nThis is a solid conditional — not a reject, not an accept.","headline":"SMILE's clinical prior is a genuine improvement, but the Dice claims need variance estimates before they persuade.","tokens_in":11394,"tokens_out":3456,"would_cite":true,"duration_ms":39997,"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":"This paper claims that conditioning synthetic myocardial scar generation on the AHA 17-segment model produces anatomically coherent, diverse LGE images that, with joint training of generator and segmenter, increase scar segmentation Dice…","keywords":["myocardial scar segmentation","LGE cardiac MRI","diffusion models","synthetic medical images","AHA 17-segment model","data augmentation","joint training","image synthesis"],"falsifier":"Take a set of LGE cases with expert-annotated AHA segment labels, run SMILE's generation for each intended segment, and measure the agreement between the AHA segment of the synthetic scar after mapping back to subject space and the intended segment; if a substantial fraction of synthetic scars land outside the intended segment or on the opposite wall, the anatomical grounding claim is refuted.","tokens_in":10452,"feed_emoji":"🫀","tokens_out":5640,"duration_ms":59692,"temperature":0.7,"pith_summary":"The paper claims that scarce, hard-to-label myocardial scar data can be supplemented by synthesizing LGE MRI images whose scar patterns respect the clinical AHA 17-segment model of the left ventricle. It introduces CLAIM, built around the SMILE module, which registers each subject's myocardium to an AHA atlas, samples adjacent segments and scar volumes, grows textured 3-D blobs, and maps them back into subject space. A diffusion generator then renders these masks into realistic LGE images, and a segmenter is jointly trained with the generator so that segmentation loss also shapes synthesis. On the EMIDEC test set, CLAIM with joint training reaches 63.53 Dice versus 58.89 for the no-augmentation baseline, and it also beats LeFusion-based baselines in most in-domain settings. If correct, this offers a clinically interpretable route to generating training data for scar segmentation without large annotated datasets.","feed_headline":"Clinically guided synthetic scars lift LGE scar Dice to 63.53","feed_subtitle":"A diffusion model draws scar masks from AHA heart segments, then co-trains the segmenter past the baseline.","key_machinery":"The central object is SMILE (Scar Mask generation guided by cLinical knowledgE), a four-step mask-generation module that transforms each subject's myocardium into the AHA 17-segment atlas space using rigid registration followed by fast symmetric forces demons non-rigid registration, randomly selects one or more adjacent basal, middle, and apical segments, samples scar volumes per region, grows 3-D blobs with controlled anisotropy and porosity into textured masks, and maps the masks back to subject space with non-rigid then rigid registration. SMILE is the component that makes generated scar locations clinically interpretable and spatially diverse; the rest of CLAIM adapts a LeFusion-style lesion-focused diffusion model and a co-trained nnUNet segmenter.","core_discovery":"The central claim is that conditioning pathological image synthesis on standardized ventricular anatomy, rather than on arbitrary blob statistics, yields synthetic scars that are both more realistic and more useful for training a segmenter. Using the AHA 17-segment nomenclature, CLAIM/SMILE controls which myocardial regions contain scar, how large the scar is per segment, and how much texture and porosity it has; the joint training signal then biases the diffusion model toward scar textures that a nnUNet-style segmenter can actually learn. The quantitative evidence is that Dice on EMIDEC rises from 58.89 with real data only to 63.53 for CLAIM with joint training under the P+P'+N'' augmentation setting, with precision also improving and segment-level scar volume errors shrinking in the middle and apical regions on bull's-eye plots.","pith_inferences":["If registration remains the accuracy bottleneck, a stronger variant of this idea would replace subject-to-atlas registration with a learned segment classifier trained on expert-annotated LGE images, avoiding error compounding in scarred, deformed ventricles.","Segment-conditioned synthesis could generalize to other structured cardiac labels, such as coronary territories or scar transmurality, and to other hollow organs whose pathology follows a standard parcellation.","The paper's own observation that out-of-domain gains can drop suggests a concrete correction: resample the synthetic scar volumes to match the target dataset's volume distribution before augmentation.","A blinded study asking cardiologists to distinguish SMILE-generated from real LGE scars would directly test the 'clinically coherent' claim beyond Dice scores."],"forward_implications":["Synthetic scar data generated with SMILE can be added at multiple scales (N', N'', and P+P'+N''), with in-domain Dice and precision improving as the synthetic set grows.","Jointly training the generator and the segmenter yields better downstream segmentation than training the generator alone, especially for small scar regions.","Bull's-eye evaluation against the AHA 17-segment model shows the method reduces segment-level scar volume errors relative to the baseline, particularly in middle and apical segments.","The approach transfers to an out-of-domain private dataset with gains over the baseline in most settings, although the improvement can reverse when the synthetic scar volume distribution mismatches the external data.","Because SMILE exposes which AHA segment generated each mask, the synthetic augmentation can be audited against clinical nomenclature, making the training data interpretable."],"supporting_citations":[{"why":"Supplies the LeFusion lesion-focused diffusion model that CLAIM adapts and the DiffMask mask-generation baseline it is compared against.","marker":"[16]"},{"why":"Provides the AHA 17-segment standardization of left ventricular myocardium that SMILE uses to control scar location.","marker":"[21]"},{"why":"Provides the EMIDEC dataset used for training the synthesis and segmentation models and for in-domain testing.","marker":"[22]"},{"why":"Supplies the nnUNet segmentation backbone used both as the joint-training segmenter and as the baseline model.","marker":"[25]"},{"why":"Defines the denoising diffusion probabilistic model forward and reverse processes that the generator builds on.","marker":"[26]"},{"why":"Provides the bi-ventricular cardiac atlas with AHA segments that serves as the template space for SMILE's registration.","marker":"[29]"},{"why":"Supplies the fast symmetric forces demons non-rigid registration method used to map myocardium to and from the atlas space.","marker":"[30]"},{"why":"Provides the private out-of-domain LGE dataset used to evaluate robustness against domain shift.","marker":"[23, 24]"}],"fun_headline_variants":["Synthetic scars with AHA guidance lift Dice to 63.53","AHA-guided diffusion boosts LGE scar segmentation to 63.5","Clinically guided scar synthesis achieves 63.5 Dice","Anatomy-aware scar generation improves segmentation Dice","Joint training of generator and segmenter raises scar Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that registering a subject's myocardium to the AHA atlas template preserves anatomical correspondence well enough that blobs grown in template space, when mapped back, land on plausible scar locations in the subject's own heart; if registration misaligns, especially in infarcted or differently spaced ventricles, the generated scar masks land in the wrong segments and the anatomical-coherence claim fails.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic scars with AHA guidance lift Dice to 63.53","AHA-guided diffusion boosts LGE scar segmentation to 63.5","Clinically guided scar synthesis achieves 63.5 Dice","Anatomy-aware scar generation improves segmentation Dice","Joint training of generator and segmenter raises scar Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000134,"raw_usage":{"total_tokens":1152,"prompt_tokens":969,"completion_tokens":183,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":100}},"tokens_in":585,"tokens_out":183,"duration_ms":2971,"temperature":1.0,"reasoning_tokens":100,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:53:52.833139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of LGE cases with expert-annotated AHA segment labels, run SMILE's generation for each intended segment, and measure the agreement between the AHA segment of the synthetic scar after mapping back to subject space and the intended segment; if a substantial fraction of synthetic scars land outside the intended segment or on the opposite wall, the anatomical grounding claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the AHA 17-segment standardization of left ventricular myocardium that SMILE uses to control scar location."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EMIDEC dataset used for training the synthesis and segmentation models and for in-domain testing."},{"cited_title":"Isensee et al., Nnu-net: A self-configuring method for deep learning-based biomedical image segmentation,Nature methods, vol","cited_arxiv_id":null,"evidence_quote":"Supplies the nnUNet segmentation backbone used both as the joint-training segmenter and as the baseline model."},{"cited_title":"Ho et al., Denoising diffusion probabilistic models,Advances in neural infor- mation processing systems, vol","cited_arxiv_id":null,"evidence_quote":"Defines the denoising diffusion probabilistic model forward and reverse processes that the generator builds on."},{"cited_title":"Bai et al., A bi-ventricular cardiac atlas built from 1000+ high resolution mr images of healthy subjects and an analysis of shape and motion,Medical image analysis, vol","cited_arxiv_id":null,"evidence_quote":"Provides the bi-ventricular cardiac atlas with AHA segments that serves as the template space for SMILE's registration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the fast symmetric forces demons non-rigid registration method used to map myocardium to and from the atlas space."}],"review_version":1}