Pith. sign in

REVIEW 4 major objections 5 minor 30 references

CLAIM: Clinically-Guided LGE Augmentation for Realistic and Diverse Myocardial Scar Synthesis and Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read 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…

desk verdict SMILE's clinical prior is a genuine improvement, but the Dice claims need variance estimates before they persuade. read the letter →

arxiv 2506.15549 v2 pith:YFY7WI6P submitted 2025-06-18 cs.CV cs.AI

classification cs.CVcs.AI
keywords myocardialscarsegmentationLGEcardiacMRIdiffusionmodelssyntheticmedicalimagesAHA17-segmentmodeldataaugmentationjointtrainingimagesynthesis
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

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

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

  • 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.
Share X Bluesky LinkedIn Reddit HN

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 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.

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 (4)
  1. [Section 3.5, Table 2] 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.'
  2. [Section 2.2, Steps 1 and 4] 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.
  3. [Section 3.5, Fig. 5] 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.
  4. [Section 3.5, Fig. 6 and Section 2.2] 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.
minor comments (5)
  1. [Abstract] 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.
  2. [Section 3.2] 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.
  3. [Fig. 5 caption] The caption contains a typo, 'In-domian,' which should be 'In-domain.'
  4. [Table 2] 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".
  5. [Section 3.4] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CLAIM's Dice results are empirical comparisons on held-out real data, and the AHA atlas enters as an explicit generative prior rather than as a fitted target.

full rationale

CLAIM is an empirical synthesis-and-segmentation pipeline rather than a derivation, so there is no equation-level chain in which an output reduces to an input. SMILE uses the AHA 17-segment atlas to sample synthetic scar locations and volumes; this is an explicit design prior, not a parameter fitted to the test set, and the downstream Dice scores in Table 2 are measured on real EMIDEC pathological cases and on a separate 80-case private dataset against real ground-truth scar labels. Joint training supervises the segmentation network with synthetic masks that the same diffusion framework produced, but that loop does not force the reported improvement: the evaluation is on real images not generated by the model. The only self-citations (Refs. 2, 23, 24) provide background and describe the clinical private dataset; they carry no load-bearing theorem or uniqueness argument. The absence of confidence intervals or repeated-seed variance for the 10-case in-domain test is a statistical robustness limitation, not evidence of circularity, and is therefore noted as a correctness risk rather than scored here.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the LeFusion diffusion formulation (cited, not re-derived), on the AHA atlas and registration pipeline, and on hand-set SMILE parameters whose exact values are not reported. No new physical entities are introduced. The main uncertainty is whether the atlas registration and the small EMIDEC training set are representative enough.

free parameters (3)
  • Scar volume sampling bounds per AHA region = not reported
    SMILE samples scar volumes from a uniform distribution for each AHA region; the bounds are not reported and directly control synthetic scar size diversity and downstream Dice.
  • Blob texture parameters (anisotropy, porosity, erosion kernel) = anisotropy unspecified; porosity adjusted; kernel sampled from 1x1 to 7x7
    These hand-chosen texture parameters in SMILE Step 3 affect scar realism and are not optimized or ablated.
  • Synthetic dataset cardinalities in augmentation subsets = N'=56, N''=140, P'=57
    The authors chose how many synthetic images to add; downstream Dice changes with subset size, so these experimental settings partly determine the reported improvements.
assumptions (4)
  • domain assumption The Bai et al. bi-ventricular atlas with AHA 17 segments is applicable to short-axis LGE images from EMIDEC and to the private dataset.
    SMILE registers each subject's myocardium to this template (Section 2.2). The atlas was built from healthy high-resolution MR images, so its segment correspondence on LGE with pathology is not guaranteed.
  • domain assumption Rigid plus non-rigid demons registration preserves anatomical correspondence for scar placement.
    Steps 1 and 4 of SMILE depend on registration accuracy; any registration error propagates directly into synthetic scar mask locations.
  • domain assumption LeFusion's lesion-focused diffusion formulation, control histogram, and background-preserving blending carry over unchanged to cardiac LGE images.
    Section 2.1 adapts LeFusion without re-deriving or validating these components on LGE data.
  • domain assumption Myocardium masks are available for all subjects used for synthesis.
    SMILE needs LV myocardium masks for registration; the paper says these are 'usually available' in public scar datasets but does not provide them for the private set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLAIM: Clinically-Guided LGE Augmentation for Realistic and Diverse Myocardial Scar Synthesis and Segmentation." pith.science (2026). https://pith.science/paper/YFY7WI6P

@misc{pith2026250615549,
  author       = {Pith},
  title        = {Pith review of: CLAIM: Clinically-Guided LGE Augmentation for Realistic and Diverse Myocardial Scar Synthesis and Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YFY7WI6P}},
  note         = {Machine review of arXiv:2506.15549}
}
read the original abstract

Deep learning-based myocardial scar segmentation from late gadolinium enhancement (LGE) cardiac MRI has shown great potential for accurate and timely diagnosis and treatment planning for structural cardiac diseases. However, the limited availability and variability of LGE images with high-quality scar labels restrict the development of robust segmentation models. To address this, we introduce CLAIM: \textbf{C}linically-Guided \textbf{L}GE \textbf{A}ugmentation for Real\textbf{i}stic and Diverse \textbf{M}yocardial Scar Synthesis and Segmentation framework, a framework for anatomically grounded scar generation and segmentation. At its core is the SMILE module (Scar Mask generation guided by cLinical knowledgE), which conditions a diffusion-based generator on the clinically adopted AHA 17-segment model to synthesize images with anatomically consistent and spatially diverse scar patterns. In addition, CLAIM employs a joint training strategy in which the scar segmentation network is optimized alongside the generator, aiming to enhance both the realism of synthesized scars and the accuracy of the scar segmentation performance. Experimental results show that CLAIM produces anatomically coherent scar patterns and achieves higher Dice similarity with real scar distributions compared to baseline models. Our approach enables controllable and realistic myocardial scar synthesis and has demonstrated utility for downstream medical imaging task. Code is available at https://github.com/farheenjabeen/CLAIM-Scar-Synthesis.

Figures

Figures reproduced from arXiv: 2506.15549 by the authors.

Figure 1
Figure 1. Overview of CLAIM. (a) illustrates the training of a diffusion model with real [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of SMILE (Scar Masks Incorporating cLinical knowledgE) [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visualization of real and synthetic masks generated by DiffMask [16] and [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of synthetic pathological images generated by different methods [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Comparing performance of different methods for scar segmentation on EMIDEC [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Bull’s eye plots for the evaluation of scar volume distribution with respect to [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

  1. [1]

    E.R.Jenistaetal.,Revisitinghowweperformlategadoliniumenhancementcmr: Insights gleaned over 25 years of clinical practice,Journal of Cardiovascular Magnetic Resonance, vol. 25, no. 1, p. 18, 2023

  2. [2]

    Y.Kiberuetal.,Myocardialscarimaging:Viabilitybeyondrevived,CurrentCar- diovascular Imaging Reports, vol. 17, no. 10, pp. 107–114, 2024

  3. [3]

    Lalande et al., Deep learning methods for automatic evaluation of delayed enhancement-mri

    A. Lalande et al., Deep learning methods for automatic evaluation of delayed enhancement-mri. the results of the emidec challenge,Medical Image Analysis, vol. 79, p. 102428, 2022

  4. [4]

    L.Lietal.,Myops:Abenchmarkofmyocardialpathologysegmentationcombin- ingthree-sequencecardiacmagneticresonanceimages,MedicalImageAnalysis, vol. 87, p. 102808, 2023

  5. [5]

    Qayyum et al., Secure and robust machine learning for healthcare: A survey, IEEE Reviews in Biomedical Engineering, vol

    A. Qayyum et al., Secure and robust machine learning for healthcare: A survey, IEEE Reviews in Biomedical Engineering, vol. 14, pp. 156–180, 2020

  6. [6]

    Savage,Synthetic data could be better than real data, 2023

    N. Savage,Synthetic data could be better than real data, 2023

  7. [7]

    Q.Chenetal.,Towardsgeneralizabletumorsynthesis,inCVPR,2024,pp.11147– 11158

  8. [8]

    Z.Dorjsembeetal.,Conditionaldiffusionmodelsforsemantic3dbrainmrisyn- thesis,IEEE Journal of Biomedical and Health Informatics, 2024

Show all 30 references
  1. [9]

    I. Horvath et al., Metgan: Generative tumour inpainting and modality synthesis in light sheet microscopy, inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 227–237

  2. [10]

    A. K. Pishva et al., Repolyp: A framework for generating realistic colon polyps with corresponding segmentation masks using diffusion models, in2023 IEEE 36thinternationalsymposiumoncomputer-basedmedicalsystems(CBMS),IEEE, 2023, pp. 47–52

  3. [11]

    Du et al., Arsdm: Colonoscopy images synthesis with adaptive refinement se- mantic diffusion models, inMICCAI, Springer, 2023, pp

    Y. Du et al., Arsdm: Colonoscopy images synthesis with adaptive refinement se- mantic diffusion models, inMICCAI, Springer, 2023, pp. 339–349

  4. [12]

    C. Han et al., Synthesizing diverse lung nodules wherever massively: 3d multi- conditional gan-based ct image augmentation for object detection, in2019 Inter- national Conference on 3D Vision (3DV), IEEE, 2019, pp. 729–737

  5. [13]

    Jin et al., Free-form tumor synthesis in computed tomography images via richergenerativeadversarialnetwork,Knowledge-BasedSystems,vol.218,p.106753, 2021

    Q. Jin et al., Free-form tumor synthesis in computed tomography images via richergenerativeadversarialnetwork,Knowledge-BasedSystems,vol.218,p.106753, 2021

  6. [14]

    Billot et al., Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,Medical image analysis, vol

    B. Billot et al., Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,Medical image analysis, vol. 86, p. 102789, 2023

  7. [15]

    Wu et al., Freetumor: Large-scale generative tumor synthesis in computed to- mographyimagesforimprovingtumorrecognition,arXivpreprintarXiv:2502.18519, 2025

    L. Wu et al., Freetumor: Large-scale generative tumor synthesis in computed to- mographyimagesforimprovingtumorrecognition,arXivpreprintarXiv:2502.18519, 2025

  8. [16]

    Zhang et al., Lefusion: Controllable pathology synthesis via lesion-focused diffusion models,arXiv preprint arXiv:2403.14066, 2024

    H. Zhang et al., Lefusion: Controllable pathology synthesis via lesion-focused diffusion models,arXiv preprint arXiv:2403.14066, 2024

  9. [17]

    Tian et al., Lesiondiffusion: Towards text-controlled general lesion synthesis, arXiv preprint arXiv:2503.00741, 2025

    H. Tian et al., Lesiondiffusion: Towards text-controlled general lesion synthesis, arXiv preprint arXiv:2503.00741, 2025

  10. [18]

    Chou et al., Ultrasound image synthesis using generative ai for lung ultra- sound detection,arXiv preprint arXiv:2501.06356, 2025

    Y.-C. Chou et al., Ultrasound image synthesis using generative ai for lung ultra- sound detection,arXiv preprint arXiv:2501.06356, 2025. 14 F. Ramzan et al

  11. [19]

    Lugmayr et al., Repaint: Inpainting using denoising diffusion probabilistic models,inProceedingsoftheIEEE/CVFconferenceoncomputervisionandpat- tern recognition, 2022, pp

    A. Lugmayr et al., Repaint: Inpainting using denoising diffusion probabilistic models,inProceedingsoftheIEEE/CVFconferenceoncomputervisionandpat- tern recognition, 2022, pp. 11461–11471

  12. [20]

    Rombach et al., High-resolution image synthesis with latent diffusion mod- els, inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp

    R. Rombach et al., High-resolution image synthesis with latent diffusion mod- els, inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695

  13. [21]

    M. D. Cerqueira et al., Standardized myocardial segmentation and nomenclature for tomographic imaging of the heart. a statement for healthcare professionals from the cardiac imaging committee of the council on clinical cardiology of the americanheartassociation:Astatementforhe...

  14. [22]

    A.Lalandeetal.,Emidec:Adatabaseusablefortheautomaticevaluationofmy- ocardial infarction from delayed-enhancement cardiac mri,Data, vol. 5, no. 4, p. 89, 2020

  15. [23]

    N.Jathanna etal.,Thenottingham ischaemiccardiovascularmagneticresonance resource(notiscmr):Aprospectivepairedclinicalandimagingscardatabase—protocol, Journal of Cardiovascular Magnetic Resonance, vol. 25, no. 1, p. 69, 2023. [24]Creation of a cardiac magnetic resonance (CMR) s...

  16. [25]

    Isensee et al., Nnu-net: A self-configuring method for deep learning-based biomedical image segmentation,Nature methods, vol

    F. Isensee et al., Nnu-net: A self-configuring method for deep learning-based biomedical image segmentation,Nature methods, vol. 18, no. 2, pp. 203–211, 2021

  17. [26]

    Ho et al., Denoising diffusion probabilistic models,Advances in neural infor- mation processing systems, vol

    J. Ho et al., Denoising diffusion probabilistic models,Advances in neural infor- mation processing systems, vol. 33, pp. 6840–6851, 2020

  18. [27]

    O. Ronneberger et al., U-net: Convolutional networks for biomedical image seg- mentation,inMedicalimagecomputingandcomputer-assistedintervention–MICCAI 2015:18thinternationalconference,Munich,Germany,October5-9,2015,pro- ceedings, part III 18, Springer, 2015, pp. 234–241

  19. [28]

    B.S.Selvaduraietal.,Definitionofleftventricularsegmentsforcardiacmagnetic resonanceimaging,JACCCardiovascImaging,vol.11,no.6,pp.926–928,2018

  20. [29]

    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

    W. 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. 26, no. 1, pp. 133–145, 2015

  21. [30]

    Z. Tang et al., An effective non-rigid image registration method based on active demons algorithm, in2016 IEEE 29th International Symposium on Computer- Based Medical Systems (CBMS), IEEE, 2016, pp. 124–129

  22. [31]

    M.Cerqueiraetal.,Standardizedmyocardialsegmentationandnomenclaturefor tomographicimagingoftheheart:Astatementforhealthcareprofessionalsfrom thecardiacimagingcommitteeofthecouncilonclinicalcardiologyoftheamer- ican heart association,Circulation, vol. 105, no. 4, pp. 539–542, 2002

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.