LGE-Guided Cross-Modality Contrastive Learning for Gadolinium-Free Cardiomyopathy Screening in Cine CMR

Dinggang Shen; Hui Wang; Lei Chen; Lei Xu; Siqing Yuan; Yueyan Wang; Yulin Wang; Zehao Weng; Zhong Xue; Zirui Cao

REVIEW 4 major objections 4 minor 13 references

LGE-Guided Cross-Modality Contrastive Learning for Gadolinium-Free Cardiomyopathy Screening in Cine CMR

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A training-time fusion of cine and contrast-enhanced cardiac MRI lets CC-CMR screen cardiomyopathy from non-contrast cine alone, reporting 94.3% accuracy and beating the previous cine-only model by 4.3 percentage points.

desk verdict Plausible training-time LGE-to-cine transfer idea, but the reported stats don't support the headline gain; the confidence intervals are internally inconsistent. read the letter →

arxiv 2508.16927 v1 pith:5V4RR6II submitted 2025-08-23 cs.CV

Siqing Yuan , Yulin Wang , Zirui Cao , Yueyan Wang , Zehao Weng , Hui Wang , Lei Xu , Zixian Chen

show 3 more authors

Lei Chen Zhong Xue Dinggang Shen

This is my paper · ORCID

classification cs.CV

keywords cardiomyopathyscreeningcardiacmagneticresonancecontrastivelearningcross-modalalignmentlategadoliniumenhancementgadolinium-freeMRIuncertainty-awarevideoswintransformer

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 proposes CC-CMR, a framework that learns to screen for cardiomyopathy from non-contrast cine CMR sequences while using Late Gadolinium Enhancement (LGE) MRI as a training-time teacher. The central idea is to align the feature spaces of cine and LGE so that pathology visible only with gadolinium contrast—myocardial fibrosis and scar—becomes encoded into the cine representations. At inference the LGE branch is dropped; only 4-chamber and short-axis cine are needed. On 231 multi-center subjects the model reaches 94.3% accuracy and 0.946 F1, outperforming the state-of-the-art cine-only model by 4.3 accuracy points and 0.039 F1, with no gadolinium exposure. The gain matters because it points toward a practical, contrast-free screening route for cardiomyopathy in settings where contrast is risky or unavailable.

What carries the argument

The Feature Interaction Module, a dual-loss training component that couples classification with cross-modal feature alignment. The classification loss (focal loss) preserves per-sequence discriminability; the Feature Space Alignment loss, a label-aware adaptation of InfoNCE, aligns same-class embeddings across cine and LGE sequences while separating different classes. An uncertainty-aware mechanism with learnable uncertainty parameters σ per sequence and per pair balances the two objectives and avoids gradient conflict. The backbone is a Video Swin Transformer pretrained on Kinetics600 and fine-tuned jointly across the three sequences; at inference only the two cine encoders and a lightweigh

What would settle it

Train the identical pipeline with LGE replaced by a non-informative control, such as shuffled LGE volumes paired to the wrong patients or a different contrast sequence; if the accuracy over the cine-only baseline stays at 0.943 instead of dropping, then LGE-to-cine transfer is not the cause of the improvement.

Watch

Extended reading notes

Core claim

The paper claims that aligning the latent spaces of cine CMR and LGE sequences during training causes the cine encoder to internalize LGE-visible pathology, specifically myocardial fibrosis and scar. This is enforced by a Feature Interaction Module with two losses: a focal classification loss that keeps each sequence's features class-separable, and a label-aware InfoNCE-style alignment loss that pulls same-label cross-sequence features together while pushing different labels apart. An uncertainty-aware mechanism weights these objectives per sequence and per sequence pair, adapting during optimization. After training, LGE is discarded; the 4CH and SAX cine encoders plus a fusion MLP output th

Load-bearing premise

The entire reported gain rests on the assumption that aligning LGE and cine feature spaces during training transfers fibrosis-specific discriminative information into the cine-only encoder, rather than the gain coming from regularization, evaluation bias, or leakage.

Editorial extensions

If this is right

Cardiomyopathy screening can be performed with routine non-contrast cine CMR, removing gadolinium exposure and its contraindications in patients with renal impairment or pregnancy.
The training-time use of LGE avoids the missing-modality failure of strict multimodal fusion models while still exploiting LGE information during learning.
The uncertainty-aware weighting offers a stable way to combine classification and alignment objectives in multi-sequence medical imaging.
Because the Feature Interaction Module is modality-agnostic, the same training scheme can be extended to additional imaging sequences or modalities.
The framework also improves five-class cardiomyopathy subtyping over the state-of-the-art cine-only model, as reported in the supplemental material.

Reading between the lines

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

The reported test set has only 70 subjects, so the 4.3-point accuracy gap could partly reflect bootstrap optimism; a pre-registered evaluation on a larger, independently collected cohort is the natural next check.
A direct test of the mechanism would be to train the same pipeline with LGE replaced by a non-informative control, such as shuffled LGE volumes or a different sequence; if the accuracy gain persists, the benefit is not specifically LGE pathology transfer.
Because the FSA loss averages cosine similarities by label before the contrastive step, the alignment may be close to class-conditional distribution matching; comparing against instance-level alignment would reveal which form of correspondence carries the benefit.
If the transfer is genuine, the same 'teacher modality at training, cheap modality at inference' recipe could apply to other contrast-dependent or costly imaging signals, for example PET guiding CT-only screening.

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.

Desk Editor's Note

read the letter

The interesting idea here is using LGE as a teacher during training and dropping it at inference for gadolinium-free screening. That's clinically motivated and the pipeline is clearly described. The Feature Interaction Module is a sensible combination of InfoNCE-style alignment and classification, and the uncertainty weighting is standard Kendall et al. The ablations are internally consistent: adding the module and the uncertainty mechanism improves accuracy on the same test set.

The problem is the statistics. The bootstrap confidence intervals in Table 2 are impossible — Row 4 reports sensitivity 0.919 with CI (0.949, 0.967), and Row 1 reports specificity 0.788 with CI (0.796, 0.859), where the lower bound exceeds the point estimate. That means the bootstrap was likely done on samples or images rather than patients, or with a wrong percentile method. All CIs in the paper should be treated as unreliable. More fundamentally, the accuracy CIs for SOTA and CC-CMR overlap (0.829–0.957 vs 0.886–0.986), so the 4.3% gain is not statistically significant. On a 70-subject test set, that difference is just three correct classifications, and no paired test like McNemar is reported.

The mechanism claim — that LGE-specific fibrosis information is transferred into cine embeddings — is plausible but only supported by a t-SNE visualization on the same test set. No code, no public dataset, no external validation. The paper is not circular; the training uses LGE supervision and inference doesn't, so the direction is honest. But the evidence as presented doesn't establish superiority over the SOTA cine-only model.

This deserves a serious referee because the application is important and the training scheme is worth examining. A revision would need to fix the bootstrap methodology, report patient-level CIs and a paired test, and ideally validate on a public dataset or release code. Right now it's a solid idea with unconvincing evidence.

Referee Report

4 major / 4 minor

Summary. The paper proposes CC-CMR, a two-stage contrastive learning framework for binary cardiomyopathy screening from cine CMR only at inference. During training, three Video Swin Transformer encoders process 4-chamber cine, short-axis cine, and short-axis LGE sequences. A Feature Interaction Module applies focal classification losses and a label-based InfoNCE-style feature-space alignment loss, combined through homoscedastic uncertainty weights. At inference, only the two cine branches and a fusion MLP are used; LGE is discarded. On a 161-subject training / 70-subject test split from two centers, the authors report accuracy 0.943 (95% CI 0.886–0.986) versus 0.900 (0.829–0.957) for the SOTA cine-only model [11], and interpret t-SNE plots as evidence that the Feature Interaction Module aligns cross-sequence features. The central claim is that LGE-guided contrastive training encodes fibrosis-specific pathology into the cine embeddings, enabling gadolinium-free screening without loss of accuracy.

Significance. If the central claim were established, the contribution would be practically valuable: using routinely acquired LGE during training to build a cine-only inference model could reduce gadolinium exposure and make CMR-based cardiomyopathy screening more scalable. The idea of label-level cross-sequence contrastive alignment with uncertainty-based loss weighting is a reasonable extension of CLIP-style multimodal alignment to cardiac video, and the paper is clearly written in terms of its intended pipeline. However, the current evidence does not establish the claimed superiority over the SOTA model: the reported confidence intervals are internally impossible, the headline accuracy difference is small and not statistically tested, and the ablation does not isolate the LGE-specific transfer mechanism. The manuscript also does not provide code or data, which limits reproducibility. The core idea is plausible and the statistical issues are fixable, but the paper as submitted does not support its main claims.

major comments (4)

[Section 3.2, Tables 1 and 2] The reported 95% confidence intervals cannot be correct for a 70-subject test set. For example, Table 2 Row 4 reports sensitivity = 0.919 with CI (0.949, 0.967), where the lower bound is greater than the point estimate; Row 1 reports specificity = 0.788 with CI (0.796, 0.859), also impossible. These values indicate that the bootstrap was likely performed at the slice/frame level rather than at the patient level, or with an incorrect percentile method. Since all CIs in Tables 1–2 use the same procedure, none can be treated as valid patient-level intervals. This directly undermines the 'enhanced stability' statement and the reported performance comparison.
[Section 3.2, Table 1] The headline claim that CC-CMR outperforms the SOTA model [11] is not statistically supported. The accuracy CIs overlap substantially: SOTA 0.900 (0.829, 0.957) and CC-CMR 0.943 (0.886, 0.986). The two models are evaluated on the same 70-subject test set, so the appropriate comparison is a paired test (e.g., McNemar's test or a bootstrap over the paired accuracy difference), which is not reported. The absolute difference corresponds to only 3 out of 70 subjects, so without a valid paired test the claim of outperforming SOTA by 4.3% is not established.
[Section 2.2, Eq. (2)] The Feature Space Alignment loss is internally inconsistent. The equation defines y_ij = 1 if i and j are from the same patient, which is an instance-level positive pair definition. However, the following paragraph states that in practice the cosine similarity matrices are averaged by labels and that i and j 'actually represent the label indices.' Instance-level alignment (same patient) and label-level alignment (same class) are different objectives, with different effects on the learned representation. The paper needs to state clearly which version is used and derive the corresponding loss; otherwise the proposed mechanism cannot be reproduced or evaluated.
[Section 2.1 / Table 2] The ablation does not isolate the LGE-guided transfer mechanism. Rows 3 vs 4–5 of Table 2 compare cine-only fusion against adding the Feature Interaction Module, but this adds multiple simultaneous changes: a classification loss on the LGE branch, the feature alignment loss, and the uncertainty weighting. To support the abstract's claim that LGE-specific fibrosis information is transferred to cine embeddings, the authors should include a control in which the auxiliary modality is another cine sequence (or the same LGE input with the contrastive alignment removed). Without such a control, the observed accuracy gains could be due to generic multi-task regularization or to the extra classification supervision, rather than to LGE-specific pathology transfer.

minor comments (4)

[After Table 2] There is an apparent encoding corruption: a block of '/uni00000017/...' characters appears after Table 2. This should be removed.
[Section 3.1] The phrase 'mid-3 slices' should be reworded for clarity (e.g., 'three mid-ventricular slices'). Also, the input shape description for LGE as '(batch, RGB channels duplicated from intensity, number of slices, 224, 224)' is unusual; please clarify the channel handling.
[Section 2.3, Eq. (4)] The uncertainty weighting in Eq. (4) uses log(σ + 1) regularization, which is not the standard form in Kendall et al. The authors should justify why this form is preferred and specify whether the σ parameters are constrained positive, since the regularization term assumes σ > −1.
[Fig. 3] The t-SNE visualization is qualitative. The paper would be strengthened by a quantitative metric (e.g., alignment distance or silhouette score) to support the claim that the Feature Interaction Module improves both alignment and separability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LGE is used for training only; cine-only inference is evaluated on held-out labels, so no prediction reduces by construction to a fitted input or self-citation.

full rationale

The paper's derivation chain is not circular. CC-CMR trains encoders with a classification loss (Eq. 1), a cross-sequence contrastive/alignment loss (Eq. 2), and uncertainty-weighted losses (Eq. 4), all on a labeled training set. At inference only cine CMR is fed to the model, and the reported accuracy/F1 are computed on a held-out test set of 70 subjects with ground-truth disease labels. The LGE modality and the uncertainty parameters are training-time inputs; neither is used to define or construct the evaluation labels or metric. The claim that LGE alignment transfers fibrosis-specific pathology is an explanatory mechanism, not an equation-level identity. The comparison to the SOTA model [11] is an external empirical comparison on the same test set. Although Table 2 contains internally inconsistent bootstrap CIs (e.g., Row 4 sensitivity 0.919 with CI (0.949,0.967)), this is a statistical reporting/validity problem, not circular reasoning. There is no fitting of a parameter to the test result, no target quantity in the loss, no uniqueness theorem, and no load-bearing self-citation. Hence the paper is self-contained against its external benchmark; score 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard transfer learning assumptions plus the clinical validity of LGE as fibrosis ground truth. The only fitted parameters beyond ordinary network weights are the uncertainty weights sigma, and the bootstrap validity assumption is demonstrably questionable. No new physical entities are introduced.

free parameters (3)

uncertainty parameters sigma (per-task and per-pair)
Learned during training; balance CLS and FSA losses in Eq. 4. Values not reported.
focal loss focusing parameter gamma = 2
Adopted from Lin et al. [5]; fixed by prior work, not fitted.
input slice counts (1 4CH, 3 SAX cine, 6 SAX LGE)
Hand-selected mid-ventricular slices; affects model input and results.

assumptions (5)

domain assumption Kinetics600-pretrained Video Swin Transformer provides useful motion/feature representation for cardiac cine MRI
Used as initialization in Section 3.1; transferability from natural video to cardiac MRI is assumed.
domain assumption LGE hyperintensity is a valid proxy for myocardial fibrosis and is learnable by the cine encoder through cross-modal alignment
Core premise of the method; no mechanistic or clinical validation is provided beyond test accuracy.
domain assumption The 231-subject cohort is representative of the cardiomyopathy subtypes and scanner heterogeneity of the target population
Claims multi-center coverage and ESC 2023 subtypes; no demographic composition or external validation given.
standard math Bootstrap resampling yields valid confidence intervals for the test metrics
Efron's bootstrap [2] is used in Section 3.2, but the reported intervals for sensitivity/specificity are implausibly narrow, indicating the resampling units are not independent test subjects.
domain assumption Patient-level split between training and test contains no leakage
Not explicitly stated in the paper; necessary for unbiased evaluation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LGE-Guided Cross-Modality Contrastive Learning for Gadolinium-Free Cardiomyopathy Screening in Cine CMR." pith.science (2026). https://pith.science/paper/5V4RR6II

@misc{pith2026250816927,
  author       = {Pith},
  title        = {Pith review of: LGE-Guided Cross-Modality Contrastive Learning for Gadolinium-Free Cardiomyopathy Screening in Cine CMR},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5V4RR6II}},
  note         = {Machine review of arXiv:2508.16927}
}

read the original abstract

Cardiomyopathy, a principal contributor to heart failure and sudden cardiac mortality, demands precise early screening. Cardiac Magnetic Resonance (CMR), recognized as the diagnostic 'gold standard' through multiparametric protocols, holds the potential to serve as an accurate screening tool. However, its reliance on gadolinium contrast and labor-intensive interpretation hinders population-scale deployment. We propose CC-CMR, a Contrastive Learning and Cross-Modal alignment framework for gadolinium-free cardiomyopathy screening using cine CMR sequences. By aligning the latent spaces of cine CMR and Late Gadolinium Enhancement (LGE) sequences, our model encodes fibrosis-specific pathology into cine CMR embeddings. A Feature Interaction Module concurrently optimizes diagnostic precision and cross-modal feature congruence, augmented by an uncertainty-guided adaptive training mechanism that dynamically calibrates task-specific objectives to ensure model generalizability. Evaluated on multi-center data from 231 subjects, CC-CMR achieves accuracy of 0.943 (95% CI: 0.886-0.986), outperforming state-of-the-art cine-CMR-only models by 4.3% while eliminating gadolinium dependency, demonstrating its clinical viability for wide range of populations and healthcare environments.

Figures

Figures reproduced from arXiv: 2508.16927 by the authors.

**Figure 2.** An overview of proposed CC-CMR, including (a) model s [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** t-SNE visualization before and after using Feature I [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages

[11]

Nature Medicine 30(5), 1471–1480 (2024)

Wang, Y.R., Yang, K., Wen, Y., Wang, P., Hu, Y., Lai, Y., Wa ng, Y., Zhao, K., Tang, S., Zhang, A., et al.: Screening and diagnosis of cardi ovascular disease using artiﬁcial intelligence-enabled cardiac magnetic resonan ce imaging. Nature Medicine 30(5), 1471–1480 (2024)

work page 2024
[1]

European Heart Journal 44(37), 3503–3626 (2023)

Arbelo, E., Protonotarios, A., Gimeno, J.R., Arbustini, E., Barriales-Villa, R., Basso, C., Bezzina, C.R., Biagini, E., Blom, N.A., de Boer, R .A., et al.: 2023 ESC guidelines for the management of cardiomyopathies: Develo ped by the task force on the management of cardiomyopathies of the European Socie ty of Cardiology (ESC). European Heart Journal 44(37...

work page 2023
[2]

Chapman and Hall/CRC (1994)

Efron, B., Tibshirani, R.J.: An introduction to the boots trap. Chapman and Hall/CRC (1994)

work page 1994
[3]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning us ing uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482–7491 ( 2018)

work page 2018
[4]

arXiv preprint arXiv:1805.06334 (2018)

Liebel, L., Körner, M.: Auxiliary tasks in multi-task lea rning. arXiv preprint arXiv:1805.06334 (2018)

arXiv 2018
[5]

In: Proceedings of the IEEE International Confere nce on Computer Vision

Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Foc al loss for dense object de- tection. In: Proceedings of the IEEE International Confere nce on Computer Vision. pp. 2999–3007. IEEE (2017) 10 Authors Suppressed Due to Excessive Length

work page 2017
[6]

In: Proceedings of the IEEE/CVF conference on compu ter vision and pat- tern recognition

Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H .: Video swin trans- former. In: Proceedings of the IEEE/CVF conference on compu ter vision and pat- tern recognition. pp. 3202–3211 (2022)

work page 2022
[7]

Journal of Machine Learning Research 9(86), 2579–2605 (2008)

Van der Maaten, L., Hinton, G.: Visualizing data using t-S NE. Journal of Machine Learning Research 9(86), 2579–2605 (2008)

work page 2008

Show all 13 references

[8]

arXiv preprint arXiv:1807.03748 (2018)

Oord, A.v.d., Li, Y., Vinyals, O.: Representation learni ng with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

2018 arXiv
[9]

In: Proceedings of the 38th I nternational Conference on Machine Learning

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., A garwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transfe rable visual models from natural language supervision. In: Proceedings of the 38th I nternational Conference on Machine Learning. vo...

2021
[10]

Computers in Biology and Medicine 175, 108505 (2024)

Sun, L., Zhang, M., Lu, Y., Zhu, W., Yi, Y., Yan, F.: Nodule -CLIP: Lung nodule classiﬁcation based on multi-modal contrastive learning. Computers in Biology and Medicine 175, 108505 (2024)

2024
[12]

IEEE Transactio ns on Medical Imaging pp

Yu, X., Zhang, L., Wu, Z., Zhu, D.: Core-periphery multi- modality feature align- ment for zero-shot medical image analysis. IEEE Transactio ns on Medical Imaging pp. 1–1 (2024)

2024
[13]

Biomedical Si gnal Processing and Control 99, 106882 (2025)

Zhu, J., Liu, H., Liu, X., Chen, C., Shu, M.: Cardiovascul ar disease detection based on deep learning and multi-modal data fusion. Biomedical Si gnal Processing and Control 99, 106882 (2025)

2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [11]

Nature Medicine 30(5), 1471–1480 (2024)

Wang, Y.R., Yang, K., Wen, Y., Wang, P., Hu, Y., Lai, Y., Wa ng, Y., Zhao, K., Tang, S., Zhang, A., et al.: Screening and diagnosis of cardi ovascular disease using artiﬁcial intelligence-enabled cardiac magnetic resonan ce imaging. Nature Medicine 30(5), 1471–1480 (2024)

work page 2024

[2] [1]

European Heart Journal 44(37), 3503–3626 (2023)

Arbelo, E., Protonotarios, A., Gimeno, J.R., Arbustini, E., Barriales-Villa, R., Basso, C., Bezzina, C.R., Biagini, E., Blom, N.A., de Boer, R .A., et al.: 2023 ESC guidelines for the management of cardiomyopathies: Develo ped by the task force on the management of cardiomyopathies of the European Socie ty of Cardiology (ESC). European Heart Journal 44(37...

work page 2023

[3] [2]

Chapman and Hall/CRC (1994)

Efron, B., Tibshirani, R.J.: An introduction to the boots trap. Chapman and Hall/CRC (1994)

work page 1994

[4] [3]

In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning us ing uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 7482–7491 ( 2018)

work page 2018

[5] [4]

arXiv preprint arXiv:1805.06334 (2018)

Liebel, L., Körner, M.: Auxiliary tasks in multi-task lea rning. arXiv preprint arXiv:1805.06334 (2018)

arXiv 2018

[6] [5]

In: Proceedings of the IEEE International Confere nce on Computer Vision

Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Foc al loss for dense object de- tection. In: Proceedings of the IEEE International Confere nce on Computer Vision. pp. 2999–3007. IEEE (2017) 10 Authors Suppressed Due to Excessive Length

work page 2017

[7] [6]

In: Proceedings of the IEEE/CVF conference on compu ter vision and pat- tern recognition

Liu, Z., Ning, J., Cao, Y., Wei, Y., Zhang, Z., Lin, S., Hu, H .: Video swin trans- former. In: Proceedings of the IEEE/CVF conference on compu ter vision and pat- tern recognition. pp. 3202–3211 (2022)

work page 2022

[8] [7]

Journal of Machine Learning Research 9(86), 2579–2605 (2008)

Van der Maaten, L., Hinton, G.: Visualizing data using t-S NE. Journal of Machine Learning Research 9(86), 2579–2605 (2008)

work page 2008