{"id":"1bad20a3-f0a2-4e63-8ef3-83c35c4864bf","arxiv_id":"2508.16927","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Gadolinium-free cardiomyopathy screening with LGE-guided contrastive learning achieves 94.3% accuracy in a private 231-subject cohort, with a modest and not statistically verified edge over a cine-only baseline.","lead":"A machine learning system trained with both contrast and non-contrast heart MRI sequences can screen for cardiomyopathy using only the non-contrast cine videos at prediction time. It reports 94.3% accuracy on a 231-subject two-center dataset, but the small test set and questionable confidence intervals mean the gain over existing cine-only models is not yet established.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy gain over SOTA is not statistically supported: CIs overlap and reported bootstrap intervals are internally inconsistent (e.g., Table 2 Row 4).","rationale":"The reader's verdict of CONDITIONAL is appropriate, but the most load-bearing concern is not the mechanism of cross-modal transfer per se; it is the statistical validity of the reported performance gain. If the improvement is not statistically significant, then even a perfectly sound mechanism does not support the central claim of outperforming SOTA. The reported CIs are not merely wide—they are internally impossible, indicating a concrete error in the evaluation methodology. This directly undermines the quantitative headline and the inference of enhanced stability. The reader did note that leakage in bootstrap evaluation could collapse the claim, which aligns with this concern, but framed the weakest assumption as the mechanism transfer. My emphasis is on the statistics because it is more immediately falsifiable from the paper's own data and because the apparent gain may vanish once correct patient-level analysis is done. I recommend keeping the verdict as CONDITIONAL: the authors must fix the bootstrap, perform paired significance testing, and ideally release code/data to verify both the statistical claims and the underlying mechanism.","tokens_in":7826,"tokens_out":3740,"duration_ms":40194,"concrete_test":"Obtain the per-patient predictions of both CC-CMR and the SOTA cine-only baseline on the 70-subject test set. 1) Recompute 95% confidence intervals via patient-level bootstrap (resample subjects, not images). 2) Apply McNemar's paired test to the 2x2 contingency table of correct/incorrect classifications between the two models. If the corrected CIs overlap and/or McNemar's p > 0.05, the claimed superiority is not statistically supported. Additionally, verify all reported CIs contain their point estimates; flag any that do not as evidence of a methodological error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—CC-CMR outperforming the SOTA cine-only model by 4.3% accuracy and 0.039 F1—rests on a 70-subject test set (33 healthy, 37 patients). The difference is only 3 correct classifications (0.943 vs 0.900). The reported 95% CIs for accuracy overlap substantially (SOTA: 0.829–0.957; CC-CMR: 0.886–0.986), so the improvement is not significant at the 5% level. More seriously, the reported bootstrap CIs are internally inconsistent: Table 2 Row 4 reports sensitivity = 0.919 with CI (0.949, 0.967), where the lower bound exceeds the point estimate; Row 1 reports specificity = 0.788 with CI (0.796, 0.859), again impossible. This indicates the bootstrap was likely performed on samples/images rather than patients, or with an incorrect percentile method, rendering all confidence intervals and the claimed 'enhanced stability' unreliable. Without a valid paired test (e.g., McNemar's) and proper patient-level CIs, the headline outperformance claim is not established. Even if the LGE-guided contrastive mechanism works as intended, the statistical evidence must be corrected before the central claim can be accepted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8185,"tokens_out":4899,"duration_ms":61045,"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":[{"comment":"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":"Section 3.2, Tables 1 and 2"},{"comment":"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":"Section 3.2, Table 1"},{"comment":"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":"Section 2.2, Eq. (2)"},{"comment":"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.","section":"Section 2.1 / Table 2"}],"minor_comments":[{"comment":"There is an apparent encoding corruption: a block of '/uni00000017/...' characters appears after Table 2. This should be removed.","section":"After Table 2"},{"comment":"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":"Section 3.1"},{"comment":"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.","section":"Section 2.3, Eq. (4)"},{"comment":"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.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the paper is generally well-structured, but the statistical presentation is not acceptable in its current form: impossible confidence intervals and overlapping intervals undermine the headline comparison. The central mechanistic claim about LGE-guided transfer is also not isolated by the ablation. These are fixable with additional analyses (valid patient-level CIs, a paired test, and control ablations), so I do not recommend rejection. However, the authors should also provide code or detailed implementation details to make the label-based loss reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nThe 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.\n\nThis 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.","headline":"Plausible training-time LGE-to-cine transfer idea, but the reported stats don't support the headline gain; the confidence intervals are internally inconsistent.","tokens_in":8662,"tokens_out":1718,"would_cite":false,"duration_ms":19051,"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":"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.","keywords":["cardiomyopathy screening","cardiac magnetic resonance","contrastive learning","cross-modal alignment","late gadolinium enhancement","gadolinium-free MRI","uncertainty-aware learning","video swin transformer"],"falsifier":"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.","tokens_in":7754,"feed_emoji":"🫀","tokens_out":7016,"duration_ms":74603,"temperature":0.7,"pith_summary":"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.","feed_headline":"Hits 94.3% accuracy on contrast-free cardiomyopathy screening","feed_subtitle":"LGE-guided contrastive training lets a cine-only model spot fibrosis, beating the prior cine-only state of the art by 4.3 points.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the state-of-the-art cine-CMR-only baseline that CC-CMR is compared against and that it outperforms.","marker":"[11]"},{"why":"Supplies the cross-modal contrastive alignment paradigm and the InfoNCE-style objective on which the FSA loss is built.","marker":"[9]"},{"why":"Supplies the focal loss used as the classification loss in the Feature Interaction Module.","marker":"[5]"},{"why":"Supplies the Video Swin Transformer backbone used as the feature extractor for each sequence.","marker":"[6]"},{"why":"Supplies the homoscedastic uncertainty weighting mechanism adapted for balancing the task losses.","marker":"[3]"},{"why":"Supplies the Efron bootstrap method used to estimate 95% confidence intervals for all reported metrics.","marker":"[2]"},{"why":"Supplies the contrastive predictive coding formulation that underlies the FSA loss equation.","marker":"[8]"},{"why":"Defines the ESC 2023 cardiomyopathy subtypes that the study cohort covers.","marker":"[1]"}],"fun_headline_variants":["Cine CMR learns fibrosis without gadolinium, hits 94.3% accuracy","Contrastive learning lets cine CMR spot fibrosis, no contrast needed","Gadolinium-free cardiomyopathy screening: 94.3% accuracy via LGE-aligned cine","Training with LGE teaches cine CMR to spot fibrosis, 94.3% accuracy"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Cine CMR learns fibrosis without gadolinium, hits 94.3% accuracy","Contrastive learning lets cine CMR spot fibrosis, no contrast needed","Gadolinium-free cardiomyopathy screening: 94.3% accuracy via LGE-aligned cine","Training with LGE teaches cine CMR to spot fibrosis, 94.3% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":2924,"prompt_tokens":751,"completion_tokens":2173,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":2089}},"tokens_in":495,"tokens_out":2173,"duration_ms":15462,"temperature":1.0,"reasoning_tokens":2089,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:07:10.956365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Nature Medicine 30(5), 1471–1480 (2024)","cited_arxiv_id":null,"evidence_quote":"Supplies the state-of-the-art cine-CMR-only baseline that CC-CMR is compared against and that it outperforms."},{"cited_title":"In: Proceedings of the 38th I nternational Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-modal contrastive alignment paradigm and the InfoNCE-style objective on which the FSA loss is built."},{"cited_title":"In: Proceedings of the IEEE International Confere nce on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Supplies the focal loss used as the classification loss in the Feature Interaction Module."},{"cited_title":"In: Proceedings of the IEEE/CVF conference on compu ter vision and pat- tern recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the Video Swin Transformer backbone used as the feature extractor for each sequence."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the homoscedastic uncertainty weighting mechanism adapted for balancing the task losses."},{"cited_title":"Chapman and Hall/CRC (1994)","cited_arxiv_id":null,"evidence_quote":"Supplies the Efron bootstrap method used to estimate 95% confidence intervals for all reported metrics."},{"cited_title":"European Heart Journal 44(37), 3503–3626 (2023)","cited_arxiv_id":null,"evidence_quote":"Defines the ESC 2023 cardiomyopathy subtypes that the study cohort covers."}],"review_version":1}