{"id":"f664ac62-e7fb-4238-83b6-365c62f068c2","arxiv_id":"2507.22024","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-stage MAE-plus-contrastive foundation model, Cardiac-CLIP aligns 3D cardiac CT with structured radiology text and reports state-of-the-art zero-shot and fine-tuned performance on classification, retrieval, and clinical risk prediction.","lead":"Cardiac-CLIP is a vision-language model for 3D cardiac CT that learns from over 130,000 scans and their radiology reports, then reports state-of-the-art results on cardiac abnormality classification, image-text retrieval, and clinical prediction tasks including future acute coronary syndrome. A generalist should read it because it suggests that a single pretrained model might reduce annotation burden and enable zero-shot cardiac screening across hospitals.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The zero-shot ACS AUROC (0.742) is reported without its text prompt; if the prompt is one of the seven trained abnormality templates, the 'emergent ACS' claim reduces to a trained correlate.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the zero-shot ACS result is presented without its text query, making it impossible to distinguish a genuinely emergent capability from a trained correlate. The paper has real strengths: a large private cardiac CT dataset, dual public-data evaluation (NLST and CT-RATE), and a reproducible code link. But the strongest claim in the abstract and Discussion depends on the ACS result, and the missing prompt directly determines whether that claim is supported. Because the reader already conditioned the verdict on this issue, my stress-test does not move the verdict; it reinforces it. If the authors disclose the prompt and confirm it is 'acute coronary syndrome' (or otherwise reproduce the result with a non-template prompt), the condition is satisfiable and the verdict could be upgraded. The loss-function ambiguity in Eq. 6 and the 'external' labeling of NLST are additional concerns, but the prompt question is the most load-bearing for the paper's central claim.","tokens_in":21879,"tokens_out":4774,"duration_ms":58671,"concrete_test":"Inspect the released GitHub inference code (github.com/Sliver-g/Cardiac-CLIP) for the exact ACS zero-shot prompt; then rerun the PD-2 (Jiangsu Province Hospital) split under two prompt families: (i) 'There is acute coronary syndrome' / 'There is no acute coronary syndrome', and (ii) each of the seven structured-report templates, most plausibly 'There is coronary stenosis' / 'There is no coronary stenosis'. If family (ii) reproduces AUROC ≈ 0.742 while family (i) is near chance, the emergent claim fails; if family (i) alone reproduces the result, the claim stands. Use the same 650-case split and zero-shot protocol as Sec. 3.4.1, and additionally verify patient-level separation between event and control cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4.1 reports a zero-shot ACS AUROC of 0.742 but never discloses the prompt text passed to the textual encoder. The only zero-shot prompts defined in the paper are the seven structured-report templates from Sec. 5.3.1 (e.g., 'There is coronary stenosis' / 'There is no coronary stenosis'). 'Acute coronary syndrome' is not one of those seven abnormality names, and the Discussion explicitly says the term is absent from the pre-training data. If the zero-shot ACS run used a trained template such as 'There is coronary stenosis' or 'There is coronary calcification', the result is a learned imaging correlate of future ACS, not evidence of emergent zero-shot diagnosis. The abstract, introduction, and Discussion all rest the strongest generalization claim on this 'emergent' ACS result, so the missing prompt is load-bearing. The absence of confidence intervals or a patient-level split description in Fig. 5(a) is secondary; the prompt ambiguity is the more direct threat.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents Cardiac-CLIP, a 3D vision-language foundation model for cardiac CT built through a two-stage pretraining pipeline: a 3D masked autoencoder trained on 130,889 CT volumes, followed by contrastive alignment of cardiac CT images with radiology reports. Free-text reports are structured into seven abnormality templates, converted into +/-1 pathology vectors, and used to build a soft-label affinity matrix for contrastive learning. The model is evaluated on abnormality classification in zero-shot and fine-tuning settings, image-text retrieval, and three clinical tasks: prospective two-year ACS prediction, functional coronary stenosis diagnosis, and CAC grading, using internal and external multi-center data. The paper reports consistent AUROC and recall improvements over Merlin, CT-CLIP, M3D, and 3D-ViT, and claims emergent zero-shot capability for ACS and CAC grading.","tokens_in":22101,"tokens_out":7548,"duration_ms":88825,"significance":"Cardiac-CLIP addresses a real gap: there is no large-scale cardiac CT vision-language foundation model, and the authors assemble an unusually large private multi-center dataset and make code available. If the claims hold, the model would offer zero-shot abnormality classification, clinically useful retrieval, and prospective risk prediction from 3D CT, which would be valuable. The soft-label contrastive idea is a plausible way to leverage structured diagnostic attributes, and the evaluation breadth across 12 institutions, public NLST and CT-RATE data, a temporal-shift cohort, and clinical endpoints is a strength. However, the strongest generalization claims currently rest on two under-specified components: the ACS zero-shot prompt and the contrastive loss target, together with dataset-separation ambiguities that must be resolved before the central conclusions can be relied upon.","major_comments":[{"comment":"The zero-shot ACS result (AUROC 0.742) is reported without stating the text prompt passed to the textual encoder. The only zero-shot prompts defined in the paper are the seven abnormality templates from Sec. 5.3.1, and Sec. 4 explicitly says that the term 'acute coronary syndrome' is not in the training data. If the prompt was one of the seven trained templates, such as 'There is coronary stenosis', the result is a learned correlate and not an emergent zero-shot diagnosis of ACS; if it was a novel phrase, the paper must supply the exact prompt and demonstrate robustness to prompt wording. This is load-bearing for the abstract's 'emergent capabilities' claim and for the Discussion's ACS paragraph. Please report the exact prompt, the event/control retrieval protocol, and confidence intervals for the 650-patient cohort.","section":"Section 3.4.1 / Fig. 5(a)"},{"comment":"The contrastive loss is written as CE(Softmax(S/tau), A) with A_ij = y_i . y_j / (||y_i|| ||y_j||). Since the pathology vectors have +/-1 entries, A entries lie in [-1,1] and rows are not normalized to sum to 1, so A is not a valid target distribution for cross-entropy. In addition, the diagonal of A is always 1, so the loss may be dominated by the diagonal rather than by the intended off-diagonal soft supervision. Please replace Eq. (6) with a well-defined objective, for example by row-normalizing and clipping the soft targets or by using a squared-error loss, and state the exact conversion from A to a probability target.","section":"Section 5.3.3, Eq. (6)"},{"comment":"The paper does not state that the 1,663 contemporaneous evaluation images from 946 patients are disjoint from the 16,641 pretraining images from 4,708 patients. Both cohorts are from Jinling Hospital over the same period, so without an explicit patient-level split the internal AUROC numbers in Fig. 3(c) and the ablation study in Fig. 6 could be inflated by training/evaluation overlap. Please specify the split and report evaluation metrics at the patient level.","section":"Section 2.2, PD-1"},{"comment":"NLST is described as an external dataset for classification, but all NLST cardiac-cropped images except the held-out test set are included in the first-stage MAE pretraining (114,248 images). A model pretrained on unlabeled NLST images is not external to NLST for the purpose of measuring transfer; any NLST gain could reflect in-distribution representation learning rather than generalization. Please either retrain a version without NLST for the external evaluation, or explicitly describe the NLST classification result as same-distribution pretraining rather than external validation.","section":"Section 3.2.2 and Section 2.1 (OD-1)"},{"comment":"The zero-shot CAC grading protocol uses the prompt 'There is Coronary Artery Calcium' as a proxy for calcification severity, but the paper does not explain the relationship between this prompt and the seven pretraining abnormality names, nor why a cosine similarity to a single presence/absence prompt should be monotone in Agatston grade. As written, the ordinal AUROC table in Fig. 5(c) cannot be reproduced, and the comparison with other foundation models under unspecified zero-shot protocols is not apples-to-apples. Please provide the exact prompts, thresholds, and normalization used for all models.","section":"Section 3.4.3 and Section 5.3.1"}],"minor_comments":[{"comment":"The total number of pretraining scans is given as 130,889 in the introduction, 130,899 in Section 2.1, and 130,889 in the abstract; please make the counts consistent.","section":"Section 1 and Section 2.1"},{"comment":"The text says 'As shown in Fig. 3(d)' for the fine-tuning evaluation, but in the caption Fig. 3(d) is the external zero-shot panel and Fig. 3(e) is the fine-tuning panel; please fix the cross-references.","section":"Section 3.2.2"},{"comment":"The ablation table does not show which rows correspond to which combination of MAE, structured report, and soft matrix; add checkmarks or a configuration column so the four configurations can be read directly.","section":"Figure 6"},{"comment":"The structured reports are generated with GPT-4o, but no quality-control or error analysis is reported; a small validation of the structuring on a random sample would help assess label noise.","section":"Section 5.3.1"},{"comment":"The paper alternates between '3D-ViT' and 'M3DNet/M3D' naming and between 'AUC' and 'AUROC'; please unify the terminology.","section":"Throughout"},{"comment":"Reference [43] is formatted with 'null null' as the author list; use the actual National Lung Screening Trial Research Team author string.","section":"References"},{"comment":"In Eq. (2), 'construction loss' should be 'reconstruction loss'.","section":"Eq. (2)"},{"comment":"The statement that the radiology reports contain related terminology such as 'coronary', 'stenosis', and 'calcification' is helpful, but it should be connected to the prompt disclosure requested for Section 3.4.1 rather than appearing only in the Discussion.","section":"Section 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a strong candidate for a medical imaging venue if the technical definitions are fixed. The undisclosed ACS prompt and the ill-defined loss in Eq. (6) are addressable in revision, and the dataset/pretraining overlap can be clarified. I would not recommend rejection, because the central two-stage idea is plausible and the empirical breadth is substantial. Please also verify that the GitHub link contains the actual code and that the use of GPT-4o on protected health information satisfies the IRB and data-sharing conditions of the contributing hospitals."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The genuinely new piece is the first dedicated 3D cardiac CT vision-language foundation model, trained on 11k real cardiac CT-report pairs plus chest CT for MAE, and the soft-label pathology-vector affinity matrix. That soft-label supervision is the paper's most original idea, and the ablation suggests it helps (avg AUROC 0.807 vs 0.771 on CCTA without it). The multi-center evaluation across 12 institutions and three hard clinical tasks is ambitious, and the fine-tuning results on functional coronary stenosis and CAC grading are credible evidence that a cardiac-specific pretraining recipe beats generic 3D models like Merlin, CT-CLIP, and M3D.\n\nThe soft spots are real but fixable. The biggest is the zero-shot ACS claim. Section 3.4.1 reports AUROC 0.742 without saying what text prompt was used. If it was one of the seven trained templates, e.g. 'There is coronary stenosis,' the model is being tested on a learned correlate, not on the unseen concept 'acute coronary syndrome.' The Discussion's 'emergent' language depends entirely on this, so the prompt must be disclosed. The stress-test note is right that this is load-bearing.\n\nEq. 6 is also sloppy: cross-entropy requires target rows to be valid probability distributions, but the affinity matrix contains negative cosine values and rows that don't sum to one. It is presumably meant as row-normalized soft targets or a distance-weighted contrastive loss, but as printed it is not well-defined. Calling NLST 'external' for classification is misleading, since NLST images were used in stage-one MAE pretraining; excluding labels does not make it external in the usual sense. And most headline numbers in Figs. 3-5 come without error bars or patient-level split details, so the precision of the comparisons is hard to judge.\n\nNone of this kills the central pretraining idea, and the internal contemporaneous/temporal results plus the FCS and CAC results look reasonably solid. Cardiac-imaging researchers and anyone building domain-specific medical VLMs will get value from this paper, especially if the authors release the code and weights as promised. It deserves serious refereeing, but only after the authors disclose the ACS prompt, fix the loss definition, and re-label NLST.","headline":"First dedicated 3D cardiac CT vision-language model with a genuinely new soft-label contrastive loss, but the zero-shot ACS claim needs prompt disclosure before the 'emergent' language can be taken at face value.","tokens_in":22675,"tokens_out":2561,"would_cite":true,"duration_ms":28786,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Cardiac-CLIP: a single 3D cardiac CT vision-language model that, the paper argues, outperforms existing 3D medical foundation models on classification, retrieval, and clinical risk prediction, including zero-shot two-year acute coronary…","keywords":["3D cardiac CT","vision-language pretraining","masked autoencoder","contrastive learning","soft-label matrix","zero-shot classification","acute coronary syndrome prediction","coronary artery calcium grading"],"falsifier":"Inspect the exact text query used for the zero-shot ACS experiment in Section 3.4.1. If the query is one of the structured templates (for example, 'There is coronary stenosis') or includes the phrase 'acute coronary syndrome', then the result is a trained correlate rather than an emergent capability; a definitive test is to retrain the second stage with all ACS-related wording removed from the reports and check whether the zero-shot AUROC still exceeds 0.7.","tokens_in":21673,"feed_emoji":"🫀","tokens_out":9102,"duration_ms":88908,"temperature":0.7,"pith_summary":"Cardiac-CLIP is a vision-language model built for 3D cardiac CT volumes paired with radiology reports, and the paper claims that one pretrained model can classify cardiovascular abnormalities, retrieve relevant reports and images, and support clinical risk prediction across hospitals without per-site retraining. The central claim is that a two-stage recipe—self-supervised 3D masked autoencoding followed by soft-label contrastive alignment with standardized reports—produces representations that transfer better than existing 3D medical foundation models in both zero-shot and fine-tuned settings. The strongest evidence is the clinical analysis: prospective two-year acute coronary syndrome risk reaches 0.742 AUROC zero-shot and 0.802 after fine-tuning, functional coronary stenosis reaches 0.782, and coronary artery calcium grading reaches an average ordinal AUROC of 0.844 zero-shot. If these results hold, a single cardiac CT foundation model could reduce the annotation burden and domain-shift problems that currently limit AI-based cardiovascular diagnostics.","feed_headline":"A 3D cardiac CT model predicts future acute coronary syndrome","feed_subtitle":"Pretrained on 130k scans and reports, it also grades calcium and detects stenosis without task-specific labels.","key_machinery":"The load-bearing component is the soft-label affinity matrix that replaces the usual one-hot contrastive target. Each scan is summarized as a pathology vector of seven binary attributes (for example, coronary stenosis, coronary calcification, cardiomegaly), and pairwise cosine similarities between these vectors become the supervision matrix $\\tilde{A}$ used to train the image-text similarity matrix. The visual encoder is a vision transformer with 3D patch embeddings, pretrained with a 75%-masked autoencoder on 130k volumes, and the text encoder is a biomedical language model. Free-text reports are first standardized into 'There is [abnormality]' and 'There is no [abnormality]' statements, and the contrastive loss is symmetrized cross-entropy between $\\text{Softmax}(S/\\tau)$ and $\\tilde{A}$. This machinery lets the model treat cases with partially overlapping pathology as softly positive pairs, which is what the authors credit for fine-grained cardiovascular semantics.","core_discovery":"The paper's central discovery is that contrastive pretraining against standardized radiology reports with soft labels gives a cardiac-CT vision-language model substantially better zero-shot and transfer performance than prior 3D medical foundation models. On internal and external classification, Cardiac-CLIP consistently leads across seven abnormalities; on a public chest CT dataset never used in pretraining, it outperforms a model trained directly on that dataset. The paper identifies the clinical-analysis results as emergent capabilities: although terms like 'acute coronary syndrome' and 'functional stenosis' and explicit calcification grades were not part of the training text, the model still separates ACS events from controls, identifies hemodynamically significant stenosis, and ranks calcium severity. The reported numbers—0.742 zero-shot AUROC for ACS, 0.782 for functional stenosis, and 0.844 average ordinal AUROC for calcium grading—are the quantitative core of the claim.","pith_inferences":["The paper does not disclose the exact text query used for the zero-shot ACS prediction; if that query is one of the seven trained report templates, then the 0.742 AUROC would be a trained correlate rather than an emergent zero-shot ability, so the claim is testable immediately.","Because the soft-label matrix is built from only seven abnormality attributes, the apparent emergence of ACS and functional-stenosis knowledge may be mediated entirely by those attributes; ablating each attribute from the pathology vectors would reveal which cues carry the prognostic signal.","The same soft-label contrastive recipe could transfer to other 3D modalities with paired reports, such as abdominal or neuro CT, provided a small set of report attributes can be standardized; the paper's comparisons suggest that domain-specific attribute construction, rather than model scale alone, drives the gains.","Real-world retrieval would likely use free-text queries rather than the structured templates; testing with raw reports and clinician-typed phrases would clarify whether the alignment survives realistic input."],"forward_implications":["A single Cardiac-CLIP pretraining provides zero-shot classification of seven cardiovascular abnormalities across contemporaneous, temporally shifted, and external datasets, with the largest gains over prior 3D foundation models exceeding 30 AUROC points on cardiomegaly and pericardial effusion.","The visual encoder transfers with little supervision: fine-tuning on only 10% of labeled data already gives competitive average AUROC on the two large public CT evaluation sets.","Zero-shot keyword-guided retrieval, using phrases like 'There is coronary stenosis', retrieves the correct cardiac CTs better than prior models on internal and external pools, including on a public chest CT dataset whose training split was never seen in pretraining.","Prospective two-year acute coronary syndrome risk can be predicted from a prior CCTA at 0.742 AUROC with no fine-tuning, and at 0.802 after fine-tuning, while a fully supervised 3D vision transformer baseline reaches only 0.529.","Coronary calcium severity is inferable from non-contrast CT without explicit score labels: the zero-shot ordinal AUROC averages 0.844 across grade thresholds."],"supporting_citations":[{"why":"Supplies the contrastive vision-language learning paradigm that Cardiac-CLIP extends to 3D cardiac CT.","marker":"[22]"},{"why":"Masked autoencoder used in first-stage pretraining to learn visual representations from unlabeled volumes.","marker":"[44]"},{"why":"Provides a public chest CT dataset with paired reports used for external evaluation and as a comparison model.","marker":"[37]"},{"why":"Provides cardiac-region cropping and the public lung-screening evaluation split used in the classification experiments.","marker":"[12]"},{"why":"Large public CT screening dataset used for first-stage MAE pretraining and external evaluation.","marker":"[43]"},{"why":"Used to standardize free-text radiology reports into structured abnormality statements for contrastive training.","marker":"[57]"},{"why":"Biomedical text encoder that produces the textual features aligned with the visual encoder.","marker":"[59,60]"}],"fun_headline_variants":["Cardiac-CLIP zero-shot heart CT model beats task-trained rivals","Heart CT vision-language model predicts acute coronary syndrome zero-shot","130k scans teach 3D cardiac CT model to spot unseen abnormalities","Soft-label contrastive pretraining enables zero-shot ACS on cardiac CT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the acute coronary syndrome zero-shot result is genuinely zero-shot: the text query used to get the 0.742 AUROC is not one of the seven trained report templates and does not otherwise expose the model to ACS-related wording when it learns to match images to reports.","fun_headline_variants_meta":{"raw":{"variants":["Cardiac-CLIP zero-shot heart CT model beats task-trained rivals","Heart CT vision-language model predicts acute coronary syndrome zero-shot","130k scans teach 3D cardiac CT model to spot unseen abnormalities","Soft-label contrastive pretraining enables zero-shot ACS on cardiac CT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000925,"raw_usage":{"total_tokens":4002,"prompt_tokens":1018,"completion_tokens":2984,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":2909}},"tokens_in":634,"tokens_out":2984,"duration_ms":24804,"temperature":1.0,"reasoning_tokens":2909,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:06:10.351152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the exact text query used for the zero-shot ACS experiment in Section 3.4.1. If the query is one of the structured templates (for example, 'There is coronary stenosis') or includes the phrase 'acute coronary syndrome', then the result is a trained correlate rather than an emergent capability; a definitive test is to retrain the second stage with all ACS-related wording removed from the reports and check whether the zero-shot AUROC still exceeds 0.7.","supporting_citations":[{"cited_title":"Masked autoencoders are scalable vision learners","cited_arxiv_id":null,"evidence_quote":"Masked autoencoder used in first-stage pretraining to learn visual representations from unlabeled volumes."},{"cited_title":"Developing generalist foundation models from a multimodal dataset for 3d computed tomography, 2024","cited_arxiv_id":null,"evidence_quote":"Provides a public chest CT dataset with paired reports used for external evaluation and as a comparison model."},{"cited_title":"Deeplearningpredictscardiovascular disease risks from lung cancer screening low dose computed tomography.Nature Communications, 12(1), May 2021","cited_arxiv_id":null,"evidence_quote":"Provides cardiac-region cropping and the public lung-screening evaluation split used in the classification experiments."},{"cited_title":"Reduced lung-cancer mortality with low-dose computed tomographic screening.New England Journal of Medicine, 365(5):395–409, August 2011","cited_arxiv_id":null,"evidence_quote":"Large public CT screening dataset used for first-stage MAE pretraining and external evaluation."}],"review_version":1}