Abstract

Cardiovascular diseases are the leading cause of death. Cardiac phenotypes (CPs), e.g., ejection fraction, are the gold standard for assessing cardiac health, but they’re derived from cine cardiac magnetic resonance imaging (CMR), which is costly and requires high spatio-temporal resolution. Every magnetic resonance (MR) examination begins with rapid and coarse localizers for scan planning, which are discarded thereafter. Despite non-diagnostic image quality and lack of temporal information, localizers can provide valuable structural information rapidly. In addition to imaging, patient-level information, including demographics and lifestyle, influence the cardiac health assessment. Electrocardiograms (ECGs) are inexpensive, routinely ordered in clinical practice, and capture the temporal activity of the heart. Here, we introduce C-TRIP (Cardiac Tri-modal Representations for Imaging Phenotypes), a multi-modal framework that aligns localizer MRI, ECG signals, and tabular metadata to learn a robust latent space and predict CPs using localizer images as an opportunistic alternative to CMR. By combining these three modalities, we leverage cheap spatial and temporal information from localizers, and ECG, respectively while benefiting from patient-specific context provided by tabular data. Our pipeline consists of three stages. First, encoders are trained independently to learn uni-modal representations. The second stage fuses the pre-trained encoders to unify the latent space. The final stage uses the enriched representation space for CP prediction, with inference performed exclusively on localizer MRI. Proposed C-TRIP yields accurate functional CPs, and high correlations for structural CPs. Since localizers are inherently rapid and low-cost, our C-TRIP framework could enable better accessibility for CP estimation.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3631_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/basrazey/C-TRIP

Link to the Dataset(s)

UK Biobank Dataset: https://www.ukbiobank.ac.uk/

BibTex

@InProceedings{ZeyBus_Opportunistic_MICCAI2026,
        author = { Zeybek, Busra Nur AND Turgut, Özgün AND Zhang, Yundi AND Pan, Jiazhen AND Graf, Robert AND Starck, Sophie AND Rueckert, Daniel AND Kafali, Sevgi Gokce},
        title = { { Opportunistic Cardiac Health Assessment: Estimating Phenotypes from Localizer MRI Through Multi-modal Representations } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The author proposes a three-stage model for generating cardiac phenotypes from MRI localiser scans using contrained from ECG and tabular data. The pipeline includes separate encoders for imaging, ECG, and tabular data. The authors introduce a localiser-centric contrastive learning framework in which the localiser latent space is aligned with ECG and tabular representations, encouraging embeddings from the same patient to be closer in the shared latent space. The localiser embedding is then used to predict cardiac phenotypes, and only the localiser images is required at inference. The model was trained and validated on UK Biobank data, but the specific image inputs are not clearly described despite the dataset’s range of available cardiac scans. Overall, phenotype prediction using the localiser appears comparable to the baseline supervised model based on dynamic CMR sequences.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The network architecture and optimisation method are clearly described.
    • The proposed method could be of considerable clinical interest for public health and patient management.
    • The proposed multimodal contrastive learning approach appears flexible and extensible, and may be of interest for other applications.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • There is no discussion of model performance for patient diagnosis. The proposed method does not appear to perform well for lower-range LVEF and RVEF prediction, which are among the most important biomarkers. Performance analysis and discussion within the diagnostic range should be provided to improve transparency and strengthen the clinical relevance of the method, particularly given its intended use for opportunistic screening.
    • The evaluation metrics do not provide sufficient insight for clinical application. The mean difference and limits of agreement provide some insight into prediction stability for models comparison, but they may also be misleading. The absolute differences between the predicted CPs and the ground truth should also be reported for a more robust evaluation.
    • The description of the input localiser images is ambiguous. The author should clearly state which data were used as localiser images and whether scouting images were included. If so, the specific scouting sequences should be identified. Since the UK Biobank dataset follows a defined imaging protocol, specifying the exact image sequence used as input would remove any ambiguity regarding the data used for each model.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The author proposed interesting approach for combining patient multimodal data as contrained in the model optimisation. However, the author did not show sufficient proof/analysis/discussion supporting its main utilisation in opportunistic screening.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The author has adequately justified the limitations of the proposed method. Although the method still requires further development in its current form, it can serve as a foundation for future work aimed at achieving the clinical goal the author set out to address. With revised wording to clarify the method’s limitations, the proposed method provides a solid basis for future work in opportunistic screening.



Review #2

  • Please describe the contribution of the paper

    This paper proposes C-TRIP, a localizer-centric multimodal contrastive learning framework for estimating cardiac phenotypes from routine MRI localizers. The method first pretrains modality-specific encoders for localizer MRI, ECG, and tabular data using masked data modeling, then aligns the localizer representation with ECG and tabular embeddings through a contrastive objective, and finally predicts cardiac phenotypes using localizer MRI only at inference time.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper addresses a clinically relevant and original problem: estimating imaging-derived cardiac phenotypes from routinely acquired, non-diagnostic localizer MRI.
    • The localizer-centric tri-modal alignment is an interesting methodological contribution, and the multimodal training with single-modality inference is well motivated.
    • The manuscript is generally clear and well organized, and the figures support the narrative well, particularly the attention maps and latent-space visualization.
    • The evaluation is thoughtful. First, the authors evaluate their method against Cine MR (CMR), which is the standard technique for cardiac structure and function measurements. Second, the evaluation goes beyond a single regression metric by reporting mean difference (MD), limits of agreement (LoA), and Pearson correlation.
    • The results support the main claim that the proposed method is particularly promising for structural phenotypes, while CMR remains strongest for functional phenotypes, which is an important and appropriately nuanced conclusion.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Choice of auxiliary modalities in contrastive alignment: The method aligns localizer MRI with ECG and tabular data, but not with CMR, even though CMR is available in the dataset and is the reference modality for phenotype extraction. Please clarify the rationale for excluding CMR as an auxiliary alignment modality. Since the paper aims to enrich localizer representations, it would be helpful to understand whether this was a deliberate conceptual choice, a fairness consideration, or a practical design decision.
    • Missing explicit details about technical choices:
    • Please clarify whether the modality-specific encoders are frozen or fine-tuned during Stage II contrastive alignment. The manuscript says the three pre-trained encoders are fused and optimized with a contrastive loss, but does not clearly specify which components remain trainable.
    • Please clarify the exact input to the Stage III regression head during training. The manuscript states that phenotype prediction is performed solely from localizers during inference, but no details about its training are provided.
    • Please provide the architectural details of the supervised baselines used for comparison. The current description of the baselines is too brief to assess whether comparisons are fully controlled.
    • Potential unfairness in the CMR baseline design: The CMR supervised baseline uses only three short-axis time frames from the middle baso-apical slice. This seems sufficient to provide some functional information, but it may be suboptimal for structural phenotype estimation compared with the localizer input, which includes three consecutive oblique-transverse slices and may provide richer spatial cues despite lower image quality. Please comment on the rationale for this restricted CMR representation and whether it may disadvantage the CMR baseline, especially for structural phenotypes.
    • Table 2 presentation: Please verify that the boldface and shading in Table 2 consistently correspond to the best absolute MD and tightest LoA, as stated in the caption.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
    • Please briefly clarify what is meant by “demographically balanced” train/validation/test splits. A one-sentence explanation would be sufficient.
    • Please add one sentence describing how the UMAPs were computed, including the source embeddings and any key settings if relevant.
  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I support a Weak Accept for this paper because it addresses a clinically relevant and original problem: estimating cardiac phenotypes from routinely acquired localizer MRI. The paper proposes an interesting methodology, using localizer-centric multimodal contrastive learning to align localizer, ECG, and tabular representations while keeping localizer-only inference at test time, which is both methodologically meaningful and practically relevant as an opportunistic low-cost screening strategy. The paper is generally clear, well organized, and supported by informative figures and a thoughtful evaluation. The results support the authors’ main claim in a balanced way: the method appears especially promising for structural phenotypes, while cine CMR remains stronger for functional phenotypes, which I find credible. My score is not higher because some important implementation details remain unclear. Additionally, I believe the rationale for excluding CMR as an auxiliary modality in the contrastive learning stage should be clarified, since incorporating the reference imaging modality might plausibly further strengthen the learned representation and improve performance. More importantly, I believe that the restrictive preprocessing of the CMR data for baseline evaluation may make the comparison less convincing for structural phenotype estimation. Overall, I believe the paper makes a worthwhile contribution that would interest the MICCAI audience, but it still has enough methodological ambiguity that Weak Accept is the most appropriate recommendation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    The rebuttal successfully addressed my questions and comments. While some limitations remain, the authors acknowledge them appropriately and provide reasonable clarifications. Overall, the paper remains novel, relevant, and of interest to the MICCAI community, and I support acceptance.



Review #3

  • Please describe the contribution of the paper

    This paper proposes C-TRIP, a tri-modal contrastive learning framework that aligns 1) localizer MRI (cheap, structural), 2) ECG (temporal), 3) tabular metadata (context) to generate an embedding space that enables predicting cardiac phenotypes (CPs). At inference only localizer MRI are used. The core idea, that routinely discarded localizer scans should be used for opportunistic screening is potentially clinically meaningful.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper is motivated by a clear and clinically relevant problem, namely the cost and complexity of CMR, and proposes a novel and practical idea of repurposing routinely acquired localizer scans within an opportunistic screening framework during MRI scans. The experimental design is comprehensive, including a range of unimodal and multimodal baselines, targeted ablations such as the tabular leakage setting, and a valuable analysis of performance in low-data regimes, which is often missing in similar work. Additionally, the authors make a meaningful effort toward interpretability, with attention maps demonstrating improved anatomical focus over baselines and UMAP visualizations providing insight into the learned representation space.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Functional Estimation Without Temporal Information is Fundamentally Underdetermined The proposed method seeks to estimate functional cardiac phenotypes (e.g., EF), which are defined by temporal dynamics between end-diastolic and end-systolic states, from static localizer images that lack phase information. This renders the problem fundamentally underdetermined, as the underlying physiological signal is not directly observable. Consequently, the model must rely on indirect structural correlates and population-level priors, rather than true functional measurement. This limitation is reflected in Fig. 3, where predictions exhibit regression-to-the-mean behavior, particularly for functional metrics. While the approach aims to approximate cine-derived biomarkers, these quantities are intrinsically temporal and cannot be recovered from static inputs. Moreover, even volumetric measures such as EDV and ESV are not directly observable from localizers and are instead inferred via proxy signals, making them fundamentally distinct from cine-derived measurements.

    Limited Evaluation on a Healthy-Skewed Cohort Masks Model Limitations The use of the UK Biobank cohort likely obscures the extent of the method’s limitations. This dataset is predominantly composed of healthy individuals, resulting in a relatively narrow distribution of functional metrics such as EF and a low prevalence of severe pathology. In this setting, regression-to-the-mean behavior is not strongly penalized, as predicting values within the typical range (55–65% EF) yields superficially reasonable performance. Consequently, the reported correlations may overstate the model’s practical utility. Crucially, the model is not evaluated under conditions that reflect clinically relevant variability, and its ability to detect pathological states remains unclear. It is therefore unknown whether the method can reliably identify heart failure or distinguish between preserved and reduced EF, which are central requirements for any screening or diagnostic application.

    Excluding ECG at Inference is Poorly Justified The decision to exclude ECG at inference is not well justified. ECG is a low-cost and widely available modality that directly captures cardiac temporal dynamics, which are essential for functional assessment. By discarding ECG at test time, the model is forced to infer functional phenotypes from static structural images, an inherently limited and ill-posed setting reflected in the regression-to-the-mean behavior observed in Fig. 3.Incorporating ECG at inference would likely improve both performance and clinical relevance; at minimum, this setting should be evaluated.

    Recommendations on Overstated Clinical Claims Relative to Evidence “Enable reliable CP estimation when comprehensive CMR cannot be performed.” This implies clinically actionable, patient-level accuracy, which is not supported given the regression-to-the-mean behavior and limited functional performance; revise to state that the method provides approximate, population-level estimates rather than reliable clinical measurements. “Accurate functional CPs” and “cardiac health assessment.” These claims overstate capability because functional metrics require temporal information absent at inference, meaning the model estimates proxies rather than true function; reframe explicitly as indirect or proxy estimation of functional phenotypes. “Opportunistic low-cost screening tool.” This is not substantiated, as no evaluation is performed on diseased populations or using diagnostic metrics such as sensitivity or specificity; limit the claim to exploratory or preliminary assessment and avoid positioning as a screening tool without appropriate validation. “Not a replacement for CMR.” While cautious in wording, the comparison is made against a reduced CMR baseline (single slice, limited frames), understating the true performance gap to full cine MRI; clarify this limitation and avoid implying proximity to the clinical gold standard.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a novel and interesting embedding-based approach, clearly motivated by addressing a relevant clinical problem, namely the use of routinely acquired localizer MRI for opportunistic cardiac assessment. The experimental design is strong, with well-structured and informative ablation studies, and analysis in low-data regimes, all of which support the technical contribution and make the work of interest to the community. However, the central claims regarding clinical applicability and reliability (particularly for functional cardiac phenotype estimation and screening) are not fully supported by the presented evidence. The absence of temporal information at inference and the observed regression-to-the-mean behavior raise concerns about the validity of functional predictions, and the exclusion of ECG at test time is not well justified. These limitations do not invalidate the contribution but suggest that the clinical claims should be more carefully framed and substantiated during rebuttal, supporting a weak accept recommendation.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all the reviewers for valuable feedback.

Major comments (R1, R2, R3) “Limitations on functional CP estimation & healthy volunteer bias” -We acknowledge the reduced performance of estimating functional CPs from localizers. Although localizers are static, the model still learns phenotype-relevant temporal relations implicitly through multimodal training with ECG. We’ll state this more clearly

-We agree; including ECG at inference could improve functional CP estimation due to explicit temporal information, and remains a future work as new results are not allowed. Yet, ECG is not temporally aligned or jointly acquired with MRI in UK Biobank, which may limit the benefit of joint inference

-We decided to include only localizers at inference to mimic clinically realistic/limited scenarios with readily-available imaging and to push the limits of the localizer’s capability to estimate CPs. Each localizer acquisition spans ~1.1 seconds, exceeding a typical cardiac cycle (~0.8–1s), and may thus preserve weak phase-averaged cues related to cardiac function. When combined with ECG during training, the cardiac phase information is implicitly learnt by C-TRIP, as shown by increased performance w.r.t. localizer baseline

-Despite the UKBiobank’s healthy volunteer bias (Fry et al., 2017), its scale helps capture robust correlations with ~30% of our dataset exhibiting reduced/mildly reduced LVEF (<55%). Crucially, ablation studies show C-TRIP improves regression-to-the-mean behavior seen in tabular baseline, suggesting it does not rely solely on population averages. Besides, our aim is opportunistic assessment from routine readily-available data, which involves scanning healthy populations to extract any incidental findings

“Clinical claims, real-world use cases, patient diagnosis” -We will frame our contribution as an opportunistic assessment of cardiac function and structure rather than comprehensive opportunistic screening (OS). While we did not evaluate disease classification tasks, estimating CPs from routinely acquired localizers is still relevant for the initial steps of OS, especially to detect potential signs of cardiac dysfunction. The next steps will include diagnosis with risk classification according to EF and disease detection. Our high accuracy in LVM estimation could especially be useful in detecting LV hypertrophy, a result of coronary artery disease. Other real-world applications may include preliminary assessment in resource-limited settings, more importantly a scenario where the model enables real-time decision-making on the necessity of a full CMR protocol

Minor comments R1 Metrics Following the Bland–Altman analysis, we report MD & LoA, which are prioritized in clinical applications (Obuchowski et al., 2015). If allowed, we’ll add MAE, which showed similar trends, to our results

R1 Localizers The localizers are 3 mid-slices from the 18-slice “AA_Heart_Scout” scan

R2 CMR as auxiliary modality Excluding CMR was a deliberate choice to test whether low-cost and non-imaging modalities (ECG/tabular) could enrich localizers. We’ll state that incorporating CMR during pre-training is future work, as previous work proved that CMR-enriched ECG representations can predict CPs (Turgut et al., 2025)

R2 Implementation Stage II: Encoders are fully fine-tuned end-to-end Stage III: the regression head is trained using only localizer embeddings extracted via the Stage II aligned encoder Baselines use the same regression head as C-TRIP, but are trained from scratch without contrastive learning

R2 CMR baseline The CMR baseline uses three neighboring SA time frames from the same slice while keeping the number of input channels identical across all models for architectural fairness & computational simplicity

R2 Formatting We’ll fix Table 2 Demographically balanced: train/val/test splits have statistically similar distributions of age, sex, LVEF UMAPs used L2-normalized joint embeddings, cosine dist., 15 neighbors, min_dist=0.1




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper proposes C‑TRIP, a multimodal contrastive learning framework that aligns localizer MRI with ECG and tabular metadata during training to enable cardiac phenotype estimation using localizer MRI alone at inference.

    The idea is that localizer scans are routinely acquired but discarded, and could therefore be repurposed for opportunistic cardiac assessment. The method is evaluated on the UK Biobank dataset with extensive baselines, ablations, interpretability analyses, and low-data regime experiments.

    A major concern raised is that the estimation of functional cardiac phenotypes from static localizer images is limited due to the absence of temporal information, with observed performance likely driven by regression to population priors on a predominantly healthy UK Biobank cohort. Additionally, there are ambiguities in methodological details pointed out by the reviewers that need to be addressed.

    I also struggle to see a real-world clinical scenario in which the proposed method would be used. In routine clinical practice, cardiac MRI examinations are explicitly performed to obtain cine imaging, and cine sequences are acquired in essentially all patients undergoing CMR. Situations in which localizer MRI is available but cine data is not and cardiac phenotyping is required are rare. Can the authors further explain their vision for potential use cases?

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    All reviewers agree on recommending acceptance after the rebuttal. The limitations and framing should be updated in the final version.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    This study proposes a promising and interesting network C-TRIP to predict cardiac phenotypes (CPs) using localizer MRI images, ECG signals, and clinical data. The method consists of three stages: pre-training single-modality encoders, multi-modal contrastive learning, and prediction inference. Main concerns raised by reviewers include the evaluation metrics, application in clinical practice, rational of the techniques used, and some overstated claims. The authors’ response acknowledged them and explained properly.



back to top