List of Papers Browse by Subject Areas Author List
Abstract
Myocardial infarction (MI) is a significant health burden globally. Its precise prediction is critical yet complicated by the functional complexities of the heart and heterogeneous clinical presentations. Although learning-based methods that model the 3D heart anatomy have been widely studied, improving cardiac embeddings with localized substructures in a multi-task setting, remains under-explored. In this work, we present a novel deep learning model that produces explainable embeddings with high relevance to cardiac function via multi-task learning. Its transformer-based architecture contains modules for both MI classification and cardiac substructure prediction. By jointly learning these tasks with shared embeddings, the model is able to better capture 3D cardiac geometries and deformation across cardiac phases, enhancing its predictive ability. We evaluate the proposed method on cardiac anatomies captured during end-diastolic and end-systolic phases from the UK Biobank study. Compared to the existing learning-based benchmarks, our method exhibits high predictive performance, achieving an area under the receiver operating characteristic curve for MI prediction of 0.802. We also demonstrate the strong explainability of our model by showing that the latent features generated under the proposed multi-task setting have a strong and statistically significant correlation with key clinical markers, such as ejection fraction.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3121_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
UK Biobank Dataset: https://www.ukbiobank.ac.uk/
BibTex
@InProceedings{PenJia_An_MICCAI2025,
author = { Peng, Jiachuan and Beetz, Marcel and Banerjee, Abhirup and Chen, Min and Grau, Vicente},
title = { { An Anatomical Significance-Aware Architecture for Explainable Myocardial Infarction Prediction via Multi-Task Learning } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15973},
month = {September},
page = {13 -- 23}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents a transformer-based multi-task learning framework for myocardial infarction (MI) prediction from 3D cardiac anatomical point clouds. The core contribution lies in the joint optimization of MI classification and cardiac substructure prediction, using shared latent embeddings. This auxiliary task is designed to enforce spatial awareness and improve the interpretability of the learned representations. Compared to prior geometry-based point cloud models, the proposed framework is shown to achieve state-of-the-art performance on UK Biobank data, with strong clinical relevance supported by correlation to functional biomarkers (e.g., ejection fraction, myocardial mass). The approach advances cardiac representation learning through anatomically-informed, explainable deep learning.
- 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.
1). Novel multi-task formulation: employs Farthest Point Sampling (FPS) and K-Nearest Neighbors (KNN) to generate spatially coherent patches from biventricular point clouds. Conventional method takes the whole heart as the input. 2). The authors go beyond standard metrics and provide SHAP-based feature importance, correlation with clinical biomarkers, and attention map visualizations, demonstrating that the model focuses on clinically relevant myocardial regions. 3). Clear architecture and rationale: The paper is well-organized with solid motivations for using transformers, patchifying 3D anatomy, and using multi-task training to counter overfitting and sparse labels.
- 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.
but the paper also misses some technical detail, making it leave some space for readers to imagine. my concerns are mainly about the patch and postprocess. 1)How does the model ensure that the g × k patch points cover all input 3D points? The paper sets g × k = 2N, which implies redundant point coverage to increase representation density However, it does not explicitly state whether all N points are guaranteed to be included or if there may be unrepresented regions. 2). What if overlapping patches assign different substructure labels to the same point? this will have influences when computing the clinical metrics. 3).Key architecture details are missing, sush as Patch embedding dimension, Hidden size of transformer layers, MLP head structure, which makes it difficult for reproducibility.
- 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 submission does not provide sufficient information for reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Given that MI is strongly related to the myocardium, the authors could further enhance explainability by reporting average attention weights assigned to each anatomical region (LV epi, LV endo, RV endo) to verify whether the model truly focuses on the myocardial region.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a solid and innovative contribution to the domain of 3D cardiac disease modeling with a method that integrates anatomical priors via substructure prediction. It is clinically relevant, well-executed, and yields competitive results. However, the paper lacks a few important implementation and methodological details (e.g., label generation, model architecture specs), which weakens its reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Review #2
- Please describe the contribution of the paper
The paper at hand describes the prediction of a future myocardial infarction using geometric deep learning in a multi-task learning setting. With the proposed approach the authors report an improved state of the art for the task at hand.
- 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.
- Leveraging the substructure prediction to enhance the overall performance seems like an elegant way to induce prior knowledge into the model.
- Details on the implementation and hyperparameterization are clearly stated.
- The performance achieved is quite impressive.
- The conducted evaluations are sound and interesting.
- 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.
- In general, the task of future myocardial infarction could be formulated as a survival regression task and it would be interesting to also report the linked metrics for this task.
- Furthermore, reporting the metrics for the substructure predictions would be interesting.
- 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
With the given novelty, good performance and performed evaluation I see no reason why not to accept this paper.
- 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
Review #3
- Please describe the contribution of the paper
This paper presents a transformer-based framework for myocardial infarction (MI) classification and cardiac substructure prediction. By employing a multi-task learning approach, the authors demonstrate that incorporating substructure prediction enhances MI classification accuracy. The latent features generated under this multi-task setting are shown to be explainable, and the proposed method outperforms existing state-of-the-art techniques.
- 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.
A key contribution is the demonstration that the auxiliary task of substructure prediction enhances the performance of the primary MI classification task, indicating effective feature enrichment through multi-task supervision. The substructure prediction module encourages the model to attend to localized spatial relationships between cardiac regions, thereby guiding the transformer encoder to learn more refined and discriminative representations. Furthermore, the authors show that the learned latent embeddings are correlated with clinical features, supporting the clinical feasibility of their approach.
- 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.
While the proposed method is well-executed, one limitation is that it builds closely upon reference [4], which also employed a multi-task architecture for MI prediction and anatomical modeling. Although the authors provide a brief distinction in their design—arguing that their approach alleviates the burden on substructure classification compared to [4] and has improved MI classification accuracy—it remains a little unclear how substantial the architectural and algorithmic differences truly are beyond replacing the encoder with a Vision Transformer (ViT). A more detailed discussion or empirical comparison with [4] would help clarify the novelty and justify the design choices.
- 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 submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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/2025/en/REVIEWER-GUIDELINES.html
Just a suggestion: while the model is clearly described in the paper, sharing the code—if feasible—would be a valuable addition that could further enhance reproducibility and benefit the community.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper is well-structured and presents a solid amount of work with impressive performance improvements. The proposed multi-task framework is thoughtfully designed, and the learned representations are shown to be clinically relevant and explainable. It would be even stronger if the authors could elaborate more clearly on the architectural differences compared to the reference [4], but overall this is a valuable contribution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [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
Reviewer #1 Thank you for your constructive feedback. Regarding the three questions you listed in the weakness section, we respectively make the following responses: 1) We have experimented with different sets of g (number of centroid points) and k (number of closet points). We discovered that g × k = 2N provides across the 3D heart, which we verified by visualizing all the visible patches. 2) This is a very good concern. We agree that label consistency across overlapping patches is important. However, to clarify, in our formulation, Overlapping patches lie on the same cardiac surfaces (i.e., substructures), and hence will be assigned the same label. For example, in a point set with 36,000 points, the first 12,000 points belong to one substructure and the next 12,000 to another, which means the process of substructure label assignment can be easily done. 3) We appreciate your suggestion regarding architectural details. We will incorporate all currently missing implementation details, including those you pointed out, in the final camera-ready version to improve reproducibility and clarity.
Reviewer #2 We really appreciate your recognition to the novelty of our work. We fully agree that incident or future MI event prediction can be formulated as a survival regression task. It would enable more nuanced risk stratification and help identify high- and low-risk populations. In this paper, our primary objective is to evaluate the effectiveness of our multi-task learning framework and its interpretability in the context of cardiac data. While survival regression is beyond the current scope, we see it as a promising direction for future work and appreciate you highlighting this important perspective.
Reviewer #3 We appreciate your comments and would like to clarify that our work differs from [4] in two key aspects: (1) the way the model processes the input 3D heart data, and (2) our explicit use of cardiac substructure labels in the learning process. These distinctions contribute significantly to the novelty of our method. We will try to include more explanation in the camera-ready file.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A