List of Papers Browse by Subject Areas Author List
Abstract
Major adverse cardiac events (MACE) are among the leading causes of death worldwide. Quantitative myocardium perfusion imaging (MPI) based on rubidium (Rb) PET allows measuring myocardial blood flow (MBF) in the left ventricle (LV), which proved to be predictive for MACE. Clinical routine relies on an approximate 3D LV centerline (CL) flattened to a 2D polar map (PM). PM-based artificial intelligences (AI) were proposed to predict MACE outperforming traditional non-image-based risk scores but discarding the true LV geometry and epi/endocardium perfusion. Using our fully automated 3D Rb PET processing pipeline, we developed tissue‑depth–specific graph neural networks (GNNs) operating in the initial 3D PET space for our 232 patients cohort, with 47 MACE events. We compared various models using a five-fold cross-validation and fold-wise bootstrapping confidence interval estimation. Because of the substantial class imbalance, evaluation focused on average precision (AP). As an initial baseline, we compared MBF intensity along the CL to train a logistic regression classifier. When extracting min, mean and max intensity on the CL, an AP of 45.9[44.0,48.0] is reached, which was outperformed by adding epi/endocardium values with AP of 54.3[52.0,56.5]. A CNN trained on the CL-based PM reached an AP of 50.7[48.6,52.9]. The proposed GNN leveraging multi-depth epicardium, CL and endocardium perfusion (MBF, uptake and activity) in the 3D PET space reached top AP of 60.3[58.3, 62.3], AUC 82.5[81.2, 83.8] and C-Index of 78.2[77.0, 79.4]. This approach could allow more accurate identification of patients at risk of MACE when compared to current clinical assessment based global LV CL MBF.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5371_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)
N/A
BibTex
@InProceedings{CheArt_Multidepth_MICCAI2026,
author = { Chevalley, Arthur AND Asiain, Maria M. AND Moulton, Eric AND Klein, Ran AND DeKemp, Robert AND Prior, John O. AND Jreige, Mario AND Kamani, Christel H. AND Depeursinge, Adrien},
title = { { Multi-depth Rubidium PET Graph Neural Network for Major Adverse Cardiac Event Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16887},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of the paper is to describe and evaluate a new graph-based method to improve the prediction of cardiovascular events from myocardium perfusion imaging.
- 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) Positive results: the results validate the authors’ assumptions that: (a) tracking the perfusion at multiple depths is better than at a single depth; (b) graph neural network is more fitted for the geometrical nature of the task.
2) Details for reproduction: the authors provide a lot of details about their method and experiments, which can be assumed to facilitate reproduction.
3) Possible clinical impact: although this study was retrospective, any improvement in cardiovascular risk prediction has a chance to benefit the patients.
- 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.
1) Readability. I find that the paper reads poorly, one of the main reasons being the sheer number of acronyms (10 in the abstract alone).
Second paragraph of the introduction is too long; authors should clarify in a concise fashion the limitations of current methods and explicitly state their hypotheses and strategy in another paragraph (in particular, the multi-depth aspect is not well explained). Also, the Introduction and Related work sections overlap a lot.
Results: a results section should summarize the data and indicate the global trends (which is actually done in the discussion section). Instead, the current results section appears as a caption of Table 1.
Discussion: in it’s current form (except the last paragraph) it is more a results section. All the key elements of a good discussion are lacking: main results stated in a concise fashion, replacement in the context of the litterature, explanation of results (why does your method works better?), limitations, clinical relevance of the work. The conclusion is fine.
2) Unclear methodology. I had a hard time understanding the second paragraph of section 3.2, which contains the core of the method. The last two sentences are particularly confusing. Maybe divide it into a list of steps? Also, figure 1, or a new figure (there is only one in all the paper) could better clarify how multiple depths are extracted. Also, could the authors describe the GNN input? What do the nodes and vertices contain exactly?
- Please rate the clarity and organization of this paper
Poor
- 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.
- 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 method is sound and the problematic important, but the current form of the paper does not match the required standard for publication.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
I am not satisfied with the rebuttal, as the authors dismissed my comments on the important readability issues of the paper. I must conclude that no improvements will be made in that regard.
Review #2
- Please describe the contribution of the paper
This paper proposes a novel framework for predicting major adverse cardiac events (MACE) from Rubidium-82 PET myocardial perfusion imaging (MPI). The authors introduce a fully automated pipeline to extract multi-depth (epicardial, centerline, endocardial) perfusion signals in 3D, and model them using graph neural networks (GNNs). The approach avoids the common polar map (PM) representation and preserves the left ventricle geometry. Experiments on a cohort of 232 patients (47 MACE events) demonstrate that the proposed multi-depth GNN outperforms feature-based models and CNNs trained on PMs, achieving the best performance
- 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.Originality and relevance
- The paper addresses an important clinical problem (MACE prediction) with strong translational potential.
- The idea of multi-depth perfusion modeling is well-motivated and novel to me in the PET MPI context.
- Using GNNs to preserve cardiac geometry instead of polar maps is a meaningful methodological contribution.
2.Technical contributions
- A fully automated 3D pipeline for LV segmentation and multi-depth sampling is well described and clinically relevant.
- The graph construction strategy is appropriate for irregular cardiac structures.
- Joint learning of classification and survival (DeepSurv-style loss) is a strong design choice.
3.Experimental design
- Comparison with multiple baselines (logistic regression, radiomics, CNNs) is thorough.
- Use of appropriate metrics for imbalance (AP, MCC) is commendable.
- Confidence intervals via bootstrapping strengthen the statistical validity.
4.Results
- Consistent improvements of multi-depth over centerline-only features support the main hypothesis.
- GNN clearly outperforms PM-based CNNs when multi-depth information is used.
- Survival analysis (C-index) adds additional clinical value.
- 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.
1.Limited dataset size
- The cohort (232 patients, 47 events) is relatively small for deep learning, especially GNNs.
- Risk of overfitting remains, despite cross-validation and regularization.
2.Reproducibility concerns
- Code is only “available upon request” and dataset is not publicly available.
- Some pipeline steps (e.g., LV segmentation, FlowQuant usage) rely on external tools or prior work with limited detail.
3.Impact of the clinical features
- Clinical features (CF) show an important impact (Table 1) but the way they are integrated in the model is not discussed in depth.
4.Comparison fairness
- CNN baseline is limited to polar maps, while their GNN relies on 3D segmentation. No 3D CNN baseline is included.
- Feature-based models do not use segment information, while GNN does—this may bias comparisons.
5.Clarity and writing
- Some sections are dense and difficult to follow (for example section 3.2 - Fully Automatic Multi-Depth MPI Processing Pipeline)
- Notation in Eq. (2) is heavy and could be simplified.
- 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 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.
(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 presents a novel and well-motivated approach combining multi-depth cardiac perfusion analysis with graph neural networks, addressing clear limitations of polar map representations. The methodological contribution is meaningful and supported by consistent experimental improvements.
- 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 paper presents a novel and clinically meaningful approach to MACE prediction based on automated transmural myocardial perfusion analysis. The use of a graph-based representation to explicitly encode transmural structure constitutes a key strength of the work. This design introduces an interpretable inductive bias that distinguishes the approach from standard black-box deep learning models and is well aligned with clinical reasoning.
The concerns raised during the review process have been satisfactorily addressed:
Dataset size and endpoint definition: The authors clarified the clinical rationale behind the use of composite MACE endpoints and acknowledged the limitations related to dataset size. This limitation will be clearly stated in the final manuscript.
Absence of 3D CNN baselines: While such baselines would provide additional context, the authors provided a clear and well-justified rationale for focusing on transmural perfusion modeling and interpretability. This is consistent with the core contribution of the paper.
Methodological clarity: The additional details provided in the rebuttal significantly improve understanding of the graph construction and modeling choices. The commitment to clarifying these aspects in the final version is appreciated.
Presentation and structure: The authors acknowledged readability concerns and provided a rationale for their organization. Minor improvements in clarity are expected in the final manuscript.
Reproducibility: While data sharing is constrained by legal considerations, the authors’ willingness to share code upon reasonable request is noted.
We encourage the authors to incorporate the clarifications and additional details from the rebuttal into the final version to further strengthen the manuscript.
Review #3
- Please describe the contribution of the paper
The paper proposes a fully automatic multi-depth pipeline for processing ⁸²Rb cardiac PET that replaces the conventional 2D polar map (PM) representation with a 3D point-cloud/graph representation preserving perfusion structure (epicardium, mid-myocardial centerline, endocardium). Three contributions are bundled: (i) an automated centerline-generation pipeline using nnU-Net segmentation followed by KNN-based separation of inner/outer borders to extract multiple depth-specific centerlines, removing the user/software dependence that currently contaminates clinical myocardial blood flow (MBF) quantification; (ii) a graph construction procedure over the sampled centerline point clouds, with minimum spanning tree simplification and RANSAC-based rest-stress registration to enable combined rest-stress feature learning despite deformation; and (iii) a TransformerConv-based GNN jointly trained for MACE classification and Cox survival prediction on ⁸²Rb SiPM PET/CT from 232 patients (47 MACE events), compared against feature-based logistic regression and a CNN baseline operating on conventional polar maps. The central empirical claim is that multi-depth perfusion information, accessible only through the 3D representation, carries MACE-predictive signal that is lost in standard polar-map pipelines.
- 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.
Polar maps do distort LV geometry, do collapse transmural information, and MBF quantification does depend meaningfully on centerline fitting and operator choice. This is a real, unsolved problem in cardiac PET, and proposing an automatic geometry-preserving alternative addresses it head-on. The pipeline design is thoughtful: combining nnU-Net segmentation with a convex-hull-based blood-pool extraction and recursive centerline generation at multiple depths is an elegant way to obtain transmural sampling without manual delineation. The graph construction choices are principled, particularly the minimum spanning tree simplification to mitigate over-smoothing (a real and underappreciated GNN failure mode) and the use of RANSAC registration to handle rest-stress deformation. The experimental design is rigorous by the standards of small-cohort cardiac imaging work: stratified five-fold cross-validation, Optuna for hyperparameter selection within matched splits to avoid leakage, double bootstrap for confidence intervals, and a carefully chosen metric set (AUC, AP, F1, MCC, C-index) that correctly acknowledges imbalance and combines classification with survival. Joint classification-plus-Cox training using the DeepSurv partial likelihood loss is the right way to handle the dual formulation. The baseline set (feature-based LR and radiomics at each depth, CNN on PMs) is well-chosen to isolate what the 3D graph representation actually contributes.
- 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.
Combining cardiac death, delayed revascularization (>6 months), MI, congestive heart failure, and de novo stable angina into a single binary outcome mixes hard endpoints with softer, more subjective ones. The distribution of event subtypes within the 47 cases is not reported, and whether the GNN predicts hard events equally well as soft ones is the clinically important question. A subgroup analysis on hard MACE only would substantially strengthen the clinical claim. Several methodological choices need justification. The downsampling of graphs by a factor of ten is stated without ablation. This is a substantial reduction, and whether predictive signal is preserved through downsampling should be shown. The MST simplification discards edges that may carry perfusion-pattern information (a transmural shortcut between epi and endo, for instance) and its effect on performance relative to the full k-NN graph should be reported. The RANSAC rest-stress registration is a reasonable substitute for direct MFR computation, but the paper does not quantify how well this registration actually aligns corresponding myocardial regions, nor whether misregistration systematically affects certain LV segments. The mapping of the 17-segment label into the graph as a node feature is used in the GNN but not in the feature-based models, which is a non-trivial asymmetry in what information each model sees. The claim that the approach removes user/software dependence in centerline fitting is made but not demonstrated. The natural validation is to compare the automatic pipeline’s MBF values against FlowQuant across multiple operators and show reduced variability, or to show that the automatic CL produces clinically concordant segment-level MBF. Without this, the deployability advantage is asserted rather than established. Reporting segmentation accuracy of the nnU-Net step and stability of the derived centerlines on test-retest scans, if available, would strengthen the pipeline contribution considerably. Other concerns: The time-to-event window (2 to 862 days) is very wide, and the Cox formulation should be accompanied by a Kaplan-Meier stratification. Minor presentation issues (“investing” for “investigating”, inconsistent formatting).
- 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 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 clinical problem is real and important, the proposed pipeline and representation are principled and well-engineered, the experimental protocol (cross-validation, bootstrap CIs, joint classification-survival) is rigorous, and the baselines are appropriate. The central ideas are genuine technical contributions that should enter the cardiac PET literature. What holds the paper back is the small event count (limiting statistical power for the headline claim), the absence of significance testing between the GNN and the PM-CNN baseline, the heterogeneous MACE endpoint with no subgroup analysis on hard events, the information asymmetry between the GNN and feature-based baselines via the segment label, and the lack of Kaplan-Meier/log-rank analysis for the survival claim. These are fixable within a revision cycle without requiring new data collection, and if addressed the paper would be solid.
- 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 would like to thank the reviewers and meta-reviewer for their careful evaluation of our manuscript. One criticism regarding this work concerns small dataset size combined with the MACE definition regrouping multiple endpoints of varying severity. While this grouping seems arbitrary, it is a standard design used in several important randomized clinical trials justified by clinical actionability, e.g. patients with higher predicted risks would be more closely monitored to limit the impact of any potential future event [PMID: 34742250]. Concerning sample size, we would like to highlight that this work is, to the best of our knowledge, the first to investigate automated transmural Rubidium perfusion analysis with a one-compartment model and reach the same conclusions as other studies showing the importance of transmural perfusion for cardiac health assessment [PMID: 36599572]. However, we will make sure to highlight the limited dataset size as a limitation in the final submission. Multiple reviewers questioned the absence of a 3D CNN model using the MBF volume as input. While we agree that it would provide an important baseline relying on deep learning, our objective is to specifically investigate the value of transmural perfusion for MACE prediction. The proposed graph approach allows us to ensure the use of transmural perfusion as an inductive bias. In addition, the resulting multi-depth perfusion graph can easily be inspected by clinicians contrary to black-box CNN models, which we will clarify. According to rebuttal guidelines we will not include this baseline but will consider it in future work. Other critics addressed the complexity of the methodology section and the limited details on design choices. We agree that this section has been stripped down of many details to keep the core component and fit the limited space available. Further in-depth description and performance analysis of the entire pipeline, i.e. MBF comparison with FlowQuant, segmentation model, is currently under publication, while we focus here on MACE prediction. Here are some details on the graph formulation that we could not include due to space constraints: -Graph nodes concatenate the perfusion (MBF/K1/activity) potentially at different depths and regional information. Edges connect nodes without features. -Minimum Spanning Tree (MST) does not cut the transmural connections due to the concatenation stated above. -Graph downsampling was not thoroughly investigated. However, we hypothesised that averaging nearby nodes only smooth the signals to avoid potential artefact and result in a comparable number of points as the standard-of-care FlowQuant -When used, clinical features are added using late fusion -The impact of registration on segmental alignment was not investigated as the overall patient-based registration was good. In addition, the volumes are similar as the rest and stress should be aligned up to small movement. -Segmental information was not used for the feature-based models as the segment territories are approximate in 3D which leads to ill-defined regions for feature extraction. However, 17-segment features were tested with non-satisfactory results. -LV centreline, computed from the whole LV, can be used to split the LV in epiLV and endoLV. Using the same algorithm on epiLV/endoLV results in the 75%/25% depth perfusion. One reviewer also pointed out readability issues and the overall structure of the paper, especially concerning results and discussion sections. We take the opportunity to clarify that we believe a results section should present metrics in an objective manner, without interpretation, so the reader can develop his own interpretations. Then, the discussion re-states the main hypothesis before interpreting the results, relation to the hypotheses and state of the art. Finally, concerning code/data sharing we are currently not able to share them due to legal restrictions but are open to share code upon reasonable request.
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 replacing conventional polar‑map representations with a fully automated 3D pipeline that extracts multi‑depth perfusion signals and represents them as graphs processed by a graph neural network, trained jointly for classification and survival analysis. But I am not convinced that the papers contribution is as strong as the authors’ claims. As major limitations, the cohort is relatively small (232 patients, 47 events), which limits confidence in model generalisation, particularly for deep and graph‑based architectures. The heterogeneous definition of MACE combines hard endpoints with softer clinical outcomes, with no stratified analysis is provided to clarify what types of events are actually being predicted. The absence of a strong 3D CNN baseline further weakens claims about the specific advantage of the graph representation.
Stronger justification of key design choices with clearer descriptions of methods and more careful framing of conclusions relative to dataset size and endpoint heterogeneity are needed to strengthen the contribution. Making code public available would improve reproducibility.
- 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.
The evalaution is limited. The contribution is solid but incremental combining exisiting methods for this problem. But the rebuttal clarifies design choices and acknowledges dataset size and endpoint heterogeneity as limitations. And two reviewers identify important clinical and methodological substance. R1 correctly flags that the methods section is dense and difficult to follow which should still be improved. Code should be released even if data are private.
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.
The paper addresses an important clinical problem in MACE prediction and proposes a meaningful representation by preserving 3D myocardial geometry and modeling multi-depth perfusion signals with GNNs, rather than relying on conventional polar maps. Although the cohort size and event count are limited, and additional statistical analysis would strengthen the claims, the method is clinically relevant, technically well motivated, and shows consistent improvements over appropriate baselines.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The paper addresses an important clinical problem and proposes an interesting multi-depth graph-based approach for PET perfusion analysis. However, the current evidence is not sufficient to support acceptance, mainly due to the small number of events, the heterogeneous MACE endpoint without subtype analysis, the lack of a strong 3D baseline, and limited validation of key pipeline choices. The rebuttal clarified some methodological details, but it did not resolve these main concerns.