List of Papers Browse by Subject Areas Author List
Abstract
Myocardial scar detection is a clinically important task with implications for prognosis, arrhythmic risk, and therapeutic decision making. The noninvasive reference standard for scar identification is cardiac magnetic resonance (CMR) with late gadolinium enhancement, but its use is constrained by cost, limited availability, and logistical barriers. In this study, we develop a CMR-supervised multi-stage automated framework to predict segment-level myocardial scar from routine echocardiography. Using an in-house dataset at a large medical center, we analyzed a dataset of echocardiogram scans from 1,921 patients with paired CMR data within one year. Myocardial regions were segmented from apical two-, three-, and four-chamber views, covering the full 17-segment AHA model, using view-specific nnU-Net architectures trained on manual annotations across the cardiac cycle. We then compared two feature representations for per-segment scar detection: (1) temporal radiomic features extracted from segmented myocardium and (2) features derived from the encoder bottleneck of trained nnU-Net models. Using class-weighted random forest classifiers, encoder features achieved a mean AUC of 0.822 across 18 view–segment combinations, significantly outperforming radiomics (mean AUC 0.793; p < 0.05 in 8 of 18 segments). Combining both feature sets yielded the highest AUC (reaching up to 0.858), although this improvement did not generalize consistently across views. These findings demonstrate that routine echocardiography, enhanced through CMR-supervised feature learning, can serve as a scalable and effective tool for myocardial scar screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4373_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{SheMan_EchoSCAR_MICCAI2026,
author = { Shehzad, Mansoor AND Yu, Arion AND Nayaz, Eman AND Alkhaleefah, Mohammad AND Al-Kindi, Sadeer AND Balakrishnan, Guha AND Sabharwal, Ashutosh},
title = { { Echo-SCAR: Temporal Feature Learning for Myocardial Scar Detection from Routine Echocardiography } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The implementation and validation of two methods to identify scar segments from echocardiography scans, validated by comparison with LGE CMR on a large dataset.
- 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.
Up to my knowledge this is the first demonstration of the ability of machine learning methods applied to echocardiography to reproduce LGE results in a dataset of this size. The methodology is sound and the validation is well performed. The results, showing an advantage of nnuNet embeddings over radiomics, and of both over clinical baselines, are interesting. The paper is well written and clear.
- 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.
The methods used are well established; the novelty relies on the application. It could be argued that the methods are not optimal for the task either. While the radiomics methodology is appropriate, the nnUNet is not an optimal way to obtain embeddings for capturing image content. Other approaches are standard in the literature and could have been used here. While the paper includes, to some extent, temporal characteristics, these are also included using simple methods (values at ED/ES, root mean square, spectral entropy), which are likely to miss interesting temporal features. There are several methods in the literature to encode spatio-temporal representations that could have been used here. I am unsure of the relevance of reference [12].
- 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.
- 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 size of the study and the conclusions are of interest to the community. The paper is well written and the introduction and discussion are strong. However, the methods are behind state of the art, which detracts from their relevance to future research
- 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 #2
- Please describe the contribution of the paper
This paper demonstrates that echocardiographic images hold enough information (features) about presence of myocardial scar (not visible to the human eye) that can be learnt using machine learning techniques to achieve reasonably good performance for segment level classification (scar vs no scar) compared to the current gold standard which is cardiac MR.
Clinically, this is a very important question to address. Detection of myocardial scars has significant implications in several settings like post myocardial infarction, heart failure, dilated cardiomyopathy, post congenital heart disease surgery etc. Though CMR is the gold standard, it is not a scalable solution due to cost and availability. However, echocardiography (especially 2D) is widely available and cost effective. But detecting scars on echocardiography is not a straightforward task even for human experts. Therefore, techniques to detect scars from echocardiography are highly valuable from a clinical application perspective.
Technically, this paper demonstrates that features learnt by a deep learning model (nnU-Net in this study) when being trained to segment the left ventricular myocardium encode enough information to detect scars reliably eventhough the model received no classification specific (scar specific) training.
- 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.
Clinically, the major contribution is using echocardiographic images to detect myocardial scars. Most work in myocardial scar detection are focused on CMR. The universal availability and lower cost of echocardiography makes it a scalable application.
Technically, usage of features extracted from the encoder bottleneck layer of nnU-Net models trained for segmentation tasks directly as input for downstream random forest classifier without specific training for the classification task is an important contribution. The core hypothesis that a U-Net based model learns rich representations of myocardium when learning it’s boundaries for segmenation task proved valid looking at the better classification performance compared the radiomics derived featureset which is usually the go to method for texture analysis.
- 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.
Clinical: One year gap between echocardiography and CMR could introduce scar label mismatch due to various reasons (listed below in additional details section). Lack of enough details about CMR and echocardiography data acquistion and curation.
Technical: Potential data leak: In section 4.1, authors mention 5-fold CV applied to 1921 patients which is the entire dataset. Hyperparameter tuning is to be done on training data (train plus val) only. Test set needs to be separated out before any model training or tuning. This needs to be addressed. Bull’s eye plot in figure 2 shows AUC of 0.86 for apex (segment 17). But the authors mentioned in section 2 that this segment was excluded “because reliable feature extraction from this region is compromised by near-field clutter artifact and apical foreshortening inherent to transthoracic imaging”. This needs to be addressed.
The paper, after addressing concerns above, is adequate for a conference presentation.
- 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.
- 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
Most of the following concerns are from the perspective of a full paper down the line unless specified above in the major concerns section.
Concerns from a clinical perspective: The authors mention upto one year gap between echocardiography and CMR acquisition. This could potentially lead to different states of the myocardium in atleast some of the patients. For eg. if a patient was admitted for myocardial infarction during which echocardiography was conducted and undergoes a CMR a few months later because the treating physician ordered CMR for heart failure symptoms or a ventricular arrhythmia (which is a common scenario), it is very likely that the CMR captures the scar but the echocardiography was done before a mature scar could be formed. Similarly, a CMR may have been done before echocardiography by which time advanced remodeling could have changed the myocardial texture. Several such situations could be listed. From a full paper perspective, the authors could provide more details about the timing between echocardiography and CMR (median time interval and IQR) and more importantly if there were different indications for echocardiography or any new cardiac event between the two image acquisitions. A clinical adjudication committee to decide if a pair of echocardiography and CMR are appropriate for inclusion in the study could be considered to ensure correct pairing. As CMR is the gold standard used for providing labels for the classification model, a more detailed description of the process from CMR acquisition to obtaining the segment level labels for scar is desirable. Specifically, the following could be reported: whether both transmural and non-transmural scars considered; cut off, if any, used for proportion of segment with scar tissue to be considered as positive label for scar presence; compatibility between the segmentation schemes used for CMR and echocardiography; number of radiologists involved in reporting the CMR; methods used to resolve ties, if any; and use of quality assurance layer. More details on echocardiography are also desirable. The paper reports use of view classification model to classify views into A4C, A3C and A2C. The authors could mention the reason for this choice over human classification (Was it to automate the pipeline (adds an additional deep learning model to the mix) or because details about the views were not available?). Duration of each echocardiographic recording preferably reported as number of cardiac cycles (important for temporal summarization), method of ascertainment for end systolic and end diastolic frames, echocardiographic settings used for image acquisition, tool/software used for marking the 17 segments, details of manual annotation could be provided.
Concerns from a technical perspective: Major concerns (data leak and bull’s eye plot in figure 2) are mentioned above. The following are concerns from a full paper perspective. To ensure reproducibility of the work, the authors need to include more details about the technical implementation. The current paper lacks sufficient detail. For eg. A diagram for end to end pipleline starting from view classification to final scar classficiation could be included. The current figure 1 provides a brief overview only and that too starting from segmentation task. More details about about data like format (dicom or jpg/png), train test split, stratification, normalization, augmentation etc needs to be included. Details about model architecture, model size, training protocol, tuning parameters used, compute resources, time taken for inference using final model etc could be included. Explainablity techniques like SHAP for radiomics could be included. Finally, addition of an external validation set (from a different hospital, different echo/CMR vendor) would add great weight to validity.
Suggestions from presentation perspective: The authors could present the segment level and view level AUCs for various aproaches using plots.
In future work, the authors could consider developing and evaluating an end-to-end deep learning model for scar classification model. Clinically, it would be interesting to see the performance of this model in clinical workflows.
- 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?
Overall study design is strong.
From clinical perspective, usage of CMR derived labels for classification model training, large dataset (though single centre), segment and view level scar detection, usage of three echocardiographic views to capture all relevant segments of left ventricular wall and temporal summarization of segment level information contributed to the strength.
From a technical perspective, chaining multiple deep learning models (view classifier, segmentation model and it’s encoder bottle neck layer) with a classical machine learning model for classification at the end to demonstrate reasonably good scar detection on noisy echocardiographic images is a commendable achievement.
- 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
A coherent spatio-temporal encoder of echo-derived biomarkers segmented into the AHA model applied to a large private cohort of 1,921 patients validated with LGE CMR scar hyperenhancement.
- 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 major strength of the paper is in its application to a private large dataset, coupled by a clever combination of established ‘state-of-the-art’ tools (temporal encoders, image radiomics).
- It provides ‘strong’ evidence that echo is a viable imaging modality to ‘more precisely’ refer to LGE CMR.
- While often the Fusion combination is better, the difference is not significantly better. Why do you think that is? Is there an anatomical or signal acquisition (noise) reason for this?
- Methods are easy to follow, uncomplicated, rational, and explainable from a clinical / physiological point of view.
- 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.
- Please define LGE earlier.
- Wording, drop second ‘a dataset’ from ‘Using an in-house dataset at a large medical center…’ (Abstract)
- Did anything change clinically during that year you gave yourselves for the LGE referral?
- What do you mean by ‘approximately 412 radiomic’? (just before Section 4)
- In terms of clinical workflow, are there instances where these echo-informed referrals would provide less value because the MRI referral would happen anyway?
- 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.
- 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
- It would be great to see in future work, how you could take this to the next level with RWMA scores. [e.g., https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=regional+wall+motion+abnormalities&oq=regional+wall+]
- 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 is a clinically relevant, well-validated study, using well-established components and applied to a great dataset. In my opinion, this paper strikes an excellent balance of all things ‘MICCAI’, i.e., enough methodological novelty with a great clinical application, validation methods, elegant simplicity.
- 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 the reviewers for their constructive feedback and careful reading. We group the major concerns below.
1.Cross-validation is nested, not leaky (R2). We apologize for the abbreviated wording in Sec. 4.1.The procedure is in fact nested patient-stratified CV: an outer 5-fold StratifiedGroupKFold for performance estimation, with an inner 3-fold StratifiedGroupKFold inside GridSearchCV for hyperparameter selection within each outer training fold. Patient grouping by MRN keeps all videos from a patient in the same fold, eliminating train/test leakage.
2.Choice of radiomics and nnU-Net encoder (R1). Thank you. Segment-level scar prediction from echocardiography is, to our knowledge, a novel direction, and the methods were chosen to match the localized nature of the task: scarring is a localized process, radiomics is a localized feature extractor, and segmentation is a localizing task. The reported performance was achieved without scar-specific supervision, which we view as one of the more interesting observations of the study. Alternative encoders (DINOv2/MAE with echo fine-tuning, video transformers) are reasonable next steps and will be flagged. The temporal summaries capture tissue state at key moments of the cardiac cycle (ES and ED at peak-contraction and peak-relaxation) and the temporal dynamics across the full cycle (RMS for signal magnitude, spectral entropy for variation).
3.Echo to CMR timing (R2, R3). Thank you, this is an important point, and we appreciate R2’s example of the post-MI patient whose echo precedes mature scar formation. The pairing in our institution follows the typical clinical sequence: echo as the first-line study, then CMR when findings warrant tissue characterization, with the closest pair per patient selected. We accepted pairs within the window described in Section 2 because scar positives at any specific AHA segment are lower prevalence, and tighter windowing would have left too few positives for stable per-segment estimates. Some positive labels likely correspond to scar that matured between studies, which we will note in the limitations.
4.Segment 17 in bullseye (R2). Correct, and thank you for the catch. Segment 17 is excluded from the per-segment analysis and should not appear colored; the figure will be corrected.
5.CMR label provenance (R2). Labels were derived from clinical CMR-LGE reads at a large academic medical center as part of routine clinical reporting, with hyperenhancement scored per AHA segment on the 0 to 5 ordinal transmural extent scale. For this work we binarized scar-present (≥1) versus absent (0); ordinal transmurality and external validation are part of future work. AHA conventions align between CMR and apical echo views.
6.Minor items (R3). Thank you for the careful reading. LGE will be defined on first mention, the duplicated “a dataset” will be corrected, and “approximately 412 radiomic features per segment” will be clarified as 103 raw PyRadiomics features times 4 temporal summaries. On workflow value: while patients with overt indications are referred regardless, many echocardiograms are performed for non-specific indications (dyspnea, atypical chest pain, screening), where the framework may help surface patients who fall through current pathways. Reference [12] (R1) is ComBat (originally microarray); we will cite an imaging-specific adaptation.
7.Fusion view-dependence (R3). Thank you. Radiomic and encoder features carry partially overlapping and partially complementary information about the same region. Properties of the acquisition view likely contribute to the observed view-dependence, and further investigation remains to disentangle the underlying factors. Extending toward RWMA scores is a natural follow-on we plan to pursue.
These edits will be made in the final manuscript. We thank the reviewers for their time and consideration
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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.
The manuscript presents a clinically important and well-executed application study showing that routine echocardiography can predict CMR-defined myocardial scar at segment and view level. The reviewers agree that the dataset size, clinical relevance, clear presentation, and comparison of radiomics with segmentation-derived encoder features make the contribution valuable, despite the use of established methods. The remaining minor concerns are mainly clarifications about the validation protocol, echo–CMR time interval, CMR label generation, and the segment-17 figure inconsistency.