List of Papers Browse by Subject Areas Author List
Abstract
Myocardial scar segmentation from single-stack late gadolinium–enhanced cardiac magnetic resonance (LGE-CMR) imaging has been a longstanding and clinically important challenge, particularly in the presence of low tissue contrast, diffuse, and small scar regions. These challenges are further intensified by the limited availability of 3D spatial context. This paper presents CalcSeg, a Confidence-aware latent context curriculum learning framework that leverages fused 3D feature representations from single-stack 2D LGE-CMR images for robust scar segmentation. Specifically, we introduce a dynamic semi-supervised curriculum learning strategy that progressively expands training from easier to more challenging scar cases using a learned confidence-aware scoring function. Such a function integrates errors in the predicted scar maps with quantified epistemic uncertainty and scar burden estimation to automatically assess sample difficulty without requiring manual labels. To compensate for the limited spatial context in single-stack acquisitions, we then develop a latent slice-wise self-attention to capture inter-slice dependencies and infer 3D spatial representations from sparse 2D inputs. We evaluate CalcSeg on multi-center clinical LGE-CMR datasets and benchmark against existing scar segmentation networks. Experimental results show that CalcSeg consistently outperforms all competing methods, particularly with substantial improvements on clinically challenging cases. Our code is released at https://github.com/vfb8zv/CalcSeg.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5989_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/vfb8zv/CalcSeg/
Link to the Dataset(s)
automatic Evaluation of Myocardial Infarction from Delayed-Enhancement Cardiac MRI (EMIDEC): https://emidec.com/dataset
STACOM challenge (human only): https://www.doc.ic.ac.uk/~rkarim/la_lv_framework/lv_infarct/index.html
UVA Health data: private in-house LGE data
BibTex
@InProceedings{JayNiv_CalcSeg_MICCAI2026,
author = { Jayakumar, Nivetha AND Kim, Hannah AND Patel, Amit R. AND Zhang, Miaomiao},
title = { { CalcSeg: Confidence-aware 3D Latent Context Curriculum Learning For Myocardial Scar Segmentation From Single-Stack LGE-CMRs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16884},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The authors present CaucSeg, a segmentation technique for scar in LGE CMR images. The approach includes an uncertainty-based curriculum learning (“UCL”) approach and a cross-slice attention mechanism. Results on a combination of private and public dataset suggest that the proposed technique outperforms other recent SOTA methods.
- 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 quite well written and overall easy to follow.
- Whilst the parts of the pipeline are not new, the specific combination proposed is interesting, suitable to the task at hand and sound.
- The results are believable and the improvement over SOTA seems substantial.
- 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.
- Limited validation. Despite having access to different datasets (some public, some private), the authors merged them when performing their validation. This reduces the chances of independent inter-paper comparisons, which are one of the limiting factors for research in LGE segmentation.
- Questionable experimental setting. The authors state “We quantify the segmentation performance using Dice Similarity Coefficient computed over the myocardium for the complete dataset and 6 known difficult cases, and the percentage error in scar volume. ” This is very problematic to me: what was the training/validation/test split? Was there one? The paper doesn’t mention this very crucial piece of information.
- Issues with uncertainty estimation. The paper claims that epistemic uncertainty “progressively decreases with the stages” of UCL, and uses this as evidence that the curriculum is working as intended. However, in Fig. 3 the y-axis is labelled “mean uncertainty”, and the mean across the whole test set (left panel) appears approximately stable across stages while only the std shrinks. This would indicate that UCL homogenises uncertainty across samples, not that it reduces it. The authors should (i) discuss this finding thoroughly and (ii) clarify exactly how the plotted “mean uncertainty” and its spread computed (per-sample MC-Dropout std aggregated how, over what?).
- Misuse of the term “unsupervised”. The paper repeatedly describes the proposed curriculum as “unsupervised”, but the difficulty score relies on the per-sample errors against the ground-truth segmentation. This requires labels and is therefore not unsupervised in the sense the term is normally used in the curriculum learning literature, where “unsupervised CL” refers to difficulty estimation without access to labels.
- 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
Page 6: “6 known difficult cases”. Known how? Aren’t 6 too few to draw meaningful conclusions and numerical estimates?
- 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 results seem somewhat solid, despite a rather limited (and possibly very questionable) validation. More information from the authors is needed during rebuttal to clarify possible flaws with their work.
- 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 authors’ provided some needed clarifications over the experimental set-up and data split that are most reassuring. While the work is still limited on the evaluation side, it now seems solid enough for publication at MICCAI.
Review #2
- Please describe the contribution of the paper
The paper addresses the problem of myocardial scar segmentation with CaucSeg, introducing an unsupervised curriculum learning strategy that leverages predictive uncertainty to estimate sample difficulty. Additionally, it proposes a 3D latent context modeling approach to improve consistency across 2D slices within a stack.
- 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.
- Myocardial scar segmentation is a challenging and important problem for the medical imaging community.
- The presented method shows improved performance against the baselines, both quantitatively and qualitatively. The idea of using curriculum learning for the myocardial scar segmentation is promising.
- 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.
- I find the paper to be written in a very confusing way. Particularly the section on “3D Latent Context Modeling” is very hard to understand, while being one of the contributions of the paper. It is unclear whether the latent representation z represents 2D or 3D features, it’s also not represented anywhere in Fig. 1 to aid understanding. Some variables are never defined, such as C in the softmax formula, B, and F. The “disentangled slice-wise self-attention” is never clearly described. The description of this module is dense and lacks clarity regarding the data flow, with the naming not exactly matching between text and figure. This makes the architectural novelty difficult to verify or reproduce. The “custom residual decoder” used is also never described. The loss function and curriculum scoring are introduced abruptly and are not explained with sufficient clarity. Several terms are defined mathematically, but the paper does not provide enough intuition for why these components are combined, how the weights were chosen, or how the resulting optimization behaves during training.
- The experimental results section provides very little analysis of the results. The ablation study is limited. It shows that each component helps, yet it does not disentangle whether the curriculum or the attention module is responsible for most of the gain on the hard cases, missing a combination where only UCL is used. Further, there is no discussion on why CaucSeg outperforms the baselines. The paper reports improvements over ScarNet, the myocardial scar segmentation SOTA method, but the experimental section does not sufficiently explain why the proposed method is better. In particular, it is unclear whether the gains come from the confidence-aware curriculum, the 3D latent context module, or simpler training effects.
- The empirical evaluation is incomplete with respect to recent literature. In particular, the paper does not compare against or cite a recent relevant method, ScarElastic [1], which appears closely related and has reported strong performance on myocardial scar segmentation. Additionally, the manuscript includes a baseline denoted as “ScarNet (Sup. CL),” but does not provide a citation or sufficient detail to clearly identify this method. Clarifying the identity of this baseline and including relevant recent methods such as [1] would strengthen the empirical evaluation.
- In Fig. 1 the images only show the myocardium with all the other tissues being masked. Is this a preprocessing step in the pipeline? No description of such masking is present in the dataset dataset section, which makes it confusing as to what is the actual input to the model.
[1] Zou, R., Li, Y., Zhou, N., Guan, X., Li, G., Xie, D., … & Luo, C. (2025). ScarElastic: continuous elasticity field modeling for myocardial scar delineation in LGE-CMR. npj Digital Medicine.
- 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 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the method is promising and tackles an important problem, the presentation requires revisions before it becomes a meaningful contribution to the MICCAI community. In the current state, the writing of the methodology section precludes reproducibility and makes it difficult for the reader to verify the claimed technical novelty. These issues are not minor wording problems and are unlikely to be resolved in a short rebuttal. For these reasons, I recommend rejection.
- 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.
While the paper addresses the important problem of myocardial scar segmentation, there are major issues that lead me to recommend a rejection. Namely, it is unclear what the exact input to the method is and how it is obtained. In the dataset section, the authors state, “All slices were cropped and masked to the left ventricle and resized to 224^2.” However, what is presented in Fig. 1 does not reflect a left ventricle mask; only the myocardium is shown on the images, while the left ventricular blood pool cavity has been completely zeroed out. How these highly precise myocardial masks are obtained remains unexplained, and the authors failed to clarify this in their rebuttal beyond making circular references back to the dataset section. It remains unknown whether clinicians must first manually delineate the myocardial area at test time, or if the masks are generated by an external automated method. If the inputs rely on a secondary segmentation method, the authors fail to address how often epicardial and endocardial scars are inadvertently excluded from the input images due to prior segmentation errors. Because these critical preprocessing details are absent from the manuscript, the pipeline is non-reproducible. Furthermore, such a restrictive masking step is not standard practice in prior works, where the concurrent segmentation of healthy myocardium, blood pool, and scar is typically part of the multi-class objective. Ultimately, it is unclear how isolating the input domain in this manner affects the performance of the baseline methods, which significantly downgrades the reliability of the presented comparative results.
Review #3
- Please describe the contribution of the paper
The paper introduces CaucSeg, a framework designed for myocardial scar segmentation from single-stack LGE-CMR images, which are notoriously difficult due to sparse slice sampling and low tissue contrast. The main contributions are twofold: (1) A 3D latent context module that utilizes slice-wise self-attention and sinusoidal positional embeddings to model longitudinal anatomical consistency across 2D slices. (2) An unsupervised, confidence-aware curriculum learning strategy that dynamically adjusts the training complexity. Unlike traditional curriculum learning that relies on predefined difficulty levels, CaucSeg uses a model-adaptive difficulty score combining predictive inaccuracy (Dice and scar burden) with epistemic uncertainty (via Monte Carlo Dropout). The method is validated on a large-scale, multi-center dataset of 976 patients, significantly outperforming existing CNN and Transformer-based baselines.
- 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.
he slice-wise self-attention module is an intelligent solution to the “anisotropy problem” in CMR. By performing latent fusion on 2D features rather than using full 3D convolutions, the authors effectively capture out-of-plane context without the computational overhead or sensitivity to breath-hold displacements common in 3D-UNet architectures. The shift from supervised to model-adaptive curriculum learning is a timely contribution. The integration of epistemic uncertainty into the difficulty score allows the model to prioritize samples where the “ground truth” might be ambiguous (e.g., the gray zone of fibrosis), which is a common clinical reality in LGE-CMR.
- 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.
Terminology Ambiguity: The authors refer to the curriculum as “unsupervised.” While the difficulty scoring is automated, the segmentation task is supervised. This distinction should be clearer to avoid confusion with self-supervised or unsupervised segmentation methods. Parameter Sensitivity: The difficulty score D_i depends on three weighting parameters \gamma_j. The manuscript lacks a sensitivity analysis or a clear heuristic for selecting these weights. Given that curriculum learning can be sensitive to the order of sample introduction, this is a significant omission.
- 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.
(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 manuscript provides a robust and technically sound solution to myocardial scar segmentation, one of the most challenging tasks in cardiac AI. The dual innovation of 3D latent context for sparse stacks and the uncertainty-guided curriculum is highly relevant to the MICCAI community. The scale of the validation (976 patients) is exceptional for a conference submission and provides high confidence in the results. While there are minor concerns regarding terminology and parameter sensitivity, the overall performance gains, particularly on clinically challenging, low-contrast cases, demonstrate a clear advancement over established SOTA models like ScarNet and UNETR.
- 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.
N/A
Author Feedback
We thank the reviewers for all valuable feedback and constructive suggestions. Our responses to the major questions are summarized below.
[M,R1,R3] Clarification on “Unsupervised” Curriculum Learning (UCL) Our intent was to distinguish our approach from manually annotated difficulty ratings, i.e., scores are computed automatically using Dice, uncertainty and scar burden error rather than clinician-assigned labels. Although uncertainty provides an unsupervised signal, we agree that it may be confused with typical unsupervised training methods that do not use ground-truth label supervision. We will update the terminology with semi/weakly-supervised.
[M,R1] Data Splits & Evaluation Protocol All metrics are computed on the independent test set, including 6 clinically challenging cases from a site, explicitly flagged by experts. The train-val-test split was 7:1:2. The uncertainty is computed as the mean standard deviation across 100 MC-Dropout samples per subject, yielding one scalar per subject. Fig.3 shows mean and standard deviation of these uncertainty values. The overall mean (left, Fig.3) appears relatively stable since high-confidence cases dominate the test set (ratio 22:6). In contrast, when stratified by difficulty (right, Fig.3), uncertainty decreases substantially across stages for the clinically challenging subgroup. We will clarify this in the revised paper.
[M,R2] Please note that z is a 2D per-slice latent representation produced independently for each slice, with ‘F’ total slices, yielding a 3D tensor prior to self-attention (SA). The disentangled slice-wise SA is a standard transformer operation [7,22] applied across slices “performing explicit 2D-to-3D latent feature fusion on the encoded latent representation z.” The term “disentangled” describes SA alternating between spatial and slice-wise attention, following the space-time attention in Arnab et al., 2021. All notations (B=batch size, F=total slices, C=projection dimension), and the “custom residual decoder” referring to an employed ResNet-based decoder instead of the standard MedSAM decoder will be clarified.
[R1] We appreciate the suggestion to keep validation separated between public and in-house datasets wherever dataset sizes allow. However, current public LGE datasets are relatively small, particularly for difficult cases, motivating us to incorporate private institutes’ data. Fully separated validation remains valuable for future work as larger public datasets become available.
[R2] The improved performance of our method over the baselines stems from combining the two proposed modules: slice-wise SA & UCL. Tab.1 includes results from ScarNet with CL, which uses a 2D transformer backbone and is analogous to the UCL-only ablation. Tab.2 shows that our SA improves overall segmentation performance, UCL particularly improves difficult cases, and their combination yields the best overall performance. These findings suggest SOTA methods underperform since they lack richer 3D context, operating on isolated 2D slices and random batch selection.
Thanks for referring to ScarElastic. Our framework could be easily incorporated on top of ScarElastic by using it as an alternative backbone. We will include this reference in our revised manuscript, and will clarify that “ScarNet (Sup.CL)” denotes ScarNet[30] trained with CL.
Masking is a preprocessing step applied prior to training and inference, as stated in the dataset section. The loss function and curriculum scoring are independent from the architectural design, as described in the method section. The focal loss parameters followed [Lin et al., 2018] with related values provided in the parameter setting section.
[R3] Based on our empirical findings, the parameters are stable as they are automatically derived from each metric’s scale via the inverse of its order of magnitude. This ensures all metrics are in the same numerical range and no single metric dominates. We will clarify this in the revised paper.
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 a myocardial scar segmentation framework that combines confidence‑aware curriculum learning with slice‑wise latent self‑attention to better handle sparse stacks, low contrast, and highly heterogeneous scar patterns. While all reviewers recognise the importance of the results and the interesting combination of uncertainty‑guided training progression and cross‑slice context modelling, in a rebuttal, it would be important for the authors to (i) clearly specify the data splits and evaluation protocol, (ii) correct or clarify the terminology around “unsupervised” curriculum learning, (iii) provide a more careful interpretation of the uncertainty experiments, and (iv) substantially improve the clarity of the method description, especially for the slice‑wise attention and decoder design. Overall, the paper tackles an important problem with promising ideas and encouraging results, but clearer presentation and experimental clarification are needed to fully establish its contribution.
- 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 with an interesting combination of uncertainty-guided curriculum learning and slice-wise latent self-attention. The rebuttal successfully clarifies several presentation issues and also explains the uncertainty analysis more carefully, and therefore, it can be accepted. R2 maintains it is unclear whether manual myocardial delineation is required at test time and whether this masking disadvantages multi-class baselines. I would really like this answered.
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 paper proposes CaucSeg, a new framework for myocardial scar segmentation from single-stack late gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) images. The core innovation lies in introducing a “confidence-aware unsupervised curriculum learning” strategy combined with a slice-wise self-attention-based 3D latent context modeling module. The main concerns include clarification for experiment settings , results analysis, interpretation of uncertainty results, and performance comparison. The author’s response address most of them. The input of the method needs further clarification. Figure 1 shows the input of the SAX images with only the myocardium region kept, which is not consistent with the descriptions in the text.
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 reports promising results, and the rebuttal clarified several issues including the data split, uncertainty analysis, and terminology around the curriculum. However, a central concern remains unresolved: the manuscript does not clearly explain how the LV/myocardial mask used as preprocessing is obtained at inference time, nor how this affects reproducibility and fairness of the baseline comparisons. This ambiguity affects the exact task being solved and the clinical interpretation of the results.
