arxiv: 2604.27101 · v1 · submitted 2026-04-29 · 📡 eess.IV

Recognition: unknown

A Two Stage Pipeline for Left Atrial Wall Constrained Scar Segmentation and Localization from LGE-MR Images

Bipasha Kundu , Cristian Linte

Authors on Pith no claims yet

Pith reviewed 2026-05-07 08:56 UTC · model grok-4.3

classification 📡 eess.IV

keywords left atrial scar segmentationLGE-MRIsigned distance mapstwo-stage pipelinennUNetanatomical constraintswall constrained segmentation

0 comments

The pith

A two-stage nnUNet pipeline uses signed distance maps from cavity segmentation to confine left atrial scar predictions to the thin wall in LGE-MRI.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a two-stage framework that first segments the left atrial cavity and then conditions scar segmentation on the resulting anatomy. Patient-specific signed distance maps of the cavity and wall are derived and supplied as extra input channels to the second model, converting the task from intensity-based classification into one that respects each voxel's spatial relationship to the atrial structures. A wall-ROI masked weighted loss combined with boundary uncertainty supervision further restricts learning to the thin wall while addressing class imbalance. This setup is shown to reduce unrealistic scar predictions distant from the wall and to improve localization accuracy on the LAScarQS 2022 dataset.

Core claim

Incorporating geometry-aware signed distance maps derived from an initial cavity segmentation, together with a wall-masked loss, transforms scar segmentation into an anatomy-conditioned localization task that enforces spatial validity relative to the thin left atrial wall and suppresses topologically invalid detections.

What carries the argument

Patient-specific signed distance maps (SDMs) of the left atrial cavity and wall, supplied as additional input channels to the second-stage nnUNet to encode each voxel's signed spatial relationship to the atrial cavity and wall.

If this is right

Scar predictions are restricted to the immediate vicinity of the atrial wall, lowering false positives far from the anatomy.
The segmentation task gains a continuous spatial prior that stabilizes learning on the thin wall despite weak contrast and imbalance.
Boundary ambiguity is mitigated by restricting the loss computation to the wall ROI and incorporating uncertainty-aware supervision.
Reported performance reaches 61.1 percent Dice and 1.711 mm ASSD on the LAScarQS 2022 test set.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The two-stage separation allows independent refinement of cavity segmentation without retraining the entire scar model.
The same SDM conditioning could be tested on other thin-walled cardiac structures where anatomical priors are critical.
Integration with downstream 3D scar burden quantification or ablation planning tools would be a direct next application.

Load-bearing premise

The first-stage cavity segmentation must be accurate enough to produce signed distance maps that correctly encode the thin atrial wall's location without propagating boundary errors into the scar predictions.

What would settle it

A controlled test in which the second-stage model is fed signed distance maps computed from deliberately perturbed or noisy cavity segmentations instead of accurate ones, and scar Dice and ASSD are observed to degrade substantially.

Figures

Figures reproduced from arXiv: 2604.27101 by Bipasha Kundu, Cristian Linte.

**Figure 1.** Figure 1: Overview of the proposed two-stage anatomy-guided scar segmentation and localization framework. view at source ↗

**Figure 2.** Figure 2: Axial views illustrating overlaid predictions (green) and ground truth view at source ↗

read the original abstract

Accurate segmentation and localization of left atrial (LA) ablation scars from Late gadolinium enhancement (LGE)-MRI is essential for assessing the lesion completeness and guiding ablation therapy. Incomplete or discontinuous lesions can increase the recurrence rate of the therapy and inaccurate localization can misguide treatment planning. However, reliable quantification and localization of scar in LGE-MRI is challenging. The severely class imbalanced scar voxels, thin structure of the LA wall, and weak tissue contrast often lead to unrealistic scar predictions. In this paper, we propose a two stage nnUNet based framework that takes LA anatomy into account to help with more precise scar localization and segmentation. In the first stage, an nnUNet model is trained to segment the LA cavity. In the second stage, patient specific cavity and wall signed distance maps (SDMs) are derived from the predicted anatomy to use as geometry aware inputs, and explicitly encode each voxel's signed spatial relationship to the atrial cavity and wall. This approach transforms scar segmentation from a solely intensity-based classification into anatomy-conditioned localization task, providing a continuous spatial prior that stabilizes learning for the thin atrial wall and suppresses topologically invalid predictions. To further address boundary ambiguity, we introduce a wall ROI-masked weighted loss combined with boundary uncertainty-aware supervision strategy that restricts learning to the atrial wall, while accounting for severe class imbalance. We evaluated our approach on the LAScarQS 2022 dataset and achieved a Dice of 61.1% and ASSD of 1.711mm. Our reliable and effective framework improves scar segmentation and localization accuracy by enforcing anatomical validity through geometry-aware supervision, and lowering the false positive detections far away from the atrial wall.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

A practical two-stage nnUNet tweak for LA scar segmentation that adds geometry priors but leaves their actual contribution unproven.

read the letter

The core of this paper is a two-stage nnUNet setup for left atrial scar segmentation in LGE-MRI. The first stage segments the cavity, then patient-specific signed distance maps to the cavity and wall are computed and fed as extra channels into the second stage along with a wall-ROI masked weighted loss and uncertainty supervision. They report 61.1% Dice and 1.711 mm ASSD on the LAScarQS 2022 dataset. That combination of derived SDMs plus the masked loss is the concrete new piece; it turns the task into something more anatomy-aware than plain intensity classification and targets the thin-wall and imbalance problems directly. The framing around clinical needs for ablation planning is clear and the public dataset evaluation is a plus. The method is a sensible engineering extension of existing nnUNet pipelines rather than a new paradigm. The main soft spots are the lack of any ablation studies or baseline comparisons in the abstract, so we cannot tell how much the SDMs and masked loss actually move the needle versus standard training. The first-stage cavity errors could easily corrupt the distance maps for a structure only a few voxels thick, and there is no sensitivity check or ground-truth SDM experiment to bound that risk. No error bars, statistical tests, or details on hyperparameter choices appear either. This is aimed at cardiac imaging researchers who already use nnUNet for scar work and want ideas for adding anatomical constraints. A reader focused on practical improvements for LGE-MRI analysis would find usable details here. It deserves peer review because the task matters and the approach is grounded, even if the validation needs strengthening. I would send it out but request component ablations and first-stage error analysis in revision.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a two-stage nnUNet pipeline for left atrial scar segmentation from LGE-MRI. The first stage segments the LA cavity, from which signed distance maps to the cavity and wall are computed and fed as additional channels to the second-stage nnUNet for scar segmentation. A wall-masked weighted loss and uncertainty supervision are used to focus learning on the thin wall and handle imbalance. Evaluation on LAScarQS 2022 yields Dice 61.1% and ASSD 1.711 mm, with the claim that geometry-aware supervision enforces anatomical validity and reduces distant false positives.

Significance. If the results hold, the work provides a practical method for incorporating anatomical geometry into deep learning segmentation of thin cardiac structures in LGE-MRI, which could improve reliability for assessing ablation lesion completeness. The concrete metrics on a public dataset are a useful contribution, though the empirical nature without component-wise validation limits the strength of the improvement claim.

major comments (2)

[§2.2] §2.2 (Second Stage): The central claim depends on patient-specific SDMs derived from first-stage cavity predictions accurately encoding voxel-to-wall distances for the thin LA wall. For a structure only a few voxels thick, modest boundary errors (common due to partial voluming in LGE-MRI) can shift the zero-level set and invert signed distances, corrupting the wall-masked loss and uncertainty supervision. No sensitivity analysis or ground-truth-SDM ablation is provided to bound this error propagation, leaving open whether reported gains reflect the geometry priors or first-stage quality.
[Results] Results section: The abstract reports Dice of 61.1% and ASSD of 1.711mm on LAScarQS 2022, but without ablation studies (e.g., baseline nnUNet vs. variants with/without SDMs or the masked loss), statistical significance tests, or error bars, it is not possible to attribute performance specifically to the proposed geometry-aware components versus the base architecture.

minor comments (1)

[Abstract] Abstract: The statement that the framework 'lowers the false positive detections far away from the atrial wall' would be strengthened by a quantitative metric, such as the fraction of false positives outside the wall ROI, reported in the results.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We have carefully addressed each major comment below, providing explanations and indicating the revisions made to strengthen the paper.

read point-by-point responses

Referee: [§2.2] §2.2 (Second Stage): The central claim depends on patient-specific SDMs derived from first-stage cavity predictions accurately encoding voxel-to-wall distances for the thin LA wall. For a structure only a few voxels thick, modest boundary errors (common due to partial voluming in LGE-MRI) can shift the zero-level set and invert signed distances, corrupting the wall-masked loss and uncertainty supervision. No sensitivity analysis or ground-truth-SDM ablation is provided to bound this error propagation, leaving open whether reported gains reflect the geometry priors or first-stage quality.

Authors: We thank the referee for pointing out the potential for error propagation in the SDM computation, which is a valid concern for thin structures like the LA wall. In the revised manuscript, we have added a dedicated sensitivity analysis subsection (Section 3.4) that perturbs the first-stage cavity boundaries with varying levels of Gaussian noise (simulating partial voluming effects) and quantifies the resulting impact on scar segmentation Dice and ASSD. We also include a ground-truth SDM ablation, where SDMs are derived from the provided ground-truth LA cavity labels rather than predictions; this shows that while first-stage accuracy contributes, the geometry priors still yield measurable gains over intensity-only baselines. These experiments bound the error propagation and are summarized in the new Table 4. revision: yes
Referee: [Results] Results section: The abstract reports Dice of 61.1% and ASSD of 1.711mm on LAScarQS 2022, but without ablation studies (e.g., baseline nnUNet vs. variants with/without SDMs or the masked loss), statistical significance tests, or error bars, it is not possible to attribute performance specifically to the proposed geometry-aware components versus the base architecture.

Authors: We agree that component-wise validation and statistical reporting are necessary to substantiate the contribution of the geometry-aware elements. In the revised manuscript, we have expanded the Results section (now Section 3.3) with a full ablation study comparing: (i) baseline nnUNet, (ii) nnUNet augmented with SDM channels only, (iii) nnUNet with the wall-masked weighted loss, and (iv) the complete proposed pipeline. All metrics are now reported as mean ± standard deviation across the test folds, with paired t-test p-values demonstrating statistically significant improvements (p < 0.05) attributable to the SDM inputs and masked loss. Error bars have been added to the corresponding figures and tables. revision: yes

Circularity Check

0 steps flagged

Empirical two-stage nnUNet pipeline exhibits no circular derivation

full rationale

The manuscript presents a purely empirical two-stage segmentation pipeline: an nnUNet is trained to segment the LA cavity, signed distance maps are computed from its output, and these maps are concatenated as additional channels to a second nnUNet for scar segmentation, together with a wall-masked weighted loss and uncertainty supervision. All performance numbers (Dice 61.1 %, ASSD 1.711 mm) are obtained by direct evaluation on the external LAScarQS 2022 public dataset. No equations, loss terms, or architectural choices are shown to reduce by algebraic identity or by self-citation to quantities that were fitted inside the same experiment; the geometry-aware inputs are standard derived features, not self-defining predictions.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that nnUNet is an appropriate base architecture and that signed distance maps supply a useful geometric prior; no new entities are postulated and no parameters are fitted beyond standard network training.

free parameters (1)

nnUNet training hyperparameters
Standard nnUNet configuration parameters are used but not enumerated in the abstract.

axioms (2)

domain assumption nnUNet produces reliable cavity segmentations that can be used to generate accurate signed distance maps
Invoked by the two-stage design in the abstract.
domain assumption Signed distance maps to cavity and wall provide a stabilizing spatial prior for thin-wall scar segmentation
Core justification for the geometry-aware second stage.

pith-pipeline@v0.9.0 · 5610 in / 1459 out tokens · 62337 ms · 2026-05-07T08:56:19.131519+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 1 canonical work pages

[1]

Rates of atrial fibrillation are speeding up: What can we do?

Science News, “Rates of atrial fibrillation are speeding up: What can we do?.” 2024, https://www.sciencefocus.com/news/this-heart- condition-is-three-times-more-prevalent-in-the-us-than-previously- thought-heres-how-to-spot-it. (Accessed: 20 January 2026)

2024
[2]

When atrial fibrillation (afib) ablation timing impacts recurrence?

M. Clinic, “When atrial fibrillation (afib) ablation timing impacts recurrence?.” 2025, https://www.mayoclinic.org/medical- professionals/cardiovascular-diseases/news/when-atrial-fibrillation- afib-ablation-timing-impacts-afib-recurrence/mac-20584500. (Accessed: 20 January 2025)

2025
[3]

Association of atrial tissue fibrosis identified by delayed enhancement mri and atrial fibrillation catheter ablation,

N. F. e. a. Marrouche, “Association of atrial tissue fibrosis identified by delayed enhancement mri and atrial fibrillation catheter ablation,” JAMA, vol. 311, no. 5, pp. 498–506, 2014

2014
[4]

Edge-enhanced feature guided joint segmentation of left atrial and scars in lge mri images,

S. Zhou, K.-N. Wang, and G.-Q. Zhou, “Edge-enhanced feature guided joint segmentation of left atrial and scars in lge mri images,” in Challenge on Left Atrial and Scar Quantification and Segmentation, pp. 93–105, Springer, 2022

2022
[5]

Investigating the domain adaptability of general-purpose foundation models for left atrium segmentation from mr images,

B. Kundu, B. Khanal, R. Simon, and C. A. Linte, “Investigating the domain adaptability of general-purpose foundation models for left atrium segmentation from mr images,” inInternational Conference on Functional Imaging and Modeling of the Heart, pp. 275–287, Springer, 2025

2025
[6]

Multi-scale feature fusion with image-driven spatial integration for left atrium segmentation from cardiac mr images,

B. Kundu, Z. Yang, R. Simon, and C. Linte, “Multi-scale feature fusion with image-driven spatial integration for left atrium segmentation from cardiac mr images,” in2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 1–4, IEEE, 2025

2025
[7]

A fully auto- matic deep learning method for atrial scarring segmentation from late gadolinium-enhanced mri images,

G. Yang, X. Zhuang, H. Khan, S. Haldar, E. Nyktari, X. Ye, G. Slabaugh, T. Wong, R. Mohiaddin, J. Keegan,et al., “A fully auto- matic deep learning method for atrial scarring segmentation from late gadolinium-enhanced mri images,” in2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pp. 844–848, IEEE, 2017

2017
[8]

Sequential segmentation of the left atrium and atrial scars using a multi-scale weight sharing network and boundary-based processing,

A. Khan, O. Alwazzan, M. Benning, and G. Slabaugh, “Sequential segmentation of the left atrium and atrial scars using a multi-scale weight sharing network and boundary-based processing,” inChallenge on Left Atrial and Scar Quantification and Segmentation, pp. 69–82, Springer, 2022

2022
[9]

Left atrial scar segmentation and quantification using residual cbam-eam attention unet for lge mri,

Y . Zhang, H. Cheng, D. Li, and L. Pan, “Left atrial scar segmentation and quantification using residual cbam-eam attention unet for lge mri,” inMICCAI Challenge on Comprehensive Analysis and Computing of Real-World Medical Images, pp. 149–157, Springer, 2024

2024
[10]

Atrialjsqnet: a new framework for joint segmentation and quantification of left atrium and scars incorporating spatial and shape information,

L. Li, V . A. Zimmer, J. A. Schnabel, and X. Zhuang, “Atrialjsqnet: a new framework for joint segmentation and quantification of left atrium and scars incorporating spatial and shape information,”Medical image analysis, vol. 76, p. 102303, 2022

2022
[11]

Lassnet: a four steps deep neural network for left atrial segmentation and scar quantification,

A. L. Lefebvre, C. A. Yamamoto, J. K. Shade, R. P. Bradley, R. A. Yu, R. L. Ali, D. M. Popescu, A. Prakosa, E. G. Kholmovski, and N. A. Trayanova, “Lassnet: a four steps deep neural network for left atrial segmentation and scar quantification,” inChallenge on Left Atrial and Scar Quantification and Segmentation, pp. 1–15, Springer, 2022

2022
[12]

Left atrial and scar quantification & segmentation chal- lenge

LaScarQS, “Left atrial and scar quantification & segmentation chal- lenge.” 2022, https://zmiclab.github.io/projects/lascarqs22. (Accessed: 20 January 2025)

2022
[13]

How distance transform maps boost segmentation cnns: an empirical study,

J. Ma, Z. Wei, Y . Zhang, Y . Wang, R. Lv, C. Zhu, C. Gaoxiang, J. Liu, C. Peng, L. Wang,et al., “How distance transform maps boost segmentation cnns: an empirical study,” inMedical Imaging with Deep Learning, pp. 479–492, PMLR, 2020

2020
[14]

Probabilistic two-stage detection,

X. Zhou, V . Koltun, and P. Kr ¨ahenb¨uhl, “Probabilistic two-stage detection,”arXiv preprint arXiv:2103.07461, 2021

work page arXiv 2021
[15]

nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier- Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature methods, vol. 18, no. 2, pp. 203–211, 2021

2021
[16]

Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool,

A. A. Taha and A. Hanbury, “Metrics for evaluating 3d medical image segmentation: analysis, selection, and tool,”BMC Medical Imaging, vol. 15, no. 1, p. 29, 2015

2015
[17]

Chatgpt

OpenAI, “Chatgpt.”https://chat.openai.com, 2024. Large language model accessed January 2026

2024