Abstract

Automatic left atrial (LA) scar segmentation from late gadolinium enhancement MRI (LGE-MRI) is challenging due to extreme class imbalance, thin atrial walls, and multi-center domain shift. Prior work commonly adopts a two-phase pipeline where scar predictions are spatially constrained using cavity-derived information. This paper systematically evaluates whether such constraints are beneficial on the Comprehensive Analysis of Real-world medical images (CARE) 2026 Left Atrium track. We find that both hard and soft cavity-derived constraints are detrimental: hard dilation clipping suppresses valid scar predictions, while soft signed distance map (SDM) inputs improve training performance but generalize poorly. A properly sized 6-stage U-Net achieves near-optimal performance, with further increases in model complexity bringing only marginal gains. Beyond spatial constraints, a Gaussian spatial weighting loss derived from scar ground truth improves segmentation, whereas a cavity-wall-derived weighting of identical form brings no benefit, indicating that the benefit of a spatial prior depends on what anatomical structure it is derived from. Built on the nnU-Net framework, our method achieves competitive results on the CARE~2026 benchmark using only the challenge-provided training data, without external datasets or pre-trained models.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/CARE_023.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=y3ZT4z7IKW

BibTex

@InProceedings{WenHao_Beyond_MICCAISAT2026,
        author = { Wen, Hao AND Kang, Jingsu},
        title = { { Beyond Cavity Constraints: A Systematic Evaluation of Spatial Priors in Cascaded Left Atrial Scar Segmentation } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17273},
        month = {pending},
        page = {pending}
}


back to top