CalcSeg: Confidence-aware 3D Latent Context Curriculum Learning for Myocardial Scar Segmentation

Forum topic · 小凯 · 2026-08-22

Summary

CalcSeg is a confidence-aware latent context curriculum learning framework for myocardial scar segmentation from single-stacked late gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) imaging. Scar segmentation in LGE-CMR is clinically important but difficult due to low tissue contrast, diffuse and small scar regions, and limited 3D spatial context in single-stacked acquisitions. The method introduces a dynamic semi-supervised curriculum learning strategy that progressively trains from easier to more challenging scar cases using a learned confidence-aware scoring function. This function combines errors in predicted scar maps, quantified epistemic uncertainty, and scar burden estimates to automatically assess sample difficulty without manual labels. To compensate for the limited spatial context of single-stacked acquisition, CalcSeg develops latent slice-wise self-attention that captures inter-slice dependencies and infers 3D spatial representations from sparse 2D inputs. Evaluations on a multi-center clinical LGE-CMR dataset show that CalcSeg consistently outperforms all competing methods. Paper: arXiv 2608.20305.

Paper Overview

Research area: Computer Vision (CV) Authors: Nivetha Jayakumar, Hannah Kim, Amit R. Patel, Miaomiao Zhang Published: 2026-08-22 arXiv: 2608.20305

Abstract

Myocardial scar segmentation from single-stacked late gadolinium-enhanced cardiac magnetic resonance (LGE-CMR) imaging is a long-standing and clinically important challenge, especially in the presence of low tissue contrast, diffuse and small scar regions. These challenges are further exacerbated by limited 3D spatial context. This paper proposes CalcSeg, a confidence-aware latent context curriculum learning framework that leverages fused 3D feature representations from single-stacked 2D LGE-CMR images for robust scar segmentation.

Specifically, the authors 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. This function integrates errors in the predicted scar maps, quantified epistemic uncertainty, and scar burden estimates to automatically assess sample difficulty without requiring manual labels.

To compensate for the limited spatial context of single-stacked acquisition, the method develops latent slice-wise self-attention that captures inter-slice dependencies and infers 3D spatial representations from sparse 2D inputs.

Results

Evaluation on a multi-center clinical LGE-CMR dataset shows that CalcSeg consistently outperforms all competing methods.

---

*Auto-collected on 2026-08-22*