Disagreement-Driven Uncertainty Quantification in Late Gadolinium Enhancement Cardiac MRI

Matthias Schwab, Markus Haltmeier, Agnes Mayr

Keywords: Cardiac MRI · Late gadolinium enhancement · Myocardial infarction · Uncertainty estimation · Deep learning · Segmentation.

Abstract:

Accurate segmentation of infarcted myocardium and microvascular
obstruction (MVO) in late gadolinium enhancement (LGE) cardiac
magnetic resonance (CMR) imaging is critical for risk assessment
in myocardial infarction patients. However, the task is challenging due
to anisotropic CMR resolution, complex enhancement patterns, and severe
class imbalance. In this work, we propose a cascaded deep learning
framework that combines a 2D slice-wise segmentation network with a
3D correction network to provide enhanced voxel-wise uncertainty estimation.
We introduce a novel uncertainty estimation approach that
leverages disagreement between the 2D and 3D models as a proxy for
segmentation uncertainty. We quantify this via a Soft Correction Score
(SCS), based on probabilistic discrepancies, and a Discrete Correction
Map (DCM), which encodes interpretable label corrections between networks.
We evaluate our framework on the publicly available EMIDEC
dataset and on a large in-house clinical dataset. Across both datasets,
our framework achieves superior segmentation accuracy and provides
uncertainty estimates comparable to established methods such as Monte
Carlo Dropout, test-time augmentation, and deep ensembles. The proposed
uncertainty measures correlate strongly with prediction errors and
offer interpretable insights into ambiguous regions, enhancing both the
reliability and clinical utility of automated LGE CMR analysis.

Submission Number: 3