Prognostic value of end-to-end deep learning assessment of myocardial scar and microvascular obstruction on late gadolinium enhancement cardiovascular magnetic resonance.

Abstract

Background: Late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) is the reference standard for assessing myocardial scar and microvascular obstruction (MVO), and is a strong predictor of post-acute myocardial infarction (AMI) outcomes. However, manual segmentation is time-consuming and subject to inter-observer variability, limiting clinical scalability. This study develops and validates LGE-CMRnet, an end-to-end deep learning pipeline for automated scar and MVO segmentation on LGE CMR, and evaluates its prognostic value in AMI patients.

Methods: A total of 3874 LGE images from 567 AMI patients (409 for training/internal stress-test cohort; 158 for external testing) were analyzed. LGE-CMRnet integrates YOLOv8 for heart localization and nnU-Net for simultaneous segmentation of myocardium, scar, and MVO. Performance was evaluated using Dice similarity coefficient (DSC), correlation, and Bland-Altman analysis against expert annotations. Prognostic value was assessed using Cox regression for major adverse cardiac events (MACE) over a median follow-up of 24.4 months.

Results: LGE-CMRnet achieved rapid processing (0.05 s per image) and high segmentation accuracy. In the external validation cohort, the model achieved mean DSC of 0.83 ± 0.11 for scar and 0.88 ± 0.11 for MVO at the patient level, with strong volumetric correlations to expert reference segmentations (scar: r = 0.90; MVO: r = 0.98, both P<0.0001). Bland-Altman analysis showed minimal bias in volumetric measurements (scar: 2.5 ± 8.9 cm³; MVO: -0.20 ± 0.89 cm³). Among the 158 patients in the external validation cohort (age 57 ± 10 years, 127 male), 35 (22.2%) experienced MACE. LGE-CMRnet-derived %MVO (hazard ratio [HR], 1.06; 95% confidence interval [CI]: 1.02 to 1.09; P = 0.003) and %Scar (HR, 1.05; 95% CI: 1.02 to 1.08; P = 0.001) were independent predictors of MACE after adjustment for established risk factors. Furthermore, LGE-CMRnet-derived metrics demonstrated non-inferior discrimination for MACE prediction compared with expert analysis. The differences in C-index were 0.02 for %MVO and 0.01 for %Scar, with the lower bounds of the 95% CIs remaining above the pre-specified non-inferiority margin.

Conclusion: LGE-CMRnet enables fast and accurate scar and MVO quantification, with prognostic performance comparable to expert analysis, supporting its potential for automated clinical risk stratification after AMI.