Objectives: To evaluate the feasibility of a non-contrast cardiac magnetic resonance (CMR)-based deep learning (DL) model for predicting left ventricular adverse remodeling (LVAR) in patients with acute ST-segment elevation myocardial infarction (STEMI).
Methods: A retrospective study included 252 patients with STEMI from two centers, randomized into training (n = 176) and testing (n = 76) cohorts. A two-stage DL framework was employed: (1) An architecture for coarse-to-fine myocardial localization and segmentation based on a 3D U-shaped network and (2) a classification model integrating imaging, morphological, and motion features extracted from cine CMR. The performance of different models was evaluated using the area under the curve (AUC), accuracy, sensitivity, specificity, and the F1 score. Regions influencing the decision-making process of the DL model were highlighted using guided gradient-weighted class activation mapping.
Results: The DL model demonstrated robust ability to predict LVAR, with an AUC of 0.865 (95% CI: 0.755-0.956), accuracy of 82.9%, sensitivity of 77.3%, specificity of 85.2% and F1 score 0.723 in the testing set. In multivariable analysis, conventional CMR parameters, including global longitudinal strain, left atrial reservoir strain, and infarct size, remained as independent predictors of LVAR. A combined model integrating DL features with conventional non-contrast CMR parameters improved the predictive performance (AUC: 0.889, 95% CI: 0.803-0.974 in the testing set), significantly outperforming both conventional non-contrast and contrast-enhanced CMR models.
Conclusion: A non-contrast DL-CMR model effectively predicts LVAR in patients with STEMI, providing a gadolinium-free tool for risk stratification and personalized management.
Keywords: Adverse remodeling; Cardiac magnetic resonance; Deep learning; Myocardial infarction; Strain.
Copyright © 2026. Published by Elsevier B.V.