Spatiotemporal deep learning for scar screening in cardiovascular magnetic resonance: Toward selective use of gadolinium

BACKGROUND: Late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) is the gold standard for assessing myocardial scar. However, a substantial proportion of patients referred for CMR scar assessment are ultimately found to have no evidence of myocardial scarring. PURPOSE: To develop and evaluate a deep learning (DL) model capable of identifying patients without myocardial scar using cine imaging alone, thereby obviating the need for contrast administration and LGE imaging in these individuals. MATERIALS AND METHODS: We developed a novel spatiotemporal DL architecture to identify patients unlikely to have myocardial scar using contrast-free cine images. The model was trained on short-axis cine images from a consecutive cohort of 3000 patients (1753 males; mean age 54 ± 18 years) undergoing CMR for evaluation of known or suspected cardiovascular disease, using 1.5 and 3T scanners from Siemens and GE. External validation was performed in an independent multicenter cohort of 1792 patients where images were acquired on 1.5T and 3T scanners from Siemens and Philips. The architecture utilizes factorized convolutions to extract spatial and temporal features and incorporates residual attention mechanisms to emphasize features most predictive of scar presence on LGE imaging. To evaluate the incremental value of incorporating temporal information, a second model was developed that excluded the temporal kernel from the DL architecture. Both models were trained and optimized using the same training dataset and were evaluated based on similar internal and external testing cohorts. Model performance in identifying patients without myocardial scar was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. RESULTS: The spatiotemporal model had a higher AUC compared to the spatial model in the internal (0.79 ± 0.02 vs. 0.70 ± 0.05, p