Quantification of Extracellular Volume Fraction in Cardiac MRI without Blood Sampling Using Multi-Stage Training Deep Learning

Motivation: Blood sampling for hematocrit (HCT) measurement limits clinical use of extracellular volume fraction (ECV) in diagnosing myocardial diseases. Goal(s): Develop and evaluate a deep learning (DL) model to predict HCT from cardiovascular magnetic resonance (CMR) data, exploring if additional features improve predictability across multiple centers. Approach: Trained a multi-stage DL model using multi-center CMR T1 values and clinical features to predict HCT without blood sampling. Results: The DL model identified native blood-pool T1 and gender as optimal features, achieving higher correlation with true HCT (R=0.65) than linear regression (R=0.59) and strong agreement between synthetic and true ECV (R=0.95). Impact: This study shows that incorporating additional features in a DL model enhances HCT prediction from CMR data, eliminating the need for blood sampling. This advancement could streamline ECV measurement, making it more accessible for diagnosing myocardial diseases in clinical settings.