A Generative Framework for Predicting Myocardial Strain from Cine-Cardiac Magnetic Resonance Imaging

Abstract

Myocardial strain is an important measure of cardiac performance, which can be altered when ejection fraction (EF) and other ventricular volumetric indices remain normal, providing an additional indicator for early detection of cardiac dysfunction. Cardiac tagging MRI is the gold standard for myocardial strain quantification but requires additional sequence acquisition and relatively complex post-processing procedures, which limit its clinical application. In this paper, we propose a framework for learning a joint latent representation of cine MRI and tagging MRI, such that tagging MRI can be synthesised and used to derive myocardial strain, given just cine MRI as inputs. Specifically, we use a multi-channel variational autoencoder to simultaneously learn features from tagging MRI and cine MRI, and project the information from these distinct channels into a common latent space to jointly analyse the multi-sequence data information. The inference process generates tagging MRI using only cine MRI as input, by conditionally sampling from the learned latent representation. Finally, automated tag tracking was performed using a cardiac motion tag tracking network on the generated tagging MRI, and myocardial strain was estimated. Experiments on the UK Biobank dataset show that our proposed framework can generate tagging images from cine images alone, and in turn, can be used to estimate myocardial strain effectively.