Automatic analysis of three-dimensional cardiac tagged magnetic resonance images using neural networks trained on synthetic data.

BACKGROUND: Three-dimensional (3D) tagged magnetic resonance (MR) imaging enables in-vivo quantification of cardiac motion. While deep learning methods have been developed to analyze these images, they have been restricted to two-dimensional datasets. We present a deep learning approach specifically designed for displacement analysis of 3D cardiac tagged MR images.

METHODS: We developed two neural networks to predict left-ventricular motion throughout the cardiac cycle. Networks were trained using synthetic 3D tagged MR images, generated by combining a biophysical left-ventricular model with an analytical MR signal model. Network performance was initially validated on synthetic data, including assessment of signal-to-noise ratio sensitivity. The networks were then retrospectively evaluated on an in-vivo external validation human dataset and an in-vivo porcine study.

RESULTS: For the external validation dataset, predicted displacements deviated from manual tracking by median (interquartile range) values of 0.72 (1.17), 0.81 (1.64), and 1.12 (4.17) mm in x, y, and z directions, respectively. In the porcine dataset, strain measurements showed median (interquartile range) differences from manual annotations of 0.01 (0.04), 0.01 (0.06), and -0.01 (0.18) for circumferential, longitudinal, and radial components, respectively. These strain values are within physiological ranges and demonstrate superior performance of the network approach compared to existing 3D tagged image analysis methods.

CONCLUSION: The method enables rapid analysis times of approximately 10 s per cardiac phase, making it suitable for large cohort investigations.