Learning Fast Diffeomorphic Registration for Cardiac Motion Estimation in 3D Echocardiography

Abstract

Abstract Echocardiography is a well-established technique for diagnosing and monitoring cardiovascular diseases. Myocardial regional motion analysis has the potential to offer a comprehensive understanding of cardiac health. Accurate and efficient quantification of cardiac motion is crucial for clinical tasks related to cardiac diagnosis and prognosis. Although various methods have been proposed for cardiac motion tracking, they often suffer from long inference times or require numerous adjustable parameters. In this study, we propose a diffeomorphic registration network (DRN) to take advantage of deep neural networks’ capability for registration purposes: their ability to learn complex representations and predict the desired output in a single step. Diffeomorphic image registration offers distinct advantages, such as generating invertible and topology-preserving deformation fields, which are key attributes in the context of cardiac motion. This approach enhances the accuracy and realism of deformation field estimation. The DRN framework employs a neural network to estimate a stationary velocity field, from which the deformation field is derived through the Scaling and Squaring method. Evaluation of methods on a synthetic echocardiography dataset with ground truth displacement vectors demonstrates the superiority of our approach in terms of accuracy. Importantly, it maintains a practical inference time of 316.63 ms per frame, making it suitable for clinical applications. Notably, DRN trained with segmentation labels achieves the best results, with a mean error of 0.81 mm for global tracking error.