Deep relative motion analysis for the identification and phenotyping of scarred myocardium using cine-MRI

The identification and phenotyping of scarred myocardium using cine magnetic resonance imaging (cine-MRI) play a pivotal role in the diagnosis and treatment of cardiovascular diseases. Myocardial motion tracking has garnered widespread attention for cine-MRI analysis. However, the complex myocardial motion and motion-related deformations limit the performance of existing methods. In this paper, we present a deep relative motion representation and learning framework. Our relative motion descriptor focuses on two aspects: static features and dynamic features. Particularly, we first project rays at specific angles from the geometric center of the blood pool in each frame, intersecting with the endocardium and epicardium. Subsequently, we represent the myocardial motion features based on the displacement and curvature of intersection points relative to the geometric center point in different frames. To further explore the motion features, we also introduce a Multi-Orientated Spatio-Temporal Multi-Layer Perception (MOST-MLP) for myocardial motion encoding. The proposed MOST-MLP is evaluated on one private dataset comprising 450 subjects and two public datasets (ACDC and M&Ms), with its strong performance across these benchmarks demonstrating the method’s effectiveness and superiority. Our code and pre-trained models are available at https://github.com/gaoningn/MOST-MLP to facilitate reproducibility and further research.