Cardiac Motion Scoring Based on CNN with Attention Mechanism

Motion scoring of cardiac myocardium is essential for early detection and diagnosis of various cardiac diseases. Existing work on the myocardium motion mainly focuses on binary abnormality detection, while the myocardium motion can be clinically classified into four types: normal, hypokinetic, akinetic and dyskinetic, which has greater significance and is more challenging to predict. The state-of-the-art demonstrated that the method for cardiac motion scoring from MR sequences based on deep convolution neural network (CNN) with non-local attention has great potential. However, due to the complex myocardium deformation and subtle inter-class difference of motion patterns, the performance is still not satisfactory. In this paper, we introduce two types of "attention mechanism" to enhance the ability of the CNN network by effectively extracting the dependency in time-wise, space-wise and cardiac segment-wise. Experiment on 1440 myocardium segments of 90 subjects from short-axis MR sequences of multiple lengths prove that our method's prediction is more precise and consistent which paves the way to the potential implementation in clinical routine.