Shape Constraints in Deep Learning for Robust 2D Echocardiography Analysis

Abstract

2D Echocardiography is a popular and cost-efficient tool for cardiac dysfunction diagnosis. Automatic solutions that could effectively and efficiently analyse cardiac functions are highly desired in clinical situations. Segmentation and motion tracking are two important techniques to extract useful cardiac indexes, such as left ventricle ejection fraction (LVEF), global longitudinal strain (GLS), etc. However, these tasks are non-trivial since ultrasound images usually suffer from poor signal-to-noise ratio, boundary ambiguity and out of view problem. In this paper, we explore how to introduce shape constraints from global, regional and pixel level into a baseline U-Net model for better segmentation and landmark tracking. Our experiments show that all the three propositions perform similarly as the baseline model in terms of geometrical scores, while our pixel-level model, which uses a multi-class contour loss, reduces segmentation outliers and improves the tracking accuracy of 3 landmarks used for GLS computation. With appropriate augmentation techniques, our models also show a good generalisation performance when testing on a larger unseen cohort.