A Shape Attentive Convolutional Neural Network for Improving the Generalizability of CMR Image Segmentation

Motivation: Collecting contoured medical data is costly, and the performance of deep learning networks can depend on the data acquisition site. Goal(s): Design a novel generalizable network to automatically segment the left and right ventricles and myocardium from cardiac magnetic resonance (CMR) data. Approach: : Combine two CNN-based attention mechanisms and a Transformer attention mechanism to improve performance, test the network on a publicly available CMR dataset measuring segmentation accuracy for each structure and at base, mid, and apex levels. Results: Our network performs better than previous methods averaged over each structure and considerably better at the base and apex levels. Impact: Our network can be trained and validated on CMR data from one site and can accurately segment CMR data from other sites.