Automatic Segmentation of the Left Atrium from LGE-MRI Based on U-Net and Bidirectional Convolutional LSTM

Abstract

Atrial fibrillation (AF) is the most common cardiac arrhythmia causing morbidity and mortality. The segmentation of the left atrium (LA) is very important for image-guided ablation of AF and quantification of left atrial fibrosis. However, manual segmentation is labor-intensive and highly subjective. Therefore, the automatic segmentation of the left atrium is of great significance. In this study, we developed a U-Net based network for automatic segmentation of the left atrium. Due to the high computational cost and GPU memory consumption of 3D deep learning networks, a 2D network was used for learning. However, the 2D network only learned the features in 2D slices. The spatial context information of the images was not considered and used, so bidirectional convolutional long short-term memory (LSTM) was combined to obtain the context information in the z-axis direction. This study aimed to design a two-step method based on U-Net and bidirectional convolutional LSTM for the automatic segmentation of the left atrium from LGE-MRI. The model was trained and tested on the dataset of the 2018 Atrial Segmentation Challenge. The dice coefficient obtained by the method was 0.906. By combining the context information between image slices, the segmentation results were optimized.