SA-net: A sequence aware network for the segmentation of the left atrium in cine MRI datasets

Abstract

The segmentation of the left atrium (LA) in CINE MRI is a prerequisite for the calculation of LA functional parameters and may be useful when selecting treatments for atrial fibrillation patients. CINE MRI usually captures both the LA and the left ventricle. The similarities between the LA and other cardiac structures complicate the segmentation of the LA and can lead to poor performance of standard 2D segmentation networks. In this study, we present SA-Net, a deep network which implicitly discriminates LA slices from non-LA slices during segmentation, with a sequence modulator using interslice correlations in a global context. Our experiments, conducted on an in-house dataset with 4710-mm thick bSSFP MR image stacks, show that SA-Net leads to good quality segmentation of the LA, with a mean Dice score of 0.89 and a mean Jaccard index of 0.80, outperforming the U-Net.