Two-Stage nnU-Net for Automatic Multi-class Bi-Atrial Segmentation from LGE-MRIs

Abstract: Late gadolinium enhancement magnetic resonance imaging (LGE-MRI) is used to visualise atrial fibrosis and scars, providing important information for personalised atrial fibrillation (AF) treatments. Since manual analysis and delineations of these images can be both labour-intensive and subject to variability, we develop an automatic pipeline to perform segmentation of the left atrial (LA) cavity, the right atrial (RA) cavity, and the wall of both atria on LGE-MRI. Our method is based on a two-stage nnU-Net architecture, combining 2D and 3D convolutional networks, and incorporates adaptive histogram equalisation to improve tissue contrast in the input images and morphological operations on the output segmentation maps. We achieve Dice similarity coefficients of 0.92 +/- 0.03, 0.93 +/- 0.03, 0.71 +/- 0.05 and 95% Hausdorff distances of (3.89 +/- 6.67) mm, (4.42 +/- 1.66) mm and (3.94 +/- 1.83) mm for LA, RA, and wall, respectively. The accurate delineation of the LA, RA and the myocardial wall is the first step in analysing atrial structure in cardiovascular patients, especially those with AF. This can allow clinicians to provide adequate and personalised treatment plans in a timely manner.