Fig. 1

The proposed PoinUNet model for segmentation of the left atrium (LA) and the surrounding epicardial adipose tissue (EAT) by using water and fat images and fat fraction maps obtained from Dixon MRI sequences. The pipeline comprises two main steps: (Step1) Training a 3D UNet for finding a region of interest selection to actually crop out the predicted region and the feed only that to the step 2 network using fat images as the input; (Step 2) Training a PoinUNet for joint segmentation of EAT and LA. This involves a multi-label segmentation approach with three input channels: water, fat, and fat fraction maps.