Left atrial segmentations were performed manually by trained raters who processed the LGE-MRIs using Seg3D software (Scientific Computing and Imaging Institute, University of Utah, Salt Lake City, UT) as described previously 2 (link), 3 (link). The segmentation was edited manually to cover the whole wall in each slice in different areas accounting for atrial wall with varying thickness. The two different scans for each patient were segmented by different raters, each with more than 2 years of experience. All the segmentations were checked at the end by a more experienced rater with more than 7 years of experience to make sure that segmentations were consistent and accurate. The left atrial wall was later masked out on the MRIs for further analysis (Figure 1A-B in supplementary material show the process).
In theory, the intensity histogram of homogenous tissue in an MRI with some noise can be modeled with a Rician distribution 13 (link). Using MATLAB and Statistics Toolbox (Release 2016a, The MathWorks, Inc., Natick, Massachusetts, United States), we removed the outliers from the wall intensity histograms (lowest and highest 1 percentile) and normalized the intensity values to be in the range of 0 to 1, fitted Rician distributions on the intensity histograms and calculated mean intensity value, skewness and standard deviation (SD) of the distributions.