Fig. 3

Performance evaluation of segmentation. Boxplots of Dice coefficients and maximum Hausdorff-distance (in mm) values for left atrium area segmentation at end-systole in the validation (sample B) and test (sample C) subsets. In blue radiologist₁ vs. segmentation algorithm, in orange radiologist₂ vs. segmentation algorithm and in green radiologist₁ vs. radiologist₂. Latter comparison was performed for sample B. The higher variability for sample B compared to C can be explained on the one hand by the difference in sampling type (B mostly included cardiac diseased subjects, while C was random) and on the other hand by the age of the exams: B was selected from older retrospective studies with heterogenous acquisition parameters, while C was selected from a more homogeneous sample acquired within a year. LAx = long-axis