Autonomous epicardial and endocardial boundary detection in echocardiographic short-axis images

An autonomous endocardial and epicardial boundary detection (ABD) method is reported. One hundred ten cycles from 55 clinical studies were selected retrospectively. Image sequences were digitized at 512 x 480 pixel resolution. The point-by-point boundary positions of the ABD and the areas enclosed were compared with positions and enclosed areas drawn by three independent observers. Correlation coefficients for epicardial end-diastolic (ED) and end-systolic (ES) areas, endocardial ED and ES areas, muscle area, and fractional area change were 0.970, 0.976, 0.951, 0.985, 0.887, and 0.878, respectively. Bland-Altman analysis showed negligible biases with standard deviations comparable to those of the observers. The mean difference between the ABD border and the consensus observer border positions in 64 directions falls within the mean range of interobserver border positions, suggesting that shape is also well defined by the ABD.