Improved accuracy of echocardiographic endocardial borders by spatiotemporal filtered Fourier reconstruction: description of the method and optimization of filter cutoffs.

The usefulness of digitized echocardiographic borders in quantitative regional left ventricular function analysis has been limited by the wide reported range for normal wall motion with this technique. We postulated that random error in endocardial border positioning is a major cause of this limitation. To test this hypothesis, we traced the endocardial borders field by field from 17 complete echocardiographic cycles in six dogs. These cycles showed a great deal of random movement, with each endocardial point reversing its motion an average of 18.5 times per cardiac cycle. Spatiotemporal Fourier analysis of these sequences demonstrated that most of the valid information on endocardial motion was contained in the first four temporal harmonics and the first seven spatial harmonics and that beyond these points the Fourier transform has the spectral characteristics of noise. Reconstruction of these 17 cycles eliminating all Fourier components above the sixth temporal and eighth spatial harmonics reduced the mean number of endocardial reversals per cycle to 2.3 (p less than .00001). To derive the optimal temporal and spatial cutoffs, we compared reconstructions of each of the 17 cycles with three M mode echocardiograms obtained simultaneously with the cross-sectional images. Fourier cutoffs were varied between two and 20 harmonics and demonstrated that the optimal temporal cutoff was 5.5 harmonics and optimal spatial cutoff 6.9. With optimal filtering, the correlation between ventricular diameter derived from the M mode and from the cross-sectional images was r = .965, compared with .877 for the M mode vs unfiltered cross-sectional data (p less than .0001). We conclude that two-dimensional filtered Fourier reconstruction significantly improves the accuracy of traced echocardiographic borders. This technique should be useful in the postprocessing of endocardial borders extracted by automated edge detection schemes and should also be applicable to cardiac images derived from modalities other than echocardiography.