Performance of User Independent Echocardiographic Border Detection Algorithm: Comparison with Human Observer Variability

Abstract

We evaluated a method for autonomous, user-independent automated border delineation (ABD) developed by Geiser and Wilson, by comparing the accuracy of ABD relative to manual border tracing. Short axis echocardiographic images of 84 patients from 3 clinical sites were analyzed using ABD and by manual tracing performed by two observers at each site and two observers at a core laboratory. The centerline method was used to measure the distance between each pair of computer-generated and hand-traced borders. Cardiac parameters were also measured from all sets of borders: LV area, fractional area change, antero-posterior diameter, wall motion, and wall thickening. The distance between computer-generated and hand-traced borders was slightly but significantly greater than human interobserver variability between the clinical sites and the core laboratory (0.34+/-0.25 (N = 328) vs. 0.26+/-0.16 (N = 320) cm for the endocardium at end diastole, p = 0.0001). Measurements of LV area and fractional area change were similar by ABD and manual tracing. Other cardiac parameters showed greater deviation between ABD and manually traced borders than between human observers. Autonomous ABD provides accurate measurements of LV area and area-derived indices. However measurements dependent on border point location deviate more by ABD.