Clinical Validation of an Edge Detection Algorithm for Two-Dimensional Echocardiographic Short-Axis Images

Abstract

The purpose of this study was to validate an edge detection algorithm for short-axis two-dimensional echocardiographic studies in a protocol that stimulated its implementation at multiple clinical laboratories. Six short-axis two-dimensional echocardiographic studies were solicited from each of five clinical laboratories. A single cardiac cycle from each of the resulting 30 studies was entered into the computer system. Five expert observers came to the laboratory on separate occasions and traced endocardial borders from the short-axis studies on 2 separate days. The computer algorithm generated borders on each frame of the cardiac cycles on the basis of regions of search defined by the observers. Of the 30 original studies, five were considered excellent, seven were good, nine were poor, and nine were technically inadequate by consensus of the five observers. The correlation coefficient for computer-defined borders with manually defined borders in the excellent quality studies was 0.985. Interobserver variability was expressed as the mean percent area difference for all possible pairings of observers. The mean percent area differences were decreased from +/- 9.8% to +/- 5.3%, +/- 12.5% to +/- 8.4%, and +/- 17.4% to +/- 15.6% when comparing observer with computer-generated borders in the excellent, good, and poor quality studies, respectively. Intraobserver variability was expressed as decrease in mean percent area difference on corresponding frames between days 1 and 2. Intraobserver variability was decreased from +/- 6.5% to +/- 4.5%, +/- 10.8% to +/- 7.0%, and +/- 14.0% to +/- 11.9%, respectively. All reductions in variability were statistically significant at p less than 0.01. Observer acceptance of computer-defined borders was estimated at 94%, 93%, and 97% for excellent, good, and poor quality studies, respectively. Once the observer defined a region of search, computer process time to generate all borders in the cardiac cycle was approximately 4 minutes. The conclusion is that the algorithm produces accurate, reliable, and acceptable borders.