Automatic Detection of Left Ventricle in Echocardiographic Sequences using Radial Search and Temporal Smoothing

We present a fully automatic method for obtaining an initial estimate of endocardial border in short axis echo cardiographic videos at two different levels of left ventricle (LV). The geometry of the acoustic window along with circular Hough transform and image statistics is used to robustly identify the region of interest, which encloses left ventricle, irrespective of image quality and level of left ventricle at which image was acquired. Again in the region of interest, left ventricle center point (LVCP) was identified by image statistics. The approximate LV border was detected by radial gradient search with magnitude as well as direction and then smoothing it locally and temporally. The results were validated by comparing computer generated boundaries to those manually outlined by one expert on ten data sets each containing 28 frames over full cardiac cycle. The mean error in the boundaries was plusmn2 mm. Directional radial gradient search along with temporal constraints from adjacent frames gives a better initial estimate of the LV boundary. The contour for each of the frame so obtained may serve a good initialization to optimize it by automated methods based on active contours.