Contour detection using artifical neuronal network presegmention

M.Süssner, M.Budil, T.Strohmer, M.Greher, G.Porenta, and T. Binder.

Visual analysis of two-dimensional echocardiograms is based on detection of the endocardial border to asses global and regional wall motion. Human experts rely on information from spatial and time domain. The purpose of this study was to apply artificial neuronal networks (ANN) and to compute the endocardial border by using time domain information. the first processing step of this semi-automatic detection system is the segmentation by extracting the tissue region, which is computed by the ANN using local texture information. A human operator must interactively define the left ventricular (LV) center and a rectangular region of interest surrounding the LV-wall. Starting at the LV-center the algorithm searches for a transition from a ``blood filled'' to a ``tissue'' region in the segmented image and decides then the position of the contour point. Since lateral tissue information is sparse between end-systole and end-diastole the detected contour points can be transformed to intermediate images by applying correlation techniques. Thus sufficient endocardial contour points can be extracted to facilitate an efficient contour linking.

Keywords: algorithms, numerics, neuronal networks
Published: In Proc. Computers in Cardiology, Vienna, 1995.

Download the paper as a GNU-zipped postscript file (471635 bytes).