Automatic detection of left ventricular shape in short axis ultrasound data using a 3D graph search algorithm

Manual tracing is usually required to assess left ventricular (LV) wall boundaries and thicknesses of diseased hearts with echocardiography, a labor-intensive process that limits clinical applicability. The authors present a 3D graph search algorithm that accurately and automatically determines LV epicardial shape from parasternal short-axis ultrasound (SAX) data. SAX images obtained for 20 subjects (11 normal, 9 abnormal LV morphology) at four levels of the LV were used for this study. Quantitative assessment of the method was performed by linear regression analysis of cross-sectional areas of expert-traced contours and computer-detected contours. An intersection-to-union ratio (IUR) metric was used to assess shape similarity. Linear regression yielded a slope and intercept of 0.96/spl plusmn/0.06 and 3.78/spl plusmn/2.83 cm/sup 2/ respectively with R=0.97, n=80. Shape analysis gave an IUR value of 0.89/spl plusmn/0.03. Normal and diseased results are also discussed.