Automated endocardial border identification in ultrasound transesophageal echocardiograms using a matched filter

An image processing algorithm was developed for automatically identifying the endocardial border of the left ventricle in short-axis transesophageal echocardiograms. The algorithm was tested on echocardiograms obtained intra-operatively, from patients with cardiac disease. The algorithm was found to be equivalent in accuracy to manual outlining and approached the measured limits of the system resolution. Automated border identification can provide quantitative estimates of global cardiac function for use in patient monitoring and clinical applications.;A maximum likelihood method was used to choose the endocardial border points from matched filter output peaks obtained from radial intensity profiles centered within the ventricle. Erroneously selected points were removed using either a circle or ellipse shape model. Using a data-derived 16-pixel signal template, the algorithm located 77.3% of the border points within the 3.3 mm uncertainty tolerance of the manual reference border. No statistically significant difference in accuracy was found between longer versus shorter, or data-derived versus step templates. Two-dimensional edge detection accuracy was found to be equivalent to simpler one-dimensional methods. Application of median pre-filtering was tested and found to reduce accuracy.;A robust ellipse model, developed for identifying outlier border points, did not improve border accuracy, when applied to short-axis views, but could be useful for border identification in long-axis views or other elliptically shaped organs.;Lateral border dropout, where tangential interaction between the ultrasound, beam and the endocardium occurs, makes border identification difficult. Border edge orientation and contrast were identified, using pairs of two-dimensional, spatially quadrature edge operators. The average relationship between border contrast and orientation to the ultrasound beam, when used to equalize border contrast, did not improve border accuracy. Application as a pre-processing step within the ultrasound instrument may prove valuable.;Matched filtering executes rapidly, does not require time-consuming pre-filtering and performs as well as other reported methods at estimating ventricular area. The algorithm, using the circle or the ellipse model, is flexible and would be capable of performing border identification in a wide variety of ultrasound image analysis applications. A comprehensive summary of the findings of the research are given in the concluding chapter.