A stochastic model for probabilistic determination of left ventricular borders.Biophysics

Left ventricular angiography has gained wide-spread clinical acceptance in cardiac catheterization centers as an important diagnostic tool in the assessment of heart functions in normal and diseased states. The extraction of left ventricular contours from angiographic recordings has facilitated quantitative and dynamic analysis of internal cardiac structure and dimensions. The time and effort required to obtain these data could be decreased dramatically with the aid of a more fully automated system for the detection and analysis of left ventricular borders. A probabilistic algorithm for automated contour detection has been developed which uses "a priori" information extracted from a variety of left ventricular angiographic images. This information consists of video intensity properties and anatomic parameters associated with the manually-defined ("true") contour in each of these images. Thus, the images function as a training set for the development as well as the evaluation of the border algorithm. The development of the algorithm is described in three parts: (1) A Flexible piece-wise parabolic template or model which approximates the shape of the left ventricle is derived using least-squares quadratic fit to given segments of the manually-defined contours in each of the training images. The template provides global guidance to the algorithmic edge-deflection process. (2) Probability functions are then generated from a quantitative comparison of the template and video data in the image with the location and video information specified by the manually-traced contours. (3) These functions are incorporated into a probability product which facilitates simultaneous application of several border-defining criteria, while at the same time, allowing local search operations to be guided by knowledge of global characteristics. Two-dimensional intensity displays of the probability functions also provide a means for qualitatively evaluating the effectiveness of each individual component in the algorithm. The algorithm is applied to single (end-diastolic) frames as well as sequential (1/60 second interval) angiocardiographic images. A quantitative measure of the accuracy of the computer-determined borders is calculated from a comparison of computed contours with those traced by hand. In this way, the same functions used to detect the ventricular outline can also be used to measure the accuracy of the computed contour. Results obtained from applying algorithm to the training images showed 95% of the computer-determined coordinates to lie within ± 4% of the manually-entered border points. Similar data was obtained from an additional test set of images. The algorithm is computationally efficient (less than 10 second per contour), and can be applied with a high degree of accuracy to the recognition of the endocrinal border in both end-diastolic and serial left ventricular angiographic images. Of particular significance in the methods described is the underlying notion of a training set of images, since these images provide a basis for the extraction, characterization, and translation of video-metric data into algorithmic form. Furthermore, since the techniques used to accomplish this task provide a means for both the detection and evaluation of computer-generated contours, a link between the tow previously-separate processes of data acquisition and data analysis is established. It is hoped that the concepts employed in this work will greatly facilitate analysis of the vast amounts of data contained in angiographic recordings and provide as well, additional information pertaining to eh dynamic dimensional and geometric changes occurring in the left ventricular endocrinal wall during contraction and relaxation of the heart.