Left Ventricle Showed by Ultrasound

To identify the cylinder in the image data, core atoms of an appropriate range of diameters Automated Identication and Measurement of Objects via Medial Primitives were collected in sample volumes on a regular lattice, and ellipsoidal voting was applied.  Crosses are shown in the cylindrical chamber of the ventricle. Due to the pre{selection of core atoms by scale, no other signi cant densities of core atoms were found. Next, the mitral valve was sought, by limiting the formation of core atoms to an appropriately smaller scale, and to orientations nearly perpendicular to the transducer. As shown in Fig. 10B, the strongest superdensities (short vertical line segments) were clustered around the center of the mitral valve, although weaker false targets were detected in the myocardium. To eliminate these false targets, a criterion was established for the formation of appropriate pairs of superdensities, in the spirit of core atoms. Only slab{like densities appropriately located and oriented with respect to cylindrical densities were accepted. These pairs were allowed to vote for their constituent superdensities, and the mean location of the winning superdensities used to establish a single mitral valve lo96 G. D. Stetten and S. M. Pizercation and a single LV cylinder location. The vector between these two locations was used to establish a cone for expected boundary points at the apex of the LV, and the mean distance to the resulting boundary points used to determine the location of the apical cap along that vector. Thus an axis between the apex and the mitral valve was established. Given this axis, LV volume was estimated by collecting boundary points around the axis. Only boundaries that faced the axis were accepted. The boundary points were organized into bins using cylindrical coordinates, in other words, disks along the axis and sectors within each disk. An average radius from the axis was established for the boundary points in each bin, creating a surface map of the endocardial surface.

The problem of empty bins was avoided by convolving the surface map with a binomial kernel in 2D until each bin had some contribution to its average radius. Volumes were then calculated by summing over all sectors. The entire procedure including identification and volume measurement of the LV was automated, and required approximately 15 seconds on a 200 MHz Silicon Graphics O2 computer. The automated volumes were compared to manual tracings performed on a stack of flat slices orthogonal to a manually {placed axis  ). This axis employed the same anatomical end{points (the ventricular apex and the center of the mitral valve) as the axis determined automatically above. The volumes and locations of the end{points were compared to those determined automatically.  They are very encouraging, particularly for the automated placement of the axis end points, which had an RMS error of approximately 1 cm. Volume calculations introduced additional errors of their own, but were still reasonable for ultrasound. Only four cases have been tried, and all are shown. The method worked in all cases.

We have described a new method for identifying anatomical structures using fundamental properties of shape extracted statistically from populations of medial primitives, and have demonstrated its feasibility by applying it under challenging conditions. Further studies are presently underway to establish reliability over a range of data. Future directions include introducing greater speficity and adaptability in the boundary thresholding, incorporating more than 2 nodes into the model, introducing variability into the model to reflect normal variation and pathologic anatomy, extending the method to the spatio {temporal domain, and applying it to visualization.