Maximum a posteriori segmentation for medical visualization

Abstract

This is a practical contouring method combining region growing, gradient edge detection, and prior shape constraints to compute contours throughout a three dimensional, computed tomography image dataset. Beginning with a sample of known object interior pixels, alternating steps of incremental region growth are followed by determination of an optimal contour, fitted simultaneously to the current region's perimeter local maxima in the gray level gradient, and to the shapes of prior contours of the object. The resulting contour corresponds to the maximum over all the iteratively-computed contours. Region growing is conducted by a supervised classifier developed on the fly for each object-section. Contours are parametric curves where the parameters are the independent variables of an objective function. The parameters also are treated as random variables whose distributions constrain future contour shapes. Both the region growing and the boundary finding are posed as maximum a posteriori problems. The method propagates contours from section to section using the texture classifier region template, and parametric shape prior probabilities from a previous section's contour to begin contour determination on a succeeding section. Initially intended as a drawing tool to speed-up interactive contouring on CT images in radiation therapy planning, the method is fully competent to run automatically as long as initial object-interior samples are provided.