Tunneling descent: A new strategy for active contour segmentation of ultrasound images

This thesis presents a new approach toward active contour segmentation in the application of ultrasound images. The specific goal is to automatically segment the endocardium in a cardiac image. Segmenting ultrasound images is known to be a challenging task due to signal drop out, attenuation and diffraction, and most importantly, the presence of speckle. Speckle is a spatial stochastic process that causes the grainy appearance of ultrasound images and low contrast of boundaries. Because speckle is random, it causes many spurious local minima in the energy function of active contours. And when active contours are evolved by classical gradient descent, they get trapped at locations far from the true boundaries.; In this thesis, we first present four empirical models for the gray level in real ultrasound images. These models capture the first order statistical properties of speckle, and are useful for segmentation. We evaluate the accuracy and classification power of these models and use them to design the MAP energy function of the active contour.; To tackle the local minima problem, this thesis introduces a strategy called tunneling descent. Tunneling descent is a deterministic evolution strategy that can detect and escape from local minima. The key idea is to evolve the contour by a sequence of constrained minimizations that move the contour into a local minimum, and subsequently out of the minimum, while growing the curve monotonically. A stopping rule is required to terminate evolution when the active contour passes the desired boundary. Several stopping rules are proposed and evaluated in this thesis.; We integrate tunneling descent into a curve based algorithm and a level set based algorithm. Each algorithm is used to segment 115 short axis ultrasound cardiac images with satisfactory results. Evaluation of the algorithms is carried out by comparing the segmentation results with two sets of manual segmentation. We also compare the performance of the tunneling descent algorithms with the Balloon algorithm.; Although the algorithms are presented here in the context of 2D ultrasound images, they are capable to segment any (first-order) textured image.