Ultrasound images are formed by recording the echo (i.e. reflection and backscatter) from an acoustic wave propagating in tissue. These images contain speckle which contributes to their grainy appearance. Speckle also makes it hard to segment these images. In this talk, we concentrate on a maximum-likelihood active contour algorithm and show that speckle creates many local minima that trap the active contour evolving under gradient descent. We propose a new evolution strategy called “tunneling descent” that overcomes the limitations of gradient descent and is able to segment short-axis cardiac ultrasound images. Experimental results validating the new evolution strategy will be presented.
Much still remains to be done, and the talk will present some of the open questions in ultrasound segmentation. A multiresolution approach is likely to be useful in addressing these questions.
Back to MGA Workshop I: Multiscale Geometry in Image Processing and Coding