Maximum-likelihood Segmentation of Ultrasound Images by Tunneling
InformationHemant D. Tagare, PhD (bio)
March 19, 2007
469 Recorded video views
0 Live stream views
Ultrasound images are notoriously difficult to segment with classical segmentation algorithms like active contours. The problem is that the energy function of active contours exhibits multiple local minima in ultrasound images. When the active contour evolves under gradient descent it gets trapped in these local minima, giving wrong segmentations.
In this talk Hemant Tagare will propose an alternate evolution strategy for active contours called tunneling descent. Tunneling descent is a deterministic evolution strategy that "tunnels out of" spurious local minima. It is designed to be a replacement for gradient descent. When used in a maximum-likelihood formulation, tunnel descent can successfully find the endocardium in short axis ultrasound images without over 100 short axis images with literally the same algorithm. I will present validation results for the segmentation as well.