Segmentation of Ultrasound Images

Yale University, New Haven CT

Investigators

Linked publications & trials

Abstract

[unreadable] DESCRIPTION (provided by applicant): Segmentation, which is the task of finding boundaries in an image, is an important stage in quantifying information in ultrasound images. The proposed research seeks to build a computational framework for segmenting human and animal cardiac ultrasound images. It is based on two ideas: the first uses a probabilistic model for the ultrasound signal and poses segmentation as a MAP estimation problem. The second uses a new optimization strategy called tunneling descent to calculate the MAP estimate. Tunneling descent has the capacity to escape from local maxima of the MAP log-likelihood function. Preliminary results, which compare tunneling descent results to manual segmentation, clearly show that tunneling descent outperforms classical active contours in segmenting short-axis ultrasound images. Experimental evaluation also shows that it is robust with respect to initialization and works reliably without tweaking. This research seeks to extend these ideas to segment more complex boundaries, boundaries with specularities, boundaries with data dropout, moving boundaries in ultrasound image sequences, 3-D ultrasound images, and r.f. ultrasound images. These extensions will be used to jointly segment the endo and epi-cardium in short axis and apical four-chamber views. Human as well as laboratory animal images will be segmented. These images will be made available by the co-investigators. Two cardiac ultrasound phantoms will be built and used for evaluating the accuracy of segmentation as well as accuracy of volume and thickening calculations that are based on segmentation. R.F. data from the phantom will also be collected and systematically manipulated in software to create B-mode images. The segmentation of the r.f. images will be compared to the segmentation of the B-mode images to understand the effect of machine processing on segmentation. The human and animal image segmentations will be compared to manual segmentation. The performance of tunneling descent will also be compared to simulated annealing. [unreadable] [unreadable] [unreadable]