| Abstract—Automated, accurate 3D segmentation is critical to achieve the full potential of 3D imaging. Three-D image volumes were acquired in the form of: (i) Fields-II generated 3D cyst images, (ii) 3D scans of tissue-mimicking phantoms with inclusions of varying shape and contrast levels, and (iii) sets of clinical 3D scans containing the prostate. Preprocessing with four different 3D speckle reduction schemes were evaluated, and Integrated BackScatter (IBS) calculations were optionally applied to the images where the RF signal was available. Segmentation was performed directly in 3D using the level set method, and the level set function was manually initialized. Balloon, curvature, and advection forces were applied to the propagating surface to minimize the energy function of the evolving surface. Relative to other segmentation methods, the level set segmentation yielded a smoother and more realistic looking segmented surface. The segmentation results were obtained in the form of rendered 3D images and numerically as volume errors and surface errors. The ground truth was known for the simulated cysts and the phantom inclusions and was for the clinical images obtained by means of hand-segmentation. The smallest RMS surface error was obtained for the Fields-II simulated cysts, in the order of 1.4 mm, while the RMS error for the 3D tissue-mimicking phantoms spread over a wider range from 1.2 mm to 5.9 mm. For the prostate images, the anisotropic diffusion filter gave a mean RMS distance between hand segmentation and the level set segmentation of 2.0 mm to 3.2 mm. A better segmentation was achieved without than with the IBS process applied. |