The Ultrasound Visualization Pipeline - Scientific Visualization: Uncertainty, Multifield, Biomedical, and Scalable Visualization - page 272

24.3.2 Data Enhancement
Ultrasound is a challenging modality for visualization due to its natural properties
such as low dynamic range, noisiness and speckle patterns [ 64 ]. Also, the geometric
resolution varies with depth and the tissue boundaries can be several pixels wide
depending on their orientation. Tissue boundaries can even disappear if they are
parallel to the ultrasound beam. 2D images are preferred without filtering and en-
hancement. Speckle patterns refer to the texture of the tissue boundary, which is
valuable information for clinicians. However, speckle in 3D brings no added value to
the visualization and is considered as an artifact the same as noise. Therefore, prior
to the rendering stage, the 3D data is filtered to enhance its quality.
For a review on early speckle reduction techniques, refer to the survey of Forsberg
et al. [ 16 ]. Belohlavek et al. [ 5 ]usethe eight hull algorithm with a geometric fil-
ter [ 10 ]. Recent techniques are based on region growing [ 9 ], adaptive filtering [ 67 ],
compression techniques [ 26 ] and anisotropic diffusion filters [ 38 ].
Systems usually employ a blend of image-processing techniques to enhance the
data. Sakas et al. listed techniques with a good trade-off between loss of informa-
tion and quality [ 64 ]. They employed Gaussian filters for noise reduction, speckle-
removal methods for contour smoothing and median filters for gap closing and noise
reduction. Median filters remove small surface artifacts and preserve the sharpness
of boundaries. There exist fast implementations where a histogram can be used to
keep track of values [ 29 ]. Still, they require a more advanced memory management,
making them less parallelizable than the evaluation of fast Gaussian filters. Lizzi
and Feleppa described a technique to increase the axial resolution by processing
the signal in the frequency domain. This resolution gain is especially valuable in
opthalmology when visualizing thin layers within the cornea [ 45 ].
24.4 Segmentation
Selecting interesting features to be visualized is important to be able to root out
the occluding elements from large datasets. For most modalities, segmentation can
be performed by extracting regions with similar data values. For instance, because
of the physical properties of x-rays, the data values in a CT scan are recorded into
Hounsfield units which provide a good basis for binary thresholding techniques for
certain tissue types. Early work indicated that binary thresholding techniques are not
very well suited for ultrasound data [ 72 ]. More sophisticated techniques are required
for satisfactory segmentation. An extensive survey on ultrasound image segmentation
was presented by Noble and Boukerroui [ 48 ] in 2006. In this section we have focused
on significant publications from recent years.
To increase robustness of the ultrasound segmentation, the various approaches are
usually tailored for specific anatomies. Carneiro et al. have developed an automatic
technique for segmenting the brain of a fetus [ 8 ]. By first detecting the cerebellum,