Proceedings Article | 12 March 2007
KEYWORDS: 3D image processing, Signal attenuation, Ultrasonography, Image segmentation, 3D image enhancement, Visualization, Image processing, 3D acquisition, Anisotropic diffusion, Image processing algorithms and systems
Ultrasound (US) is one of the most used imaging modalities today as it is cheap, reliable, safe and widely
available. There are a number of issues with US images in general. Besides reflections which is the basis
of ultrasonic imaging, other phenomena such as diffraction, refraction, attenuation, dispersion and scattering
appear when ultrasound propagates through different tissues. The generated images are therefore corrupted by
false boundaries, lack of signal for surface tangential to ultrasound propagation, large amount of noise giving
rise to local properties, and anisotropic sampling space complicating image processing tasks.
Although 3D Transrectal US (TRUS) probes are not yet widely available, within a few years they will likely be
introduced in hospitals. Therefore, the improvement of automatic segmentation from 3D TRUS images, making
the process independent of human factor is desirable. We introduce an algorithm for attenuation correction,
reducing enhancement/shadowing effects and average attenuation effects in 3D US images, taking into account
the physical properties of US. The parameters of acquisition such as logarithmic correction are unknown, therefore
no additional information is available to restore the image. As the physical properties are related to the direction
of each US ray, the 3D US data set is resampled into cylindrical coordinates using a fully automatic algorithm.
Enhancement and shadowing effects, as well as average attenuation effects, are then removed with a rescaling
process optimizing simultaneously in and perpendicular to the US ray direction. A set of tests using anisotropic
diffusion are performed to illustrate the improvement in image quality, where well defined structures are visible.
The evolution of both the entropy and the contrast show that our algorithm is a suitable pre-processing step for
segmentation tasks.
