Medical Image Processing - Signal Processing Systems - page 355

5 Image Preprocessing

Noise is inseparable from the physical processes and electronics involved in medical
image generation. Ultrasound images are characteristically noisy and laden with
speckles. Quantum noise and radiation scattering exhibit as intensity variation even
for a uniform medium in CT. PET and SPECT images, likewise, are affected by the
random statistical nature of radioactive decay and photon absorption by the body
and the external detectors. Statistical fluctuations of the FID sensed by receiver coils
and mechanical vibrations on the coils themselves similarly introduce noise in MRI.
The purpose of image preprocessing is to suppress noise and improve image contrast
such that the reviewing physician can more easily identify features of interest. Image
preprocessing is also often necessary as a data preparation step prior to applying
more advanced image processing tasks of segmentation and registration described
later.
Most classical image filtering approaches are applicable to medical images.
These include both the spatial and frequency domain filters. In the spatial do-
main, the preprocessing can be pixel- or voxel-based, or neighborhood-based.
An everyday example of pixel/voxel-based preprocessing in medical imaging is the
practice of applying window/level. CT and MRI produce 12-bit images, whereas
most displays remain 8-bit for gray-scale viewing. The necessary 12-to-eight-bit
conversion maps a narrow range of 12-bit intensities (called “window;” “level”
is the center of the window), coinciding with those of features of interest, to a
0-to-255 range. Another common technique belonging to the same category is
histogram equalization. Through intensity redistribution, the goal here is to convert
the nonuniform histogram of an image into a uniform histogram as much as possible.
Histogram equalization may help improve the visibility of a target structure in some
applications. In others, it may have an opposite effect and histogram modification,
reshaping of a histogram to mimic a desired distribution, may be more effective [ 6 ] .
Neighborhood-based preprocessing includes such low-pass spatial domain filters
as the Gaussian and rank-order median filters. The well-known edge detection and
image sharpening techniques for enhancing higher frequencies are also common in
medical image processing. Often adaptive versions of neighborhood-based filters
yield better results. Examples of these include adaptive arithmetic mean filters
as well as those for feature enhancement [ 6 ] . The adaptive anisotropic diffusion
filter [ 22 ] has been shown to be highly effective in achieving simultaneous noise
suppression and edge enhancement. The basic premise of anisotropic diffusion
filtering is to suppress noise and preserve edges by encouraging intra-region smooth-
ing while prohibiting inter-region smoothing. This is achieved by the following
iterative process. For an image I

,where v , is a vector in the 2D/3D space
and t a given point in time (i.e., iteration number), anisotropic diffusion filtering
is given by ∂ I ( v , t )
∂

( v , t )

= div ( c ( v , t ) ∇ I ( v , t )) . ∇ I ( v , t ) is the gradient image and c ( v , t )

t

is a monotonically decreasing function of the gradient ( c ). The c -

term encourages less diffusion (smoothing) in the vicinity of high gradients (strong

= f ( ∇ I ( v , t ))