Digital Signal Processing Reference
In-Depth Information
Threshold-based algorithms can also be useful in identifying structures depicted
by edge/surface points. An example is the Canny edge detector that uses the
threshold of gradient magnitude to find the potential edge pixels. Threshold-based
technique inherently do not account for spatial information. So, their effectiveness
is limited to images where large chunks of pixels/voxels have similar characteristics.
Also, these algorithms are very sensitive to image noise and partial volume
effects, which in medical imaging, refers to the scenario wherein insufficient image
resolution results in anatomical/physiological information from different tissue
types mixing within a single pixel/voxel. Due to a combination of these factors,
regions or edges detected by these algorithms can often be sets of discrete pixels, and
therefore may be incomplete or discontinuous. In such cases, it is necessary to apply
post-processing like morphological operation to connect the breaks or eliminate the
holes. Morphological operations refer to the group of image processing operations
that process images based on shapes. These operations determine the value of each
pixel in the output image based on a comparison of the corresponding pixel in the
input image with its neighbors. By appropriately selecting the size and shape of the
neighborhood and the comparison criteria, it is possible to grow/shrink/merge/split
chunks of pixels in an image to connect or eliminate the holes/breaks in an image.
Incorporation of such morphological operations is also useful in automatic threshold
detection using histogram-based approach in cases where it may be difficult to
identify significant peaks and valleys in the image to define the threshold criteria.
6.3
Deformable Model-Based
Segmentation algorithms using deformable models, also known as active contour
models, form a class of connectivity-preserving techniques that forms a very reliable
and robust approach to edge/surface-based image segmentation.
Segmentation starts with placing, either manually or through automatic initial-
ization, a template of the expected shape, modeled as a 2D contour or 3D surface, in
the vicinity of the (perceived) true borders of the object of interest. A second step,
called energy minimization step, then refines the shape template appropriately under
image-derived and shape integrity preserving constraints to make it snap to the true
borders.
The initial template can be defined parametrically (e.g., a line, circle, ellipse,
sphere, etc.) or can be a more complex model that represents the object to be
segmented (e.g., left ventricle of the heart, etc). The shape template is generally
sampled as discrete points (vertices), each of which is moved during the iterative
energy preservation step under the influence of internal and external forces acting at
that point. The internal forces are model-/template-derived forces that maintain the
integrity (smoothness, regularity) and overall shape of the template and are usually
defined through the geometric properties of the template such as length, area and
curvature. The external forces are typically derived based on the image information
(such as intensity distribution, strength of edges, etc.) and drive the template toward
