Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 223

Once an incorrect result is observed for an input image either by the user or by the
automatic evaluation procedure, the case is tagged as a bad case. In a successive step,
the best segmentation parameters for that image are determined and the attributes,
necessary for similarity determination, are computed from the image. Both the seg-
mentation parameters and the attributes calculated from the image are stored into the
case-base as a new case. In addition to that, non-image information is extracted from
the file header or from any other associated source of information and is stored to-
gether with the other information in the case-base. If the case-base is organized hier-
archically, the new case has to be stored at the position in the hierarchy suggested by
its similarity-relation to the other cases in the case-base.
During storage, case generalization is done to ensure that the case-base does not
become too large unnecessarily. Cases that are similar to each other are grouped to-
gether in a case class and, for this case-class, a prototype is computed by averaging
the values of the attributes. Case generalization ensures that a case is applicable to a
wider range of segmentation problems.
It can also happen that several cases are quite similar to each other and, therefore,
they get retrieved for a new problem at the same time. Then, learning the similarity by
updating the local weights associated to each attribute is advised.

3 The Watershed Algorithm: What Influences the Result of the
Segmentation?

An easy way to explain the idea of the Watershed Transform is the interpretation of
the image as 3D landscape. Therefore we allocate for every pixel its grey value as z-
coordinate. Now we flood the landscape from the regional minima, after having bored
the local minima and sunk the landscape into water. Lakes are created (basins, catch-
ment basins) in correspondence with the regional minima. We build dams (watershed
lines or simply watersheds), where two lakes meet. Alternatively, instead of sinking
the landscape, the latter can be flooded by rainfall and the watershed lines will be the
lines of attracting the rain that will fall on the landscape. Whichever of the paradig-
mas is used - the immersion or rainfall paradigma - to obtain its simulation two ap-
proaches are possible: either the basins are identified, then the watershed lines are
obtained by taking a set complement, or the complete image partition is computed,
then successively the watersheds are found by detecting boundaries between basins.
Many algorithms were developed for computing the Watershed Transform. For a
survey see e.g. [8]. In this work we deal mainly with the Watershed Transformation
scheme suggested by Vincent and Soille in [9], and use this scheme also for the new
implementation that we have done for the segmentation algorithm from Frucci et al.
[10], [11].
The first definition of the so called Watershed Transform by immersion was given
by Vincent and Soille [9]. Let D be a digital grey value image with

h h

and as

min max

MIN

the minimal and maximal gray value. is the union of all regional minima with

h

Furthermore, let BD

⊆

hhh ∈ [ , ].

grey value h with and suppose that B is

min max