Meta-learning for Image Processing Based on Case-Based Reasoning - Computational Intelligence in Healthcare: Advanced Methodologies - page 250

6 Discussion

The image descriptions that work well for the Watershed Transformation are the sta-
tistical and texture feature description (STDescript) and the image description based
on Central Moments (CMDescript). While the STDescript is only rotation and scale
invariant, the STDescript weighted and the CMDescript is not. The obtained groups
for the two descriptions are:

STDescript
G1={neu4; neu4_r_180; neu1; neu3}, G2={neu2}, G=3{ parrot}, G4={gan128},

G5={monroe}, and G6={cell}

STDescript weighted
G1={neu1;neu3}, G2={neu4_r180}, G3={neu2}, G4={neu4}, G5={gan128}, G6={parrot},
G7={cell}, and G8={monroe}

CMDescript
G1={Monroe, parrot}, G2={neu3, neu4}, G3={gan128}, G4={neu4_r_180},

G5={neu2}, G6={neu1}, and G7={cell}.

The obtained groups of the STDescript weighted seem to reflect better the relation-
ship between the image characteristics and the segmentation parameters than the ob-
tained groups by the CMDescript.
The computation time of both image descriptions is more or less the same.
In conclusion, we can say that the statistical and texture feature description
weighted and the central moments are the best image description which we have
found so far during our study.

7 Case Generalizations and Incremental Model Learning

For each case group we can compute a case-class representative that will be used
during case retrieval as matching candidate. The case representative can be the mean
over all cases or the median (see for example table 5). If all cases in a case group
share the same segmentation parameters, matching will stop after having found the
closest case representative among all case representatives. If not all the cases in a case
group share the same segmentation parameters, matching will proceed until the clos-
est case of all the cases in a group of cases is found (see Figure 40 for example of the
case base based on the image description of STDescript weighted. Over the course of
time new cases can be learnt during the application of the Watershed Transformation
based on CBR. The matching procedure is done based on the similarity measure de-
scribed in Section 5.1.

Table 5. Case group with its case description and prototype

ImageName Mean StdDev Kurtosis VarCoef … Energy Contrast LocHomog xCentroid yCentroid

neu1r.bmp 16,55 23,58 4,48 1,43 … 0,00748 267,45 268,45 56,98 75,67

neu3r.bmp 15,06 22,43 6,59 1,49 … 0,00558 268,04 269,04 58,55 71,86

Prototype Σ 15,80 23,01 5,54 1,46 … 0,00653 267,75 268,75 57,77 73,76