Pit Pattern Classification in Colonoscopy using Wavelets - WaveLab

3 Texture classification
After the MFNN has been trained successfully it is able to discriminate between normal and
abnormal texture regions by forming hyperplane decision boundaries in the pattern space.
In [43], [33] and [31] the ANN is fed with feature vectors, which contain wavelet coefficient
co-occurrence matrix based features based on the subbands resulting from a 1-level DWT.
Karkanis et al. use in [25] a very similar approach, but consider only the subband exhibiting
the maximum variance for feature vector creation.
According to the results in the publications the classification results are promising since
the system used has been proven capable to classify and locate regions of lesions with a
success rate of 94 up to 99%.
The approach in [44] uses an artificial neural network for classification too. The color
features extracted in this approach are used as input for a Backpropagation Neural Network
(BPNN) which is trained using various training algorithms such as resilient propagation,
scaled conjugate gradient algorithm and the Marquardt algorithm.
Depending on the training algorithm used for the BPNN and the combination of features
which is used as input for the BPNN this approach reaches an average classification accuracy
between 89 and 98%.
The last of the presented methods also using artificial neural networks is presented in [13].
The 84 features extracted with this method are used as input for a multilayer backpropagation
neural network. This results in a classification accuracy between 85 and 90%.
3.3.3 SVM
The SVM classifier, further described in [7, 20], is another, more recent technique for data
classification, which has already been used for example by Rajpoot in [36] to classify texture
using wavelet features.
The basic idea behind support vector machines is to construct classifying hyperplanes
which are optimal for separation of given data. Apart from that the hyperplanes constructed
from some training data should have the ability to classify any unknown data presented to
the classifier as well as possible.
In figure 3.2(a) an example 2-dimensional feature space with linear separable features
of two classes A (filled circles) and B (outlined circles) is shown. The black line running
through the feature space is the hyperplane separating the feature space into to half spaces.
Additionally on the left side of the hyperplane as well as on the right side of the hyperplane
two other lines can be seen - drawn in gray in figure 3.2. These lines are boundaries, which
have the same distance h to the separating hyperplane at any point.
These boundaries are important, as feature vectors are allowed to be on the boundaries,
but not inside them. Therefore all feature vectors must always satisfy the constraint, that
the distances to the hyperplane are always equal or greater than h. Since there are many
classifying hyperplanes possible the SVM algorithm now tries to maximize the value of h,
such that only one possible hyperplane remains.
30