Skin Lesion Classification - Biomedical Signal Analysis: Contemporary Methods and Applications - page 338

. . .

Figure 12.7
Basic procedure for converting the N × N image matrix into an N 2 -dimensional
image vector. The rows of the matrix are subsequently combined, resulting in an
N 2 -dimensional vector.

methods in image processing. The goal was to stress either contrast
or smoothness of the images. Thus, we generated different ensembles of
images for each cell type.

Presentation of the samples

For any analysis of the fluorescence images, a rearrangement of the image
matrices proved necessary either to calculate the covariance matrix
needed for PCA or to train the neural network when performing ICA.
Therefore consider an image X ,the N 2 pixels of which are stored in
an N N matrix according to X =( x ij ) 0<i≤N,0<j≤N . An image vector

x =[ x (1) ,...,x ( N 2 )] can then be created by subsequently concatenating
the rows of the matrix X , thus obtaining an N 2 -dimensional image
vector, as illustrated in figure 12.7.

×

Reconstruction error-based classification using PCA

First the PCA eigenimages are computed for the different ensembles of
skin lesions and the corresponding relative reconstruction error ε rel

i is
determined, following equation (12.6). In figure 12.8, ε re i for 20 fluores-
cence images of each type of skin lesion based on a reconstruction by
psoriasis eigenimages is shown. An image is then classified as belonging
to class i when the minimum ε i is below some fixed threshold Θ. For a
reliable classification, this threshold Θ has to be adapted such that the
ratio between the relative portion of the chosen class below and above
the threshold Θ is maximum.
Applying this technique to all three types of skin lesions, three
different sub-parts are defined by the corresponding thresholds, as also