Skin Lesion Classification - Biomedical Signal Analysis: Contemporary Methods and Applications - page 346

the most relevant information, and thus no further enhancement of the
classification rate might be understandable.
And there is a second reason, while PCA provides a statistical tool based
on an exact algebraic solution and independent of the probability distri-
bution, the neural implementation of the Bell-Sejnowski ICA algorithm
strongly depends on the assumed source density. Due to the computa-
tional load of the high-dimensional data, no adaptive techniques could
be applied for an accurate modeling of the source densities. Although
the distributions of the fluorescence images are super-Gaussian, and
consequently the assumption of a Laplacian source distribution proves
reasonable, slight deviations may involve errors which automatically
lower the level of accuracy, and therefore the classification rate.

Using local ICA based on Kohonen's SOMs, the cluster size proved
most essential for a reliable classification. Patches with sizes up to 8

8
pixels did not contain sucient spatially structured information to al-
low a further increase of the classification rates. However, using larger
patches (16

×

32 pixels), the obtained results clearly out-
performed the classification rates achieved by PCA and ICA. It must to
be noted that a strong simultaneous increase of the rates for all ensem-
bles was obtained when evaluating patches of 16

× 16 and 32 ×

16 pixels. Obviously,
corresponding spatial structures in the fluorescence images allow high
classification rates. However, a Fast Fourier Transformation (FFT) could
not prove this hypothesis.
For larger patch sizes (64

×

128 pixels), a further increase of
the classification rates for psoriasis and actinic keratosis could be noted.
However, at the same time, the results for basal cell carcinoma deterio-
rated considerably until the values for (global) ICA were obtained. This
might be due to the inherent structure of basal cell carcinoma, again
only an assumption which could not be evidenced by an FFT analysis.

× 64 and 128 ×

For answering the question of when to apply which classification
method, two circumstances have to be taken into account: In the case
where a high classification for a single skin lesion is needed ( Does image
i belong to class ? ), the applied method depends on the desired clas-

sification class: while for actinic keratosis PCA and ICA showed equally
high classification rates (72%), basal cell carcinoma can be identified
best by using local ICA based on SOM, resulting in an average classifi-

A