Skin Lesion Classification - Biomedical Signal Analysis: Contemporary Methods and Applications - page 343

classification based on local ICA using SOM

90

85

80

75

70

65

60

55

50

2 4 8 16 32 64 128

patch length

Figure 12.11
The classification rates for the three skin lesions strongly depends on the patch size.
A patch size of at least 16 × 16 is needed to obtain a reliable classification rate.
While for large patch sizes the classification rate for basal cell carcinoma (+)
strongly decreases, for psoriasis ( × ) a further enhancement can be noted, leading to
the ICA results.

geneous identification of the different skin lesion ensembles is achieved.
As the clustering step is added to capture nonlinear trends in the
multivariate data, various patch sizes have to be analyzed as generally
nothing is known about the inherent structure of the fluorescence images.
In figure 12.11 the classification rates for the three different skin lesion
ensembles, based on a training set of psoriasis images, are displayed,
depending on the patch sizes: At least a size of 16 16 = 256 pixels is

needed to obtain a reasonable recognition rate. For patch sizes between
16

×

32 pixels per patch, on the average a relatively good
recognition rate is achieved. Thereby the classification rate of psoriasis
images (80%) still considerably exceeds the values obtained for basal cell
carcinoma and actinic keratosis (70%). A further increase of the patch

× 16 and 32 ×