Exploiting hybrid features extracted from color spaces of dermoscopic images

In the context of detecting dangerous illnesses of the patient affected by skin lesions, this study is focused on the problem of improving the classification of melanoma vs. normal nevus considering neural network and new proposed features based on high and low pixels intensity, and densities of the pattern detected with hit-or-miss (HoM) morphological transform. A large set of features from color spaces such as RGB (red, green, and blue), HSV (hue, saturation, and value) and L*a*b* (lightness variable L*, chromaticity indices a* and b*) were initially extracted and grouped in two parts: the first group includes first and second order features, and the second group contains hybrid features (a mix of first order, second order features and proposed features). The dimensionality of the features from these two groups is considerably reduced with principal component analysis (PCA) statistical method and fed three different types of neural network (NN): Radial Basis Function Neural Network (RBFNN), Cascade-Forward Neural Network (CFNN) and Pattern Recognition Neural Network (PRNN). From this corroboration of hybrid features, color space and NN, the L*a*b* color space with CFNN provided accuracy (ACC) of 0.897, F1-score of 0.91 and area under the ROC curve (AUC) of 0.896, respectively.