Métodos de clasificación para identificar lesiones en piel a partir de espectros de reflexión difusa

Abstract

In order to differentiate between benign and malignant lesions in the human skin using diffuse reflection spectra, different classification algorithms were tested using the WEKA data mining software. In addition, due to the high dimensionality of the spectral signal, an attribute selection technique was applied to determine the variables that contribute with more information. The spectral signal classification was tested using support vector machines, neural networks and random forests, their performance was measured using the k-fold cross-validation percentages of the Kappa statistic, area under the ROC curve, specificity and sensitivity. Finally it is shown that the one layer neural network with 6 neurons and the parameters momentum and learning rate in 0.6 and 0.3 respectively, is best suited to the problem of pattern recognition, achieving correctly classify 89.89% of the cases.