Tuning Cost-Sensitive Boosting and Its Application to Melanoma Diagnosis

This paper investigates a methodology for effective model selection of cost-sensitive boosting algorithms. In many real situations, e.g. for automated medical diagnosis, it is crucial to tune the classification performance towards the sensitivity and specificity required by the user. To this purpose, for binary classification problems, we have designed a cost-sensitive variant of AdaBoost where (1) the model error function is weighted with separate costs for errors (false negative and false positives) in the two classes, and (2) the weights are updated differently for negatives and positives at each boosting step. Finally, (3) a practical search procedure allows to get into or as close as possible to the sensitivity and specificity constraints without an extensive tabulation of the ROC curve. This off-the-shelf methodology was applied for the automatic diagnosis of melanoma on a set of 152 skin lesions described by geometric and colorimetric features, out-performing, on the same data set, skilled dermatologists and a specialized automatic system based on a multiple classifier combination.